Structure-based Drug Design with Equivariant Diffusion

Models
Arne Schneuing1*† , Charles Harris2† , Yuanqi Du3† , Kieran Didi2 ,
Arian Jamasb2,9 , Ilia Igashov1 , Weitao Du4 , Carla Gomes3 ,
Tom L. Blundell2,10 , Pietro Lio2,5 , Max Welling6,11 , Michael Bronstein7,8 ,
Bruno Correia1*
arXiv:2210.13695v3 [[Link]] 23 Sep 2024

1 ÉcolePolytechnique Fédérale de Lausanne, Lausanne, Switzerland.

2 University of Cambridge, Cambridge, UK.
3 Cornell University, Ithaca, USA.
4 Chinese Academy of Mathematics and System Science, Beijing, China.
5 University of Rome “La Sapienza”, Rome, Italy.
6 Microsoft Research AI4Science, Amsterdam, Netherlands.
7 University of Oxford, Oxford, UK.
8 AITHYRA Institute, Vienna, Austria.
9 Current affiliation: Prescient Design, Genentech, Basel, Switzerland.
10 Current affiliation: Heart and Lung Research Institute, University of Cambridge,

Cambridge, UK.
11 Current affiliation: University of Amsterdam, Amsterdam, Netherlands.

*Corresponding author(s). E-mail(s): [Link]@[Link]; [Link]@[Link];

† These authors contributed equally to this work.

Abstract
Structure-based drug design (SBDD) aims to design small-molecule ligands that bind with high
affinity and specificity to pre-determined protein targets. Generative SBDD methods leverage
structural data of drugs in complex with their protein targets to propose new drug candidates.
These approaches typically place one atom at a time in an autoregressive fashion using the binding
pocket as well as previously added ligand atoms as context in each step. Recently a surge of
diffusion generative models has entered this domain which hold promise to capture the statistical
properties of natural ligands more faithfully. However, most existing methods focus exclusively on
bottom-up de novo design of compounds or tackle other drug development challenges with task-
specific models. The latter requires curation of suitable datasets, careful engineering of the models
and retraining from scratch for each task. Here we show how a single pre-trained diffusion model
can be applied to a broader range of problems, such as off-the-shelf property optimization, explicit
negative design, and partial molecular design with inpainting. We formulate SBDD as a 3D-
conditional generation problem and present DiffSBDD, an SE(3)-equivariant diffusion model that
generates novel ligands conditioned on protein pockets. Our in silico experiments demonstrate
that DiffSBDD captures the statistics of the ground truth data effectively. Furthermore, we show
how additional constraints can be used to improve the generated drug candidates according to
a variety of computational metrics. These results support the assumption that diffusion models
represent the complex distribution of structural data more accurately than previous methods,
and are able to incorporate additional design objectives and constraints changing nothing but
the sampling strategy. We anticipate that our findings may contribute to accelerate progress on
several computational drug design frontiers as more powerful distribution learners emerge, that
can be inserted into our flexible framework.

Keywords: structure-based drug design, conditioned diffusion models, equivariance

1
1 Introduction

The rational design of small-molecules with drug-like properties remains an outstanding challenge
in both fundamental and biopharmaceutical research. Structure-based drug design (SBDD) aims to
find small-molecule ligands that bind to specific three-dimensional sites in proteins with high affinity
and specificity [1]. Traditionally, SBDD campaigns are usually initiated either by high-throughput
experimental or virtual screening [2, 3] of large chemical databases. Generally, these approaches
are expensive and time-consuming, but they also restrict the exploration of the chemical space to
previously studied molecules, with a further emphasis usually placed on commercial availability [4].
Moreover, the optimization of initial lead molecules is often a biased process, with significant reliance
on human intuition [5]. Recent advances in geometric deep learning, especially in modelling geometric
structures of biomolecules [6–8], provide a promising direction for SBDD [9]. Despite remarkable
progress in the use of deep learning as surrogate docking models [10–12], deep learning-based design
of ligands that bind to target proteins remains an overarching problem in molecular modeling. Early
attempts have been made to represent molecules as atomic density maps, with variational auto-
encoders generating new atomic density maps corresponding to novel molecules [13]. However, it
is nontrivial to map atomic density maps back to molecular space, requiring an additional atom-
fitting stage. An alternative is to represent molecules as 3D graphs with atomic coordinates and
types which naturally circumvents the post-processing steps. Li et al. [14] proposed an autoregressive
generative model to sample ligands given the protein pocket as a conditioning constraint. Peng
et al. [15] improved this method by using an E(3)-equivariant graph neural network which respects
rotation and translation symmetries in 3D space. Similarly, Drotár et al. [16] and Liu et al. [17] used
autoregressive models to generate atoms sequentially and incorporate angles during the generation
process. However, the main premise of sequential generation methods may not hold in real scenarios,
since it imposes an artificial ordering scheme in the generation process and, as a result, the global
context of the generated ligands may be lost. Very recently, a number of diffusion models have been
put forward for target-specific molecule design [18–22]. These models place all atoms simultaneously,
allowing them to reason about the whole molecule at once and typically enabling faster sampling.
While this class of models has already shown great promise in de novo ligand generation, their
potential in other parts of the drug design pipeline has not been thoroughly explored.

In this study, we propose DiffSBDD, an SE(3)-equivariant 3D-conditional diffusion model for SBDD
that respects translation, rotation, and permutation symmetries. With a free de novo generation
benchmark, we first establish that the diffusion model captures molecular properties of real molecules
more accurately than previously popular autoregressive models. Secondly, we show how the design
space can be effectively constrained based on prior knowledge to improve the sample quality. Our
inpainting-inspired approach furthermore allows us to tackle a diverse set of molecular design prob-
lems, such as scaffold hopping/elaboration and fragment growing/merging, all without the need to
retrain new models. Lastly, we demonstrate how pre-trained DiffSBDD models can be used out-of-
the-box to optimize arbitrary molecular properties via a noise/denoise scheme in combination with
an evolutionary algorithm.

2
A Diffusion framework B Graph modeling
Ligand TYR

C Fixing known substructures Mask Intramolecular edges

Input atom
Input fragment Generated atom

Noise Mask
input

Intermolecular edges

Denoising Mask
step

Iterate
D Property optimization
1 2

• Drug-likeness
• Affinity
• Solubility
• Synthesizeability
• ...

Starting Query oracle/ Rank Improved Optimized

Noise/denoise
molecule score function molecules molecules molecules

Iterate

E Reflection sensitivity F Neural network architecture

Node-level operations Message passing Node-level operations

Ligand encoder
MLP Ligand decoder
MLP
R-citalopram
SE(3)-GNN

SE(3)-GNN

SE(3)-GNN

Pocket encoder Pocket decoder

MLP MLP

S-citalopram

3
Fig. 1: (A) Overview of the 3D diffusion setup. The diffusion process q yields a noised version
of the original atomic point cloud for a time step t ≤ T . The neural network model is trained to
approximate the reverse process conditioned on the target protein structure z (P ) . Once trained, an
(L)
initial noisy point cloud is sampled from a Gaussian distribution zT ∼ N (0, I) and progressively
denoised using the learned model. Covalent bonds are added to the resultant point cloud at the
end of generation. (B) Each state is processed as a graph where edges are introduced according to
P −P
distance thresholds within the ligand dL−L
max , within the pocket dmax and between ligand and pocket
nodes dL−P
max . (C) Replacement method for fixing molecular substructures. To complete the known
part of the molecule (orange) with newly generated chemical matter (green atoms), we apply the
learned denoising process to the entire molecule (orange & green), but at every step we replace the
prediction for the known part (orange) with the ground-truth noised version computed with q. The
protein context (gray) remains unchanged in every step. (D) Iterative procedure to tune molecular
features. We find variations of a starting molecule by applying small amounts of noise and running an
appropriate number of denoising steps. The new set of molecules is ranked by an arbitrary oracle and
the procedure is repeated for the strongest candidates. (E) Antidepressant Citalopram as an example
in which stereochemistry is essential for its therapeutic effect. (F) The neural network backbone is
composed of MLPs that map scalar features of ligand and pockets nodes into a joint embedding
space, and SE(3)-equivariant message passing layers that operate on these features and each node’s
coordinates. It outputs the predicted noise values for every vertex.

2 Equivariant Diffusion Models for SBDD

We leverage equivariant denoising diffusion probabilistic models (DDPMs) [23, 24] to generate
molecules and binding conformations jointly with respect to a specific protein target. Figure 1A
schematically depicts the 3D diffusion procedure. During training, varying amounts of random noise
are applied to 3D structures of real ligands and a neural network learns to predict the noise-less
features of the molecules. For sampling, these predictions are used to parameterize denoising tran-
sition probabilities which allow us to gradually move a sample from a standard normal distribution
onto the data manifold. Both the protein and the ligand are represented as 3D point clouds, where
atom types are encoded as one-hot vectors, and all objects are processed as graphs. For improved
computational efficiency, we define independently tunable distance cutoffs for intermolecular edges
between nodes of the ligand and pocket and intramolecular edges between two nodes from the same
molecule (Figure 1B). This means information is only propagated between spatially proximal atoms.
Our neural network is designed to respect natural symmetries of the molecular system, which include
rotations and translations but excludes nonsuperposable transformations. That is, we process rigid
transformations in an equivariant way but not reflections. This design choice is motivated by well-
studied examples of drugs whose sterochemistry affects their activity and toxicity. For instance,
the antidepressent Citalopram (Figure 1E) has two enantiomers but only the S-enantiomer has the
desired therapeutic effect. The difference between the S- and R-form of the molecule, however, is
only detectable by a reflection-sensitive GNN (Appendix E). To process ligand and pocket nodes
with a single graph neural network (GNN), atom types and residue types are first embedded in a
joint node embedding space by separate learnable MLPs. We also experimented with coarse-grained
Cα descriptions of the pockets to reduce processing time even further but found this representation
to be inferior in most cases (Appendix F.5). Further technical details of the diffusion framework and
equivariant neural network are described in Method Sections 7.1 and 7.3.

To condition the 3D generative model on the structure of the protein pocket, we consider two distinct
approaches. In the first approach, DiffSBDD-cond, we provide fixed three-dimensional context in
(L)
each step of the denoising process. To this end, we supplement the ligand atomic point cloud zt ,
(P )
denoted by superscript L, with protein pocket nodes zdata , denoted by superscript P , that remain
unchanged throughout the reverse diffusion process (Figure 1A). For the second method, DiffSBDD-
(L) (P )
joint, we initially train a diffusion model to approximate the joint distribution p(zdata , zdata ) of ligand-
pocket pairs, and inject information about target pockets only at inference time. The methodology
is analogous to the substructure inpainting approach described below (Section 7.4 and Figure 1C).

4
Table 1: Evaluation of generated molecules for targets from the CrossDocked and Binding MOAD
test sets. To assess how well the models capture properties of real ligands, we compute the Wasser-
stein distance between the distributions of a scores from generated molecules and the ground truth
molecules from the test sets. The best performance (i.e. lowest Wasserstein distance) is highlighted
in bold. ∗ denotes that we re-evaluate the generated ligands provided by the authors. † means infer-
ence times are taken from the original paper. ‡ means inference time estimated based on five targets.
QED: Quantitative Estimation of Drug-likeness [27]; SA: Synthetic Accessibility [28]; LogP : parti-
tion coefficient [29]; CNN : Convolutional Neural Network.
Wasserstein distance to reference distribution
QED SA LogP Lipinski Vina score CNN affinity Time (s, ↓)
∗
Pocket2Mol [15] 0.106 0.24 0.879 0.647 0.596 0.616 2504 ± 2207†
CrossDocked

ResGen [30] 0.116 0.556 0.778 0.673 2.15 1.21 ≈ 936‡

DiffSBDD-cond 0.0205 1.29 0.589 0.261 0.848 0.145 135.866 ± 51.66
DiffSBDD-joint 0.0187 1.51 1.67 0.405 0.694 0.281 160.314 ± 73.30
≈ 613‡
Binding MOAD

Pocket2Mol [15] 0.123 0.93 0.841 0.269 3.4 1.58

ResGen [30] 0.0727 0.92 0.831 0.237 6.93 1.71 ≈ 697‡
DiffSBDD-cond 0.0924 1.15 0.833 0.143 4.26 0.903 336.061 ± 85.02
DiffSBDD-joint 0.0716 0.76 1.23 0.183 9.56 0.628 369.873 ± 124.54

Both approaches are equally applicable to the small molecule design task and in practice differ only
in whether the neural networks expects the original pocket or a noisy version as input.

To evaluate our approach we first show that diffusion models are a powerful framework for learn-
ing the distribution of three-dimensional molecular data by generating new target-specific ligands de
novo without additional constraints or optimizing a particular property (Section 3). We then demon-
strate how the flexibility of diffusion models enables partial molecular redesign to incorporate specific
design constraints without needing to develop specialised models (Sections 4), and iterative improve-
ment of molecular properties measured by arbitrary oracles (Section 5). While we provide empirical
results only for our model, the methodology can be readily used in combination with other recently
published diffusion models for molecule design [18–22]. Finally, we have curated an experimentally
determined binding dataset derived from Binding MOAD [25], which supplements the commonly
used synthetic dataset CrossDocked [26], to validate our model’s performance under realistic binding
scenarios. While protein-ligand pairs of the former might contain non-native contacts, since ligands
were cross-docked into structurally similar binding pockets, our new dataset consists exclusively of
experimentally validated interactions.

3 DiffSBDD captures the data distribution faithfully

As a first test to our model, we probe its ability to accurately represent the properties of real
binders, and compare the results with Pocket2Mol [15] and ResGen [30], two recently published
autoregressive models, which represent the previous state-of-the-art class of machine learning models
for structure-based drug design. We use publicly available code and weights of the models [31, 32].

Since distribution learning capabilities in the high-dimensional space of chemical compounds are dif-
ficult to quantify directly, we instead measure a range of molecular properties that are relevant for
potential drug candidates. We then compare the distributions of these scores to the distributions we
get from the real ligands in our test set using the Wasserstein distance. These results are summarized
in Table 1 for computational scores of drug-likeness (QED), synthetic accessibility (SA), hydropho-
bicity (LogP) and two measures of target affinity, the empirical Vina scoring function and a neural
network estimation of binding affinity (CNN affinity). Both were computed by GNINA [33] after
local energy minimization to resolve minor clashes. The underlying distributions of scores are shown
in supplementary Figures F2 and F3. We perform this analysis both for the test set targets from
CrossDocked [26], a standard dataset extensively used in prior works [13, 15, 17, 30], and our newly

5
 CrossDocked Binding MOAD
A C PDB 4tos D F PDB 6c0b
1.0
0.6
0.8
Tanimoto similarity

Tanimoto similarity
0.6 0.4

0.4 Vina: -6.76 Vina: -8.33

QED: 0.60 0.2 QED: 0.87
0.2 Enamine Sim.: 0.52 Enamine Sim.: 0.36
0.0 0.0
DiffSBDD-cond DiffSBDD-joint Pocket2Mol ResGen DiffSBDD-cond DiffSBDD-joint Pocket2Mol ResGen
Model Model
20 20

15 15
Vina: -11.16 Vina: -7.69

Vina score difference

Vina score difference

10 QED: 0.85 10 QED: 0.75

Enamine Sim.: 0.50 Enamine Sim.: 0.35
5 5

0 0

−5 −5

−10 −10
DiffSBDD-cond DiffSBDD-joint Pocket2Mol ResGen DiffSBDD-cond DiffSBDD-joint Pocket2Mol ResGen
Model Model
B E
Vina: -8.52 Vina: -10.28
QED: 0.53 QED: 0.67
2.5 Enamine Sim.: 0.46 2.0 Enamine Sim.: 0.32
Reference Reference
DiffSBDD-cond DiffSBDD-cond
2.0 DiffSBDD-joint Reference DiffSBDD-joint Reference
1.5
Pocket2Mol Pocket2Mol
Avg. count

Avg. count
1.5 ResGen ResGen
1.0
1.0

0.5
0.5 Vina: -15.64 Vina: -8.07
QED: 0.47 QED: 0.36
0.0 0.0
3 4 5 6 7 >7 3 4 5 6 7 >7
Ring size Ring size

Fig. 2: Evaluation of distribution learning capabilities and generated examples. All targets are taken
from the CrossDocked and Binding MOAD test sets. (A) Comparison of generated molecules with
the reference molecule from the same pocket. We compare Tanimoto similarity of the molecular
fingerprints and compute the difference Vinagen − Vinaref between their Vina docking scores. (B)
Average number of rings of different sizes per generated molecule. (C) Example molecules generated
by DiffSBDD-cond for a pocket from the CrossDocked test set. We compared all generated molecules
with the approximately 4.2M compounds from the Enamine Screening Collection, and selected the
three closest hits with drug-likeness QED > 0.5. Vina docking score, QED drug-likeness score and
fingerprint similarity to the most similar Enamine molecules are reported in each case. (D-F) Same
analyses but for target pockets from the Binding MOAD test set.

curated dataset based on Binding MOAD [25]. Note that Pocket2Mol and ResGen were trained on
the same CrossDocked training set but not on Binding MOAD.

Generally, our diffusion models capture molecular properties of natural ligands more accurately than
the autoregressive baselines despite significantly shorter sampling times. A notable exception is the
Vina score, which Pocket2Mol matches particularly well on the CrossDocked dataset. Interestingly,
this observation is not confirmed by GNINA’s CNN affinity which estimates the same quantity.
Its distribution is better approximated by DiffSBDD. Figure 2A shows that both DiffSBDD and
Pocket2Mol Vina scores are centered around the reference but the spread is larger in the case of the
diffusion models, which means their samples contain larger fractions of low scoring molecules but also
ligands that potentially bind more tightly than the native counterparts. The greater abundance of
high scoring molecules is particularly important in anticipation of downstream design applications,
where we often look for the most competitive binder rather than average candidates. A similar obser-
vation holds for the Binding MOAD dataset with experimentally determined binding complexes.
However, unlike the CrossDocked case, docking scores are lower on average than the scores of corre-
sponding reference ligands from this dataset. We believe the reason to be twofold: the Binding MOAD
training set is much smaller and also contains more challenging ground-truth ligands (native binders)
whereas CrossDocked complexes can have unrealistic protein-ligand interactions. This hypothesis is
supported by less favorable Vina scores of reference molecules from the synthetic dataset on average
(-7.68 vs. -9.17 kcal/mol). This result underscores the importance of high quality training sets for
SBDD models that aim to design high affinity binders. Lastly, the DiffSBDD models also produce
molecules that are slightly more similar to the reference on average (Figure 2A,D) and contain a
comparable amount of 5- and 6-rings to natural ligands (Figure 2B,E). However, very small and very
large ring systems are typically over-represented in DiffSBDD molecules.

6
 A Scaffold Hopping
2gm1 F Importance of resamplings

Increasing resamplings (r)

B Scaffold Elaboration
5ndu
Input r=1 r=2 r=5 r = 10

C Fragment Merging
5rue 5spd

5rsw
D Fragment Growing
5ndu

E Fragment Linking

5ndu

Fixed atoms Output Reference

Fig. 3: Molecular inpainting design examples for scaffold hopping (A), scaffold elaboration (B), frag-
ment merging (C), fragment growing (D) and fragment linking (E) respectively. The input to our
model (the fixed atoms) is shown in blue, the outputs (designed molecules) are shown in green and
the original molecule is shown in magenta for reference. PDB codes are shown for the ground truth
structure. In the case of fragment merging, we compose fragments with two different crystal struc-
tures with PDB codes shown. (F) Importance of resampling for generating realistic and connected
molecules. Top: Visual example; inpainted region (green) finally harmonizes with molecular context
at high resamplings. Bottom: Effect of the number of resampling steps on molecular connectivity.
Means and 95% confidence intervals are plotted for 3 design tasks. For this experiment we used 20
randomly selected targets from the test set.

Panels C and F of Figure 2 present a representative selection of molecules for one target from each
test set. The selection is filtered to contain examples which are drug-like (QED > 0.5) and similar
to purchasable molecules from the Enamine Screening Collection. These filters represent favorable
properties one might look for in a drug design campaign. The target with Protein Data Bank (PDB)
identifier 6c0b, for example, is a human receptor which is involved in microbial infection [34] and
possibly tumor suppression [35]. The reference molecule, a long fatty acid (see Figure 2F, bottom
panel) that aids receptor binding [34], has too high a number of rotatable bonds and low a number
of hydrogen bond donors/acceptors to be considered a suitable drug (QED of 0.36). Our model
however, generates drug-like (QED = 0.87 in the first example) and suitably sized molecules by
adding aromatic rings connected by a few rotatable bonds, which allows the molecules to adopt a
complementary binding geometry and is entropically favourable by reducing the degrees of freedom,
a classic approach in medicinal chemistry [36]. Larger random samples of generated molecules are
presented in the Appendix (Figure F6).

4 Generating novel chemical matter from known

substructures

In drug discovery it is very common to design molecules around previously identified potent sub-
structures. For example, we may wish to design a scaffold around a set of functional groups (scaffold

7
Table 2: Evaluation of molecular inpainting for fragment linking, scaffold hopping and scaffold
elaboration across the whole Binding MOAD test set. Percentage value next to task name denotes
the proportion of atoms fixed during the design process. DiffSBDD-baseline generates a whole new
molecule from scratch, DiffSBDD-diversify is elaborating around a fixed substructure of an existing
molecule and DiffSBDD-de novo is designing new motifs around a fixed substructure from stratch. t
and r are the number of partial noising steps or resamplings for diversify and de novo respectively.
Note, here we use a subset of our original test set for which we can calculate masks for all design
tasks (n = 55).
Vina (↓) Validity (↑) Connectivity (↑) QED (↑) SA (↓) Diversity (↑)
Test set −9.86 ± 1.6 1 1 0.543 ± 0.16 3.86 ± 1.2 —
DiffSBDD-baseline −5.69 ± 6 0.964 0.589 0.419 ± 0.2 4.99 ± 1.1 0.701 ± 0.09
Fragment linking (73.43% atoms fixed)
DiffSBDD-de novo (r = 20) −7.74 ± 1.8 0.796 0.558 0.322 ± 0.17 5.38 ± 0.76 0.486 ± 0.09
DiffSBDD-diversify (t = 100) −8.72 ± 1.7 0.991 0.968 0.476 ± 0.14 4.26 ± 1.1 0.35 ± 0.089
DiffLinker [41] −6.92 ± 2.7 0.947 0.840 0.349 ± 0.19 4.72 ± 0.97 0.453 ± 0.13
Scaffold hopping (27.32% atoms fixed)
DiffSBDD-de novo (r = 20) −7.6 ± 2.5 0.782 0.663 0.39 ± 0.18 5.29 ± 0.72 0.612 ± 0.074
DiffSBDD-diversify (t = 100) −8.95 ± 1.8 0.977 0.948 0.492 ± 0.18 4.39 ± 1.0 0.479 ± 0.1
Scaffold elaboration (72.68% atoms fixed)
DiffSBDD-de novo (r = 20) −8.1 ± 1.8 0.852 0.445 0.388 ± 0.2 5.2 ± 0.7 0.397 ± 0.11
DiffSBDD-diversify (t = 100) −9.32 ± 1.7 0.995 0.971 0.516 ± 0.18 4.12 ± 1.1 0.282 ± 0.14

hopping) or extend an existing fragment to make a whole molecule (fragment growing). Generating
compounds, or parts thereof, conditioned on a given molecular context is reminiscent of inpainting,
a technique originally introduced for completing missing parts of images [37, 38] but also adopted
in other domains, including biomolecular structures [39]. We can realise a number of drug discov-
ery sub-tasks via an inpainting technique known as the ‘replacement method’ [38, 40], whereby we
add new atoms in and around fixed regions of the substructure to design whole molecules (Fig. 1C
and Methods Section 7.4). Unlike previous methods, using DiffSBDD in this way does not require
retraining a model on any specialized or synthetic datasets. Curating such datasets is often time and
labor-intensive, and typically relies on potentially sub-optimal assumptions (e.g. definition of frag-
ments) to convert a general dataset of small molecules into a task-specific dataset that can be used
to train specialised models. With our proposed approach, by contrast, the simple definition of an
arbitrary binary mask is sufficient for the diffusion model to generalize to any inpainting task whilst
using a neural network trained on all available protein-ligand data in raw form.

Examples of molecules designed with inpainting for scaffold hopping, scaffold elaboration, fragment
merging, fragment growing, and fragment linking are shown in Figure 3. All molecular inpainting
experiments use a version of DiffSBDD trained on Binding MOAD. Figure 3A shows an example
of scaffold hopping a design for a mitotic kinesin Eg5 inhibitor (PDB code 2gm1) [42] where we fix
the functional groups mediating the binding to the pocket whilst designing a new scaffold structure.
Figure 3B shows the opposite case of scaffold elaboration for a rationally designed oncology inhibitor
targeting a phosphoprotein (PDB code 5ndu) [43] where we fix the scaffold and design new functional
groups. Figure 3C shows an example of fragment merging. We successfully replicate the results of
Gahbauer et al. [44], which performed fragment merging of two fragments (PDB code 5rsw and 5rue)
identified by experimental screening [45] for the SARS-CoV-2 non-structural protein 3 (Nsp3) using
the chemoinformatics-based method Fragmenstein1 . Figure 3D shows an example of fragment growing
around the central motif of PDB entry 5ndu. Finally, Figure 3E gives an example of fragment linking
between two outer fragments of PDB entry 5ndu. Note that here we are not only designing a small
linker made of a few atoms but rather an entirely new fragment with two connecting linkers. Such
a linker design would be substantially out-of-distribution for previous methods, due to the size and
complexity of the linker relative to those used during training [46]. Further implementation details
of the fragment merging experiment are given in Appendix Section G.1.

Depending on the use case, we find it desirable to perform molecular inpainting within two regimes;
(i) designing a completely new inpainted region de novo (DiffSBDD-de novo) as to explore the entire
chemical fitness landscape, or (ii) redesigning an existing region (e.g. a scaffold) via partial noising

1
[Link]/matteoferla/Fragmenstein

8
then denoising (see Section H), thus locally exploring desired properties by exploitation (DiffSBDD-
diversify). The first case is more amenable to situations in which we have no prior information
other than the fixed substructure (e.g. fragment linking after a fragment screen), meaning that
unconstrained exploration of the chemical fitness landscape is the preferred approach for the majority
of SBDD. The second case is more relevant in scenarios where we have prior information about
the desired chemical and topological composition of the designed region which we can use to bias
generation (with the choice of t being a hyperparameter). This is particularly relevant in the case of
scaffold hopping, where we are trying to keep the properties of a molecule relatively unchanged whilst
designing a new topology [47]. To investigate the differences between both approaches quantitatively,
we further tested our method systematically for the whole Binding MOAD test set in the tasks of
linker design, scaffold hopping and scaffold elaboration (Table 2). As baselines, we provide the metrics
for the molecules in the test set, molecules only conditioned on the pocket with no fixed substructure
(DiffSBDD-baseline) and the performance of DiffLinker [41] in the fragment linking task. Due to the
fact that we can only calculate fragment and scaffold masks for larger molecules, we have reduced
the size of the test set used in Table 2 to n = 55 to ensure fair comparison between methods.

Constraining fixed regions to highly complementary substructures within the protein pocket sig-
nificantly enhances Vina scores using DiffSBDD-de novo compared to the DiffSBDD-baseline. For
molecules generated without substructure conditioning under DiffSBDD-baseline, the average Vina
score is -5.69. In contrast, DiffSBDD-de novo sees improved scores, achieving -7.74 kcal/mol when
used for linker design from starting fragments and achieves results comparable to the specialist model
DiffLinker. In the case of scaffold elaboration, DiffSBDD-de novo significantly boosts docking scores
from -5.69 kcal/mol to -8.10 kcal/mol over the baseline, by focusing on the optimal placement of
functional groups to facilitate key residue binding on a pre-existing scaffold. Moreover, there is a
notable improvement in average docking scores for scaffold hopping, with scores rising from -5.69
kcal/mol to -7.60 kcal/mol when using DiffSBDD-de novo. This enhancement is achieved despite a rel-
atively small proportion of atoms being fixed, about 27.32%, compared to other tasks. The significant
increase in docking scores is attributed to the nature of the fixed atoms, primarily functional groups
that form the pharmacophore, which are crucial for binding affinity. For detailed implementation,
see Appendix G.2.

Similar to earlier findings, the replacement method often produces poor and inconsistent outcomes.
Following the approach by Lugmayr et al. [38], we enhanced the sample quality by refining interme-
diate states iteratively before progressing in the denoising process, a technique called resampling, as
detailed in the Methods Section 7.4. This technique is crucial for seamlessly integrating modified and
original areas and proves essential in the molecular context. Minimal resampling resulted in chem-
ically valid but disjointed structures, while increased iterations led to coherent molecules, even in
complex scenarios with extensive modifications needed (Fig. 3F). Our results indicate that the effect
of resampling on molecular connectivity is particularly pronounced (Fig. 3F) but it also significantly
impacts another metrics (Appendix Figure G7B-D).

5 Molecule optimization: iterative search for better molecule

candidates

For hit identification and optimization of lead molecules in real use cases, it is not enough to just
sample molecules from the whole training data distribution. Instead, we are usually interested in the
better performing tail of the distribution, and only want to pursue the most promising candidates.
Since we could show that DiffSBDD recapitulates the chemical space of the training set including
high-scoring molecules, we should always find promising drug candidates with strong docking scores,
synthetic accessibility and other desired properties. Here we propose a simple protocol to access them
efficiently (Figure 1D).

We first noise a molecule from an experimental protein-ligand complex for t steps, where t ≪ T , using
the forward diffusion process. From this partially noised sample, we can then denoise the appropriate
number of steps with the reverse process until t = 0. The stochasticity in this quick noise/denoise

9
Fig. 4: Results on molecular optimization using DiffSBDD. (A-D) Experiments on single property
molecular optimization. (A) Starting molecule from PDB code 5NDU. (B) QED optimization over 8
generations. (C) SA optimization over 7 generations. (D) docking score optimization over 3 genera-
tions. We found that optimization over subsequent generations continuously optimized the docking
score, but that was at expense of molecular quality. (E-G) Kinase inhibitor specificity optimization
experiment. (E) Cartoon representation showing the high degree of structural similarity between our
two kinases of interest (BIKE and MPSK1). (F) Trajectory plot showing the highest scoring molecule
at each iteration during kinase inhibitor optimization. (G) Visual representation of the molecular
graphs and bound conformations of the native and final molecules with corresponding Vina docking
scores. Boxes in panels (B-D) represent the upper and lower quartile as well as the median of the
data. Whiskers denote 1.5 times the interquartile range. Outliers outside this range are shown as flier
points.

process allows us to sample new and diverse candidates of various properties whilst staying in the
same region of chemical space, assuming t is small (see Appendix Figure H8). This approach is
inspired by [48] but note this does not allow for direct optimization of specific properties. Instead,
it can be regarded as an exploration around the local chemical space whilst maintaining high shape
and chemical complementary via the conditional denoising model.

We extend this idea by combining the partial noising/denoising procedure with a simple evolutionary
algorithm that optimizes for specific molecular properties (Figure 1D). We find that our model
performs well at this task out of the box without requiring additional fine-tuning. At every stage
in the optimization process, we generate 100 new molecules (from either the previous generation or
the original molecule in the first case). Molecules are modified via partial noising/denoising with
a randomly chosen t between 10 and 150. The new molecules are then passed to an oracle/score
function (e.g. docking program or synthetic accessibility predictor) to be ranked. The top k molecules
are then selected to seed the new population. In our study, we use k = 10.

In Figure 4A-D, we optimize an inhibitor molecule found in PDB entry 5ndu [43], which has poor
SA and QED scores, 7.21 and 0.35 respectively, but high binding affinity. Over a number of rounds
of optimization, we can observe substantial increases in QED, from 0.35 to mean of 0.43, whilst still
maintaining high similarity to the original molecule. We can also rescue the low synthetic accessibility
score of the seed molecule by producing a battery of highly accessible molecules when selecting for
SA. Finally, we observe that we can perform significant optimization of binding affinity after only a
few rounds of optimization.

10
To demonstrate the power of molecular optimization with DiffSBDD, we consider the challenging
case of highly selective kinase inhibitor design (Figure 4E-G). In our experiment, we perform positive
design against our on-target kinase BIKE (PDB code 4w9w) [49] whilst simultaneously perform-
ing negative design against the structurally similar off-target kinase MPSK1 (PDB code 2buj) [50]
(Fig. 4E) by optimizing for L. Furthermore, before selecting the top molecules, we pruned any can-
didates that regress with regard to the on- and off-target docking scores of the original molecule (i.e.
those above or left of the red star in Fig. 4F) in order to bias the molecules to have high affinity to the
on-target kinase as well as specificity. The starting molecule (ChEMBL identifier CHEMBL388978)
had an on- and off-target docking score of −7.2 kcal/mol and −10.8 kcal/mol respectively. After 5
rounds of optimization, we had improved the on- and off-target docking scores for the top molecule
to −13.9 kcal/mol and −8.7 kcal/mol respectively (Fig 4G), demonstrating substantially improved
specificity and reduced off-target activity. Furthermore, all of this is achieved whilst maintaining a
high QED score (from 0.54 to 0.52) and improving the SA score from 6.31 to 4.60.

6 Conclusion

Many machine learning methods for structure-based drug design create molecules one atom at a time,
thereby imposing an arbitrary order on the generative process and preventing them from reasoning
about molecules holistically. Moreover, most existing approaches focus exclusively on the de novo
generation of new ligands from scratch, which often limits their sample quality and synthesizability,
and ultimately hinders lab validation of designs. In this work, we investigated the distribution learn-
ing capabilities of 3D-conditional diffusion models as an alternative to the autoregressive paradigm.
To this end, we developed DiffSBDD, an SE(3)-equivariant 3D-conditional diffusion model that
generates small molecule ligands for given target pocket structures, with chemical properties that
closely match those of native ligands. Since, on the application side, medicinal chemists typically have
concrete design specifications in mind, we studied different substructure redesign and optimization
techniques, and showed how these can be used to improve a range of desirable properties of candidate
compounds, such as computational docking and drug-likeness scores. Constraining the problem to
realistic substructures like fragments or scaffolds leads to better designs because it prevents the neu-
ral network from overly hallucinating. Retaining substructures of previously synthesized molecules
facilitates chemical synthesis and experimental testing. Moreover, the capability to further ‘locally’
optimize designed ligands is important in real-world drug discovery and effectively improves the qual-
ity of the initial designs. For similar applications, previous works typically resorted to specialised
models that were trained on tailored datasets and performed well only on narrowly defined tasks.
Here, we provided evidence that a powerful general diffusion model can be used as a drop-in replace-
ment for these specialised models if the sampling procedure is modified appropriately. This means
we can expect better performance in all discussed sub-tasks, solely by improving the distribution
learning capabilities and sample quality of the main model.

In general, data-driven techniques are emerging as a suitable tool to tackle the immense complexity
of the biochemical design space, which is extremely hard to navigate with mechanistic approaches.
While the purely de novo design of novel chemical matter remains challenging, we could show that
learning-based tools are ready to be incorporated in drug development pipelines if additional design
constraints are enforced. In the future, we foresee an increasing importance of machine learning
models in this domain as they are improved further. Developing more informative metrics and more
reliable benchmarks will be an important step towards reaching this goal because they reduce the
reliance on visual inspection of examples and expert judgment and thus accelerate progress. Lastly,
we envision that other iterative sampling techniques, including flow matching and Schrödinger bridge
models, will benefit from the presented inference strategies as well.

11
7 Methods

7.1 Denoising Diffusion Probabilistic Models

Denoising diffusion probabilistic models (DDPMs) [23, 51] are a class of generative models inspired
by non-equilibrium thermodynamics. Briefly, they define a Markovian chain of random diffusion steps
by slowly adding noise to sample data and then learning the reverse of this process (typically via a
neural network) to reconstruct data samples from noise.

In this work, we closely follow the framework developed by Hoogeboom et al. [24]. In our setting,
data samples are atomic point clouds zdata = [x, h] with 3D geometric coordinates x ∈ RN ×3 and
categorical features h ∈ RN ×d , where N is the number of atoms. A fixed noise process

q(zt |zdata ) = N (zt |αt zdata , σt2 I) (1)

adds noise to the data zdata and produces a latent noised representation zt for t = 0, . . . , T . αt
controls the signal-to-noise ratio SNR(t) = αt2 /σt2 and follows either a learned or pre-definedp
schedule
from α0 ≈ 1 to αT ≈ 0 [52]. We choose a variance-preserving noising process [37] with αt = 1 − σt2 .

Since the noising process is Markovian, we can write the denoising transition from time step t to
s < t in closed form as
2 2
 α σ2
t|s s αs σt|s σt|s σs2 
q(zs |zdata , zt ) = N zs zt + zdata , I (2)
σt2 σt2 σt2

with αt|s = ααst and σt|s

2 2
= σt2 − αt|s σs2 following the notation of Hoogeboom et al. [24]. This true
denoising process depends on the data sample zdata , which is not available when using the model for
generating new samples. Instead, a neural network ϕθ is used to approximate the sample ẑdata . More
specifically, we can reparameterize Equation (1) as zt = αt zdata + σt ϵ with ϵ ∼ N (0, I) and directly
predict the Gaussian noise ϵ̂θ = ϕθ (zt , t). Thus, ẑdata is simply given as ẑdata = α1t zt − ασtt ϵ̂θ .

The neural network is trained to maximize the likelihood of observed data by optimizing a variational
lower bound on the data, which is equivalent to the simplified training objective [23, 52] Ltrain =
1 2
2 ||ϵ − ϕθ (zt , t)|| up to a scale factor (see Appendix A for details).

7.2 Equivariance

Structural biology remains a rather data-sparse domain. It is therefore common practice to encode
known geometric constraints, typically equivariance to rotations and translations, directly into the
neural network architecture, thereby facilitating the learning task because possible neural opera-
tions are limited to a meaningful subset. In the 3D molecule generation setting, we explicitly exclude
reflection-equivariant operations because they would make the model blind to some aspects of stereo-
chemistry. It is known that different stereoisomers can have significantly different therapeutic effects
(e.g., [53], Figure 1E) and might even lead to unforeseen off-target activity and hence toxicity. We
therefore developed a reflection-sensitive system that is SE(3)- rather than E(3)-equivariant although
the latter is more commonly adopted in related works [18, 24, 54].

Technically, we ensure SE(3)-equivariance in the following sense2 : evaluating the likelihood of a

molecule x(L) ∈ R3×NL given the three-dimensional representation of a protein pocket x(P ) ∈ R3×NP
should not depend on global SE(3)-transformations of the system, i.e. p(Rx(L) + t|Rx(P ) + t) =
p(x(L) |x(P ) ) for orthogonal R ∈ R3×3 with RT R = I, det(R) = 1 and t ∈ R3 added column-wise.
At the same time, it should be possible to generate samples x(L) ∼ p(x(L) |x(P ) ) from this conditional
probability distribution so that equivalently transformed ligands Rx(L) +t are sampled with the same

2
We ignore node type features, which transform invariantly, for simpler notation.

12
probability if the input pocket is rotated and translated and we sample from p(Rx(L) + t|Rx(P ) + t).
This definition explicitly excludes reflections which are connected with chirality and can alter the
biomolecule’s properties.

In our set-up, equivariance to the orthogonal group O(3) (comprising rotations and reflections) is
achieved because we model both prior and transition probabilities with isotropic Gaussians where the
mean vector transforms equivariantly with respect to rotations of the context (see Hoogeboom et al.
[24] and Appendix D). Ensuring translation equivariance, however, is harder because the transition
probabilities p(zt−1 |zt ) are not inherently translation-equivariant. In order to circumvent this issue,
we follow previous works [24, 55, 56] by limiting the whole sampling process to a linear subspace
where the center of mass (CoM) of the system is zero. In practice, this is achieved by subtracting
the center of mass of the system before performing likelihood computations or denoising steps. Since
equivariance of the transition probabilities depends on the parameterization of the noise predictor
ϵ̂θ , we can make the model sensitive to reflections with a simple additive cross-product term in the
EGNN’s coordinate update as discussed in Section 7.3 and Appendix E.

7.3 SE(3)-equivariant Graph Neural Networks

A function f : X → Y is said to be equivariant w.r.t. the group G if f (g.x) = g.f (x), where g.
denotes the action of the group element g ∈ G on X and Y [57]. Graph Neural Networks (GNNs)
are learnable functions that process graph-structured data in a permutation-equivariant way, making
them particularly useful for molecular systems where nodes do not have an intrinsic order. Permuta-
tion invariance means that GNN(ΠX) = Π GNN(X) where Π is an n × n permutation matrix acting
on the node feature matrix.

Since the nodes of the molecular graph represent the 3D coordinates of atoms, we are interested in
additional equivariance w.r.t. the Euclidean group E(3) or rigid transformations. An E(3)-equivariant
GNN (EGNN) satisfies EGNN(ΠXA + b) = Π EGNN(X)A + b for an orthogonal 3 × 3 matrix A
with A⊤ A = I and some translation vector b added row-wise.

In our case, since the nodes have both geometric atomic coordinates x as well as atomic type features
h, we can use a simple implementation of EGNN proposed by Satorras et al. [54], in which the
updates for features h and coordinates x of node i at layer l are computed as follows:

mij = ϕe (hli , hlj , d2ij , aij ), ẽij = ϕatt (mij ) (3)

X
hl+1
i = ϕh (hli , ẽij mij ) (4)
j̸=i
X xli − xlj
xl+1
i = xli + ϕx (hli , hlj , d2ij , aij ) (5)
dij + 1
j̸=i

where ϕe , ϕatt , ϕh and ϕh are learnable Multi-layer Perceptrons (MLPs) and dij and aij are the
relative distances and edge features between nodes i and j respectively. Following [41], we do not
update the coordinates of nodes that belong to the pocket to ensure the three-dimensional protein
context remains fixed throughout the EGNN layers.

We can break the symmetry to reflections and thereby make the GNN layer SE(3)-equivariant by
adding a cross product-dependent term to the coordinate update, which changes sign under reflection:

X xli − xlj
xl+1
i = xli + ϕdx (hli , hlj , d2ij , aij ) (6)
dij + 1
j̸=i

(xli − x̄l ) × (xlj − x̄l )

+ ϕ× l l 2
x (hi , hj , dij , aij ). (7)
||(xli − x̄l ) × (xlj − x̄l )|| + 1

13
Here, x̄l denotes the center of mass of all nodes at layer l. ϕ×
x is an additional MLP. This modification
is discussed in more detail in Appendix E.

7.4 Inpainting

For molecular inpainting as shown in Figure 1C, a subset of all atoms is fixed and serves as the
molecular context we want to condition on. All other atoms are generated by the DDPM. To this
gen
end, we diffuse the fixed atoms at each time step and predict a new latent representation zt−1 with
the neural network. We then replace the generated atoms corresponding to fixed nodes with their
forward noised counterparts:
input
zt−1 ∼ q(zt−1 |zdata ) (8)
gen
zt−1 ∼ pθ (zt−1 |zt ) (9)
 input gen 
zt−1 = zt−1 , zt−1,i∈M
/ , (10)

where M denotes the set of mask indices used to uniquely identify nodes corresponding to fixed
atoms. In this manner, we traverse the Markov chain in reverse order from t = T to t = 0 to generate
conditional samples. Because the noise schedule decreases the noising process’s variance to almost
zero at t = 0 (Equation (1)), the final sample is guaranteed to contain an unperturbed representation
of the fixed atoms. This approach can be applied to pocket-conditioned ligand-inpainting by fixing
all pocket nodes when sampling from the joint distribution model. However, it is much more general
and allows us to mask and replace arbitrary parts of the ligand-pocket system without retraining,
and can also be combined with the conditionally trained model—an option we explore in Section 4.

Equivariance

Since the equivariant diffusion process is defined for a CoM-free system, we must ensure that this
requirement remains satisfied after the substitution step in Equation (10). To prevent a CoM shift,
we therefore translate the fixed atom representation so that its
P centerinput
of mass coincides with the
predicted representation: x̃input input 1 gen 1
P
t−1 = x t−1 + n x
i∈M t−1,i − n i∈M t−1,i before creating the new
x
input gen input input input
combined representation zt−1 = [z̃t−1 , zt−1,i∈M/ ] with z̃t−1 = [x̃t−1 , ht−1 ] and n = |M|.

Resampling

Trippe et al. [58] show that this simple replacement method inevitably introduces approximation error
that can lead to inconsistent inpainted regions. In our experiments, we observe that the inpainting
solution sometimes generates disconnected molecules that are not properly positioned in the target
pocket (see Figure C1 for an example). Trippe et al. [58] propose to address this limitation with a
particle filtering scheme that upweights more consistent samples in each denoising step. We, however,
choose to adopt the conceptually simpler idea of resampling [38], where each latent representation
is repeatedly diffused back and forth before advancing to the next time step as demonstrated in
Algorithm 1. This enables the model to harmonize its prediction for the generated part and the noisy
sample from the fixed part, which does not include any information about the generated part. We
choose r = 10 resamplings per denoising step for our experiments with DiffSBDD-joint based on
empirical results discussed in Appendix C.

7.5 Implementation details

Molecule size

As part of a sample’s overall likelihood, we compute the empirical joint distribution of ligand and
pocket nodes p(NL , NP ) observed in the training set and smooth it with a Gaussian filter (σ = 1).

14
Algorithm 1 Sampling with the replacement method and resampling iterations. r denotes the
number of resampling steps and M is a set of indices of all atoms we want to fix. Note that samples
from the generative process pθ (zt−1 |zt ) are assumed to be CoM-free.
Require: r, M
zT ∼ N (0, I)
for t = T, ..., 1 do
for k = 1, ..., r do
input
zt−1 ∼ q(zt−1 |zdata ) ▷ Sample known context
gen
zt−1 ∼ pθ (zt−1 |zt ) ▷ Sample generated part
x̃input input 1 gen 1
xinput
P P
t−1 = xt−1 + n i∈M xt−1,i − n i∈M t−1,i ▷ Adjust center of mass
input gen
zt−1 = [z̃t−1 , zt−1,i / ]
∈M ▷ Combine
if k < r then
zt ∼ q(zt |zt−1 ) ▷ Apply noise and repeat
end if
end for
end for
return z0

In the conditional generation scenario, we derive the distribution p(NL |NP ) and use it for likelihood
computations.

For sampling, we can either fix molecule sizes manually or sample the number of ligand nodes from
the same distribution given the number of nodes in the target pocket:

NL ∼ p(NL |NP ). (11)

For the experiments discussed in Section 3, we increase the mean size of sampled molecules by 5
(CrossDocked) and 10 (Binding MOAD) atoms, respectively, to approximately match the sizes of
molecules found in the test set. This modification makes the reported Vina scores more comparable
as the in silico docking score is highly correlated with the molecular size, which is demonstrated in
Figure F4. Average molecule sizes after applying the correction are shown in Table F4 together with
corresponding values for generated molecules from other methods.

Preprocessing

All molecules are expressed as graphs. The full atom model uses the same one hot encoding of atom
types for ligand and protein nodes. For the Cα only model the node features of the protein are set
as the one hot encoding of the amino acid type instead. We refrain from adding a categorical feature
for distinguishing between protein and ligand atoms and use two separate MLPs for embedding the
node features instead (see Figure 1F).

Noise schedule

We use the pre-defined polynomial noise schedule introduced in [24]:

 t 2
α̃t = 1 − , t = 0, ..., T (12)
T
α̃t 2
2


Following [24, 59], values of α̃t|s = α̃s are clipped between 0.001 and 1 for numerical stability
near t = T , and α̃t is recomputed as
t
Y
α̃t = α̃τ |τ −1 . (13)
τ =0

15
A tiny offset ϵ = 10−5 is used to avoid numerical problems at t = 0 defining the final noise schedule:

αt2 = (1 − 2ϵ) · α̃t2 + ϵ. (14)

Feature scaling

We scale the node type features h by a factor of 0.25 relative to the coordinates x which was
empirically found to improve model performance in previous work [24]. To train joint probability
models in the all-atom scenario, it was necessary to scale down the coordinates (and corresponding
distance cutoffs) by a factor of 0.2 instead in order to avoid introducing too many edges in the graph
near the end of the diffusion process at t = T .

Hyperparameters

Hyperparameters for all presented models are summarized in Table 3. Training takes about 2.5 h/3.8 h
(conditional/joint) per 100 epochs on a single NVIDIA V100 for Binding MOAD in the Cα scenario
and 11.5 h/14.7 h per 100 epochs with full atom pocket representation on two V100 GPUs. For
CrossDocked, 100 training epochs take approximately 6 h/8 h in the Cα case and 48 h/60 h per 100
epochs on a single NVIDIA A100 GPU with all atom pocket representation.

Table 3: DiffSBDD hyperparameters.

CrossDocked Binding MOAD
Cond Joint Cond (Cα ) Joint (Cα ) Cond Joint Cond (Cα ) Joint (Cα )
No. layers 5 5 6 6 6 6 5 5
Joint embedding dim. 32 32 128 128 128 128 32 32
Hidden dim. 128 128 256 256 192 192 128 128
Learning rate 10−3 10−3 10−3 10−3 5 · 10−4 5 · 10−4 5 · 10−4 5 · 10−4
Weight decay 10−12 10−12 10−12 10−12 10−12 10−12 10−12 10−12
Diffusion steps 500 500 500 500 500 500 500 500
Edges (ligand-ligand) fully connected fully connected fully connected fully connected fully connected fully connected fully connected fully connected
Edges (ligand-pocket) < 5 Å < 5 Å < 5 Å < 5 Å < 7 Å < 7 Å < 8 Å < 8 Å
Edges (pocket-pocket) < 5 Å < 5 Å < 5 Å < 5 Å < 4 Å < 4 Å < 8 Å < 8 Å
Epochs 1000 1000 1000 1000 1000 1000 1000 1000

Postprocessing

For postprocessing of generated molecules, we use a similar procedure as in [60]. Given a list of atom
types and coordinates, bonds are first added using OpenBabel [61]. We then use RDKit to sanitise
molecules and filter for the largest molecular fragment.

7.6 Experimental Set-up

Datasets

We use the CrossDocked dataset [26] with 100,000 high-quality protein-ligand pairs for training and
100 proteins for testing, following previous works [15, 60]. The data split was done by 30% sequence
identity using MMseqs2 [62].

We also evaluate our method on a curated dataset of experimentally determined complexed protein-
ligand structures from Binding MOAD [25]. We keep pockets with valid3 and moderately ‘drug-
like’ ligands with QED score > 0.3. We further discard small molecules that contain atom types
∈
/ {C, N, O, S, B, Br, Cl, P, I, F } as well as binding pockets with non-standard amino acids. We
define binding pockets as the set of residues that have any atom within 8 Å of any ligand atom.

3
as defined in [Link]

16
Ligand redundancy is reduced by randomly sampling at most 50 molecules with the same chemical
component identifier (3-letter-code). After removing corrupted entries that could not be processed,
40 344 training pairs and 130 testing pairs remain. A validation set of size 246 is used to monitor
estimated log-likelihoods during training. The split is made to ensure different sets do not contain
proteins from the same Enzyme Commission Number (EC Number) main class.

Since the ResGen baseline model did not successfully generate samples for 21 and 5 targets from the
CrossDocked and Binding MOAD test sets, respectively, our analyses are only performed for samples
from the remaining pockets.

Baselines

We select two recently published autoregressive deep-learning methods for structure-based drug
design. Pocket2Mol [15] and ResGen [30] are sequential schemes relying on graph representations of
the protein pocket and previously placed atoms to predict probabilities based on which new atoms
are added. They are currently state-of-the-art among this class of models. Similar methods (e.g. 3D-
SBDD [60] or GraphBP [17]) consistently under-performed these baselines in our initial experiments
and are therefore not included in the final comparison. For Pocket2Mol, we re-evaluate already gen-
erated ligands on the CrossDocked dataset kindly provided by the authors. All other results were
produced using the official implementations available online4 with default sampling parameters. Note
that, unlike DiffSBDD, we therefore sample for the Binding MOAD test set with Pocket2Mol and
ResGen models that have been trained on CrossDocked. Since these two sets overlap (30 test PDBs
from Binding MOAD are found in the CrossDocked training set), there is potential data leakage. In
practice, however, we do not observe significantly different results when these targets are excluded
from the analysis. We also attempted to train Pocket2Mol on Binding MOAD, but did not manage
to robustly train the model on this dataset due to instability during training. While we aimed to
sample 100 ligands per pocket for the results in Section 3, the exact number of available molecules
varies slightly due to the characteristics of the different methods (see Table F3).

Evaluation metrics

We employ widely-used metrics to assess the quality of our generated molecules [14, 15]: (1) Vina
Score is an empirical estimation of binding affinity between small molecules and their target pocket;
(2) QED is a simple quantitative estimation of drug-likeness combining several desirable molecular
properties [27]; (3) SA estimates synthetic accessibility, i.e. the difficulty of synthesis [28]; (4) Lipin-
ski measures how many rules in the Lipinski rule of five [63] are satisfied (in addition to the original
four rules we require 10 or fewer rotatable bonds); (5) Diversity is computed as the average pair-
wise dissimilarity (1 - Tanimoto similarity) between molecular fingerprints of all generated molecules
for each pocket; (6) Inference Time is the average sampling time per target. Roughly 100-130
molecules were sampled per target but the exact number varies between targets for Pocket2Mol and
ResGen. Chemical properties are calculated with RDKit [64]. Docking scores are obtained after local
minimization with an empirical force field using the GNINA implementation [33] or, if specified, after
redocking with QuickVina2 [65].

Supplementary Information. Description of method details and additional experimental results

are included in the Supplementary Information.

Code Availability. Our source codes are publicly available at [Link]

DiffSBDD. Model weights can be downloaded from Zenodo: [Link]

Data Availability. The subset of the CrossDocked dataset used in this study was curated in
a previous work and is available online: [Link]
data. The raw BindingMOAD data can be downloaded from [Link] We

4
[Link] [Link]

17
provide further instructions on how to process these data in our code repository: [Link]
arneschneuing/DiffSBDD. Sampled molecules are available on Zenodo: [Link]
8239058.

Acknowledgments. We thank Xingang Peng and Shitong Luo for providing us generated
molecules from Pocket2Mol. We thank Hannes Stärk and Joshua Southern for valuable feedback and
insightful discussions. This work was supported by the European Research Council (starting grant no.
716058), the Swiss National Science Foundation (grant no. 310030 188744), and Microsoft Research
AI4Science. Charles Harris is supported by the Cambridge Centre for AI in Medicine Studentship
which is in turn funded by AstraZeneca and GSK. Michael Bronstein is supported in part by ERC
Consolidator grant no. 724228 (LEMAN).

References
[1] Anderson, A.C.: The process of structure-based drug design. Chemistry & biology 10(9), 787–797
(2003)

[2] Lyne, P.D.: Structure-based virtual screening: an overview. Drug discovery today 7(20), 1047–
1055 (2002)

[3] Shoichet, B.K.: Virtual screening of chemical libraries. Nature 432(7019), 862–865 (2004)

[4] Irwin, J.J., Shoichet, B.K.: Zinc- a free database of commercially available compounds for virtual
screening. Journal of chemical information and modeling 45(1), 177–182 (2005)

[5] Ferreira, L.G., Dos Santos, R.N., Oliva, G., Andricopulo, A.D.: Molecular docking and structure-
based drug design strategies. Molecules 20(7), 13384–13421 (2015)

[6] Bronstein, M.M., Bruna, J., Cohen, T., Veličković, P.: Geometric deep learning: Grids, groups,
graphs, geodesics, and gauges. arXiv preprint arXiv:2104.13478 (2021)

[7] Atz, K., Grisoni, F., Schneider, G.: Geometric deep learning on molecular representations. Nature
Machine Intelligence 3(12), 1023–1032 (2021)

[8] Khakzad, H., Igashov, I., Schneuing, A., Goverde, C., Bronstein, M., Correia, B.: A new age in
protein design empowered by deep learning. Cell Systems 14(11), 925–939 (2023)

[9] Gaudelet, T., Day, B., Jamasb, A.R., Soman, J., Regep, C., Liu, G., Hayter, J.B.R., Vickers, R.,
Roberts, C., Tang, J., Roblin, D., Blundell, T.L., Bronstein, M.M., Taylor-King, J.P.: Utilizing
graph machine learning within drug discovery and development. Briefings in Bioinformatics
22(6) (2021) [Link]

[10] Lu, W., Wu, Q., Zhang, J., Rao, J., Li, C., Zheng, S.: Tankbind: Trigonometry-aware neural
networks for drug-protein binding structure prediction. bioRxiv (2022)

[11] Stärk, H., Ganea, O., Pattanaik, L., Barzilay, R., Jaakkola, T.: Equibind: Geometric deep learn-
ing for drug binding structure prediction. In: International Conference on Machine Learning,
pp. 20503–20521 (2022). PMLR

[12] Corso, G., Stärk, H., Jing, B., Barzilay, R., Jaakkola, T.: Diffdock: Diffusion steps, twists, and
turns for molecular docking. arXiv preprint arXiv:2210.01776 (2022)

[13] Ragoza, M., Masuda, T., Koes, D.R.: Generating 3d molecules conditional on receptor binding
sites with deep generative models. Chemical science 13(9), 2701–2713 (2022)

[14] Li, Y., Pei, J., Lai, L.: Structure-based de novo drug design using 3d deep generative models.
Chemical science 12(41), 13664–13675 (2021)

18
[15] Peng, X., Luo, S., Guan, J., Xie, Q., Peng, J., Ma, J.: Pocket2mol: Efficient molecular sampling
based on 3d protein pockets. arXiv preprint arXiv:2205.07249 (2022)

[16] Drotár, P., Jamasb, A.R., Day, B., Cangea, C., Liò, P.: Structure-aware generation of drug-like
molecules. arXiv preprint arXiv:2111.04107 (2021)

[17] Liu, M., Luo, Y., Uchino, K., Maruhashi, K., Ji, S.: Generating 3d molecules for target protein
binding. arXiv preprint arXiv:2204.09410 (2022)

[18] Guan, J., Qian, W.W., Peng, X., Su, Y., Peng, J., Ma, J.: 3d equivariant diffusion for target-
aware molecule generation and affinity prediction. arXiv preprint arXiv:2303.03543 (2023)

[19] Lin, H., Huang, Y., Liu, M., Li, X., Ji, S., Li, S.Z.: Diffbp: Generative diffusion of 3d molecules
for target protein binding. arXiv preprint arXiv:2211.11214 (2022)

[20] Guan, J., Zhou, X., Yang, Y., Bao, Y., Peng, J., Ma, J., Liu, Q., Wang, L., Gu, Q.: Decompdiff:
Diffusion models with decomposed priors for structure-based drug design (2023)

[21] Xu, M., Powers, A.S., Dror, R.O., Ermon, S., Leskovec, J.: Geometric latent diffusion models
for 3d molecule generation. In: International Conference on Machine Learning, pp. 38592–38610
(2023). PMLR

[22] Weiss, T., Mayo Yanes, E., Chakraborty, S., Cosmo, L., Bronstein, A.M., Gershoni-Poranne, R.:
Guided diffusion for inverse molecular design. Nature Computational Science, 1–10 (2023)

[23] Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. Advances in Neural
Information Processing Systems 33, 6840–6851 (2020)

[24] Hoogeboom, E., Satorras, V.G., Vignac, C., Welling, M.: Equivariant diffusion for molecule
generation in 3d. In: International Conference on Machine Learning, pp. 8867–8887 (2022).
PMLR

[25] Hu, L., Benson, M.L., Smith, R.D., Lerner, M.G., Carlson, H.A.: Binding moad (mother of all
databases). Proteins: Structure, Function, and Bioinformatics 60(3), 333–340 (2005)

[26] Francoeur, P.G., Masuda, T., Sunseri, J., Jia, A., Iovanisci, R.B., Snyder, I., Koes, D.R.: Three-
dimensional convolutional neural networks and a cross-docked data set for structure-based drug
design. Journal of Chemical Information and Modeling 60(9), 4200–4215 (2020)

[27] Bickerton, G.R., Paolini, G.V., Besnard, J., Muresan, S., Hopkins, A.L.: Quantifying the
chemical beauty of drugs. Nature chemistry 4(2), 90–98 (2012)

[28] Ertl, P., Schuffenhauer, A.: Estimation of synthetic accessibility score of drug-like molecules
based on molecular complexity and fragment contributions. Journal of cheminformatics 1(1),
1–11 (2009)

[29] Wildman, S.A., Crippen, G.M.: Prediction of physicochemical parameters by atomic contribu-
tions. Journal of chemical information and computer sciences 39(5), 868–873 (1999)

[30] Zhang, O., Zhang, J., Jin, J., Zhang, X., Hu, R., Shen, C., Cao, H., Du, H., Kang, Y., Deng,
Y., et al.: Resgen is a pocket-aware 3d molecular generation model based on parallel multiscale
modelling. Nature Machine Intelligence, 1–11 (2023)

[31] Peng, X., Luo, S., Guan, J., Xie, Q., Peng, J., Ma, J.: Pocket2Mol: Efficient Molecular Sampling
Based on 3D Protein Pockets. [Link]

[32] Zhang, O., Zhang, J., Jin, J., Zhang, X., Hu, R., Shen, C., Cao, H., Du, H., Kang, Y., Deng, Y.,
et al.: ResGen: A Pocket-aware 3D Molecular Generation Model Based on Parallel Multi-scale
Modeling. [Link]

[33] McNutt, A.T., Francoeur, P., Aggarwal, R., Masuda, T., Meli, R., Ragoza, M., Sunseri, J., Koes,

19
 D.R.: Gnina 1.0: molecular docking with deep learning. Journal of cheminformatics 13(1), 1–20
(2021)

[34] Chen, P., Tao, L., Wang, T., Zhang, J., He, A., Lam, K.-h., Liu, Z., He, X., Perry, K., Dong,
M., et al.: Structural basis for recognition of frizzled proteins by clostridium difficile toxin b.
Science 360(6389), 664–669 (2018)

[35] Ding, L.-C., Huang, X.-Y., Zheng, F.-F., Xie, J., She, L., Feng, Y., Su, B.-H., Zheng, D.-L.,
Lu, Y.-G.: Fzd2 inhibits the cell growth and migration of salivary adenoid cystic carcinomas.
Oncology Reports 35(2), 1006–1012 (2016)

[36] Ritchie, T.J., Macdonald, S.J.: The impact of aromatic ring count on compound developability–
are too many aromatic rings a liability in drug design? Drug discovery today 14(21-22), 1011–
1020 (2009)

[37] Song, Y., Sohl-Dickstein, J., Kingma, D.P., Kumar, A., Ermon, S., Poole, B.: Score-based
generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456
(2020)

[38] Lugmayr, A., Danelljan, M., Romero, A., Yu, F., Timofte, R., Van Gool, L.: Repaint: Inpainting
using denoising diffusion probabilistic models. In: Proceedings of the IEEE/CVF Conference on
Computer Vision and Pattern Recognition, pp. 11461–11471 (2022)

[39] Wang, J., Lisanza, S., Juergens, D., Tischer, D., Watson, J.L., Castro, K.M., Ragotte, R.,
Saragovi, A., Milles, L.F., Baek, M., et al.: Scaffolding protein functional sites using deep
learning. Science 377(6604), 387–394 (2022)

[40] Ho, J., Salimans, T., Gritsenko, A., Chan, W., Norouzi, M., Fleet, D.J.: Video diffusion models.
arXiv:2204.03458 (2022)

[41] Igashov, I., Stärk, H., Vignac, C., Schneuing, A., Satorras, V.G., Frossard, P., Welling, M., Bron-
stein, M., Correia, B.: Equivariant 3d-conditional diffusion model for molecular linker design.
Nature Machine Intelligence, 1–11 (2024)

[42] Kim, K.S., Lu, S., Cornelius, L.A., Lombardo, L.J., Borzilleri, R.M., Schroeder, G.M., Sheng,
C., Rovnyak, G., Crews, D., Schmidt, R.J., et al.: Synthesis and sar of pyrrolotriazine-4-one
based eg5 inhibitors. Bioorganic & medicinal chemistry letters 16(15), 3937–3942 (2006)

[43] Barone, M., Müller, M., Chiha, S., Ren, J., Albat, D., Soicke, A., Dohmen, S., Klein, M., Bruns,
J., Dinther, M., et al.: Designed nanomolar small-molecule inhibitors of ena/vasp evh1 inter-
action impair invasion and extravasation of breast cancer cells. Proceedings of the National
Academy of Sciences 117(47), 29684–29690 (2020)

[44] Gahbauer, S., Correy, G.J., Schuller, M., Ferla, M.P., Doruk, Y.U., Rachman, M., Wu, T.,
Diolaiti, M., Wang, S., Neitz, R.J., et al.: Iterative computational design and crystallographic
screening identifies potent inhibitors targeting the nsp3 macrodomain of sars-cov-2. Proceedings
of the National Academy of Sciences 120(2), 2212931120 (2023)

[45] Schuller, M., Correy, G.J., Gahbauer, S., Fearon, D., Wu, T., Dı́az, R.E., Young, I.D., Car-
valho Martins, L., Smith, D.H., Schulze-Gahmen, U., et al.: Fragment binding to the nsp3
macrodomain of sars-cov-2 identified through crystallographic screening and computational
docking. Science advances 7(16), 8711 (2021)

[46] Imrie, F., Bradley, A.R., Schaar, M., Deane, C.M.: Deep generative models for 3d linker design.
Journal of chemical information and modeling 60(4), 1983–1995 (2020)

[47] Böhm, H.-J., Flohr, A., Stahl, M.: Scaffold hopping. Drug discovery today: Technologies 1(3),
217–224 (2004)

[48] Luo, S., Su, Y., Peng, X., Wang, S., Peng, J., Ma, J.: Antigen-specific antibody design and

20
 optimization with diffusion-based generative models. bioRxiv (2022)

[49] Sorrell, F.J., Szklarz, M., Azeez, K.R.A., Elkins, J.M., Knapp, S.: Family-wide structural analysis
of human numb-associated protein kinases. Structure 24(3), 401–411 (2016)

[50] Debreczeni, J., Eswaran, J., Bullock, A., Filippakopoulos, P., Kavanagh, K., Amos, A., Fedorov,
O., Sobott, F., Ball, L., Von Delft, F., et al.: Crystal structure of the human serine-threonine
kinase 16 in complex with staurosporine [internet]. RCSB PDB Protein Data Bank (2005)

[51] Sohl-Dickstein, J., Weiss, E., Maheswaranathan, N., Ganguli, S.: Deep unsupervised learning
using nonequilibrium thermodynamics. In: International Conference on Machine Learning, pp.
2256–2265 (2015). PMLR

[52] Kingma, D., Salimans, T., Poole, B., Ho, J.: Variational diffusion models. Advances in neural
information processing systems 34, 21696–21707 (2021)

[53] Lepola, U., Wade, A., Andersen, H.F.: Do equivalent doses of escitalopram and citalopram have
similar efficacy? a pooled analysis of two positive placebo-controlled studies in major depressive
disorder. International clinical psychopharmacology 19(3), 149–155 (2004)

[54] Satorras, V.G., Hoogeboom, E., Welling, M.: E (n) equivariant graph neural networks. In:
International Conference on Machine Learning, pp. 9323–9332 (2021). PMLR

[55] Köhler, J., Klein, L., Noé, F.: Equivariant flows: exact likelihood generative learning for sym-
metric densities. In: International Conference on Machine Learning, pp. 5361–5370 (2020).
PMLR

[56] Xu, M., Yu, L., Song, Y., Shi, C., Ermon, S., Tang, J.: Geodiff: A geometric diffusion model for
molecular conformation generation. arXiv preprint arXiv:2203.02923 (2022)

[57] Serre, J.-P., et al.: Linear Representations of Finite Groups vol. 42. Springer, New York (1977)

[58] Trippe, B.L., Yim, J., Tischer, D., Broderick, T., Baker, D., Barzilay, R., Jaakkola, T.: Diffu-
sion probabilistic modeling of protein backbones in 3d for the motif-scaffolding problem. arXiv
preprint arXiv:2206.04119 (2022)

[59] Nichol, A.Q., Dhariwal, P.: Improved denoising diffusion probabilistic models. In: International
Conference on Machine Learning, pp. 8162–8171 (2021). PMLR

[60] Luo, S., Guan, J., Ma, J., Peng, J.: A 3d generative model for structure-based drug design.
Advances in Neural Information Processing Systems 34, 6229–6239 (2021)

[61] O’Boyle, N.M., Banck, M., James, C.A., Morley, C., Vandermeersch, T., Hutchison, G.R.: Open
babel: An open chemical toolbox. Journal of cheminformatics 3(1), 1–14 (2011)

[62] Steinegger, M., Söding, J.: MMseqs2 enables sensitive protein sequence searching for the analysis
of massive data sets. Nature Biotechnology 35(11), 1026–1028 (2017) [Link]
nbt.3988

[63] Lipinski, C.A., Lombardo, F., Dominy, B.W., Feeney, P.J.: Experimental and computational
approaches to estimate solubility and permeability in drug discovery and development settings.
Advanced drug delivery reviews 64, 4–17 (2012)

[64] Landrum, G., et al.: Rdkit: Open-source cheminformatics software (2016)

[65] Alhossary, A., Handoko, S.D., Mu, Y., Kwoh, C.-K.: Fast, accurate, and reliable molecular
docking with quickvina 2. Bioinformatics 31(13), 2214–2216 (2015)

[66] Adams, K., Pattanaik, L., Coley, C.W.: Learning 3d representations of molecular chirality with
invariance to bond rotations. arXiv preprint arXiv:2110.04383 (2021)

21
[67] Li, Q.: Application of fragment-based drug discovery to versatile targets. Frontiers in molecular
biosciences 7, 180 (2020)

[68] Gahbauer, S., Correy, G.J., Schuller, M., Ferla, M.P., Doruk, Y.U., Rachman, M., Wu, T.,
Diolaiti, M., Wang, S., Neitz, R.J., et al.: Iterative computational design and crystallographic
screening identifies potent inhibitors targeting the nsp3 macrodomain of sars-cov-2. Proceedings
of the National Academy of Sciences 120(2), 2212931120 (2023)

[69] Bemis, G.W., Murcko, M.A.: The properties of known drugs. 1. molecular frameworks. Journal
of medicinal chemistry 39(15), 2887–2893 (1996)

[70] Kong, Z., Ping, W., Huang, J., Zhao, K., Catanzaro, B.: Diffwave: A versatile diffusion model
for audio synthesis. In: International Conference on Learning Representations (2021)

[71] Luo, S., Hu, W.: Diffusion probabilistic models for 3d point cloud generation. In: Proceedings of
the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 2837–2845 (2021)

[72] Du, Y., Fu, T., Sun, J., Liu, S.: Molgensurvey: A systematic survey in machine learning models
for molecule design. arXiv preprint arXiv:2203.14500 (2022)

[73] Jing, B., Corso, G., Chang, J., Barzilay, R., Jaakkola, T.: Torsional diffusion for molecular
conformer generation. arXiv preprint arXiv:2206.01729 (2022)

[74] Anand, N., Achim, T.: Protein structure and sequence generation with equivariant denoising
diffusion probabilistic models. arXiv preprint arXiv:2205.15019 (2022)

[75] Kelley, L.A., Mezulis, S., Yates, C.M., Wass, M.N., Sternberg, M.J.: The phyre2 web portal for
protein modeling, prediction and analysis. Nature protocols 10(6), 845–858 (2015)

[76] Kalyaanamoorthy, S., Chen, Y.-P.P.: Structure-based drug design to augment hit discovery.
Drug discovery today 16(17-18), 831–839 (2011)

[77] Duvenaud, D.K., Maclaurin, D., Iparraguirre, J., Bombarell, R., Hirzel, T., Aspuru-Guzik, A.,
Adams, R.P.: Convolutional networks on graphs for learning molecular fingerprints. Advances
in neural information processing systems 28 (2015)

[78] Gilmer, J., Schoenholz, S.S., Riley, P.F., Vinyals, O., Dahl, G.E.: Neural message passing for
quantum chemistry. In: International Conference on Machine Learning, pp. 1263–1272 (2017).
PMLR

[79] Lapchevskyi, K., Miller, B., Geiger, M., Smidt, T.: Euclidean neural networks (e3nn) v1. 0.
Technical report, Lawrence Berkeley National Lab.(LBNL), Berkeley, CA (United States) (2020)

[80] Du, W., Zhang, H., Du, Y., Meng, Q., Chen, W., Zheng, N., Shao, B., Liu, T.-Y.: Se (3)
equivariant graph neural networks with complete local frames. In: International Conference on
Machine Learning, pp. 5583–5608 (2022). PMLR

[81] Schütt, K.T., Sauceda, H.E., Kindermans, P.-J., Tkatchenko, A., Müller, K.-R.: Schnet–a deep
learning architecture for molecules and materials. The Journal of Chemical Physics 148(24),
241722 (2018)

[82] Klicpera, J., Groß, J., Günnemann, S.: Directional message passing for molecular graphs. arXiv
preprint arXiv:2003.03123 (2020)

[83] Duan, C., Du, Y., Jia, H., Kulik, H.J.: Accurate transition state generation with an object-
aware equivariant elementary reaction diffusion model. Nature Computational Science 3(12),
1045–1055 (2023)

[84] Batzner, S., Musaelian, A., Sun, L., Geiger, M., Mailoa, J.P., Kornbluth, M., Molinari, N., Smidt,
T.E., Kozinsky, B.: E (3)-equivariant graph neural networks for data-efficient and accurate
interatomic potentials. Nature communications 13(1), 1–11 (2022)

22
[85] Holdijk, L., Du, Y., Hooft, F., Jaini, P., Ensing, B., Welling, M.: Path integral stochastic optimal
control for sampling transition paths. arXiv preprint arXiv:2207.02149 (2022)

23
Appendix A Note on Variational Lower Bound

To maximise the likelihood of our training data, we aim at optimising the variational lower bound
(VLB) [24, 52]

T
X

 
− log p(zdata ) ≤ DKL q(zT |zdata )||p(zT ) −Eq(z0 |zdata ) log p(zdata |z0 ) + Lt (A1)
t=1
| {z }| {z }
prior loss Lprior reconstruction loss L0 | {z }
diffusion loss

with


Lt = DKL q(zt−1 |zdata , zt )||pθ (zt−1 |ẑdata , zt ) (A2)
h 1  SNR(t − 1)  i
= Eϵ∼N (0,I) − 1 ||ϵ − ϵ̂θ ||2 (A3)
2 SNR(t)

during training. The prior loss should always be close to zero and can be computed exactly in closed
form while the reconstruction loss must be estimated as described in Hoogeboom et al. [24]. In
practice, however, we simply minimise the mean squared error Ltrain = 21 ||ϵ − ϵ̂||2 while randomly
sampling time steps t ∼ U(0, . . . , T ), which is equivalent up to a multiplicative factor.

Appendix B Note on Equivariance of the Conditional Model

The 3D-conditional model can achieve equivariance without the usual “subspace-trick”. The coor-
dinates of pocket nodes provide a reference frame for all samples that can be used to translate
P (P )
them to a unique location (e.g. such that the pocket is centered at the origin: i xi = 0). By
doing this for all training data, translation equivariance becomes irrelevant and the CoM-free sub-
space approach obsolete. To evaluate the likelihood of translated samples at inference time, we
can first subtract the pocket’s center of mass from the whole system and compute the likelihood
after this mapping. Similarly, for sampling molecules we can first generate a ligand in a CoM-
free version of the pocket and move the whole system back to the original location of the pocket
nodes to restore translation equivariance. As long as the mean of our Gaussian noise distribu-
(P ) (P )
tion p(zt |zdata ) = N (µ(zdata ), σ 2 I) depends equivariantly on the pocket node coordinates x(P ) ,
O(3)-equivariance is satisfied as well (Appendix D). Since this change did not seem to affect the
performance of the conditional model in our experiments, we decided to keep sampling in the linear
subspace to ensure that the implementation is as similar as possible to the joint model, for which
the subspace approach is necessary.

Appendix C Resampling

Here, we briefly recapitulate the resampling algorithm introduced in Ref. [38]. The key intuition is
that inpainting with the replacement method combines a generated part with an independently sam-
pled latent representation of the known part. Even though the neural network tries to reconcile these
two components in every step of the denoising trajectory, it cannot succeed because the same issue
reoccurs in the following step. Lugmayr et al. [38] thus propose to apply the neural network several
times before proceeding to the next noise level, allowing the DDPM to preserve more conditional
information and move the sample closer to the data distribution again.

24
Table C1: Evaluation of generated molecules for target pockets from the CrossDocked (C.D.) and
Binding MOAD (B.M.) test sets with the inpainting approach and Cα pocket representation for
varying numbers of resampling steps r and denoising steps T . Here, the SA scores were mapped to
the unit interval using SAnorm = (10 − SA)/9.
r T Vina Score (kcal/mol, ↓) QED (↑) SAnorm (↑) Lipinski (↑) Diversity (↑) Time (s, ↓)
1 500 −5.830 ± 2.47 0.403 ± 0.18 0.552 ± 0.13 4.620 ± 0.81 0.808 ± 0.06 97.434 ± 39.79
C.D.

5 100 −6.872 ± 2.43 0.444 ± 0.19 0.551 ± 0.12 4.654 ± 0.72 0.766 ± 0.06 96.205 ± 39.22
10 50 −7.177 ± 3.28 0.556 ± 0.20 0.729 ± 0.12 4.742 ± 0.59 0.718 ± 0.07 94.481 ± 38.86
1 500 −5.810 ± 2.00 0.468 ± 0.16 0.627 ± 0.14 4.839 ± 0.49 0.851 ± 0.04 40.298 ± 13.52
B.M.

5 100 −6.082 ± 2.01 0.537 ± 0.16 0.701 ± 0.13 4.924 ± 0.31 0.855 ± 0.05 45.074 ± 21.14
10 50 −6.192 ± 2.24 0.560 ± 0.16 0.737 ± 0.13 4.941 ± 0.27 0.859 ± 0.05 41.490 ± 14.32

12

10

8
CoM RMSD (Å)

l
na 1) 5) 10
)
itio g(
r=
g(
r= r=
nd tin tin g(
co tin
ain ain ain
inp inp inp

Fig. C1: (A) Example of a generated molecule (green) without additional resampling steps and the
reference molecule (magenta) from the target PDB 5ncf. The generated molecule is not placed in
the target pocket but in the protein core. (B) RMSD between reference molecules’ center of mass
and generated molecules’ center of mass for the conditional model and inpaining model with varying
numbers of resampling steps r. The pocket representation is Cα in all cases.

Number of resampling steps

To empirically study the effect of the number of resampling iterations applied, we generated ligands
for all test pockets with r = 1, r = 5, and r = 10 resampling steps, respectively. Because the resam-
pling strategy slows down sampling approximately by a factor of r, we used the striding technique
proposed by Nichol and Dhariwal [59] and reduced the number of denoising steps proportionally to
r. Nichol and Dhariwal [59] showed that this approach reduces the number of sampling steps signif-
icantly without sacrificing sample quality. In our case, it allows us to retain sampling speed while
increasing the number of resampling steps.

To gauge the effect of resampling for molecule generation we show the distribution of RMSD val-
ues between the center of mass of reference molecules and generated molecules in Figure C1. The
unmodified replacement method (r = 1) produces molecules that are clearly farther away from the
presumed pocket center than the conditional model. Increasing r moves the mean distance closer to
the average displacement of molecules from the conditional method. This effect seems to saturate at
r = 10 which is in line with the results obtained for images [38].

Table C1 shows that neither the additional resampling steps nor the shortened denoising trajectory
degrade the performance on the reported molecular metrics. The average docking scores even improve
slightly which might reflect better positioning of generated ligands in the pockets prior to docking.
The same model trained with T = 500 diffusion steps was used in all three cases.

25
Appendix D Proofs

In the following proofs we do not consider categorical node features h as only the positions x are
subject to equivariance constraints. Furthermore, we do not distinguish between the zeroth latent
representation x0 and data domain representations xdata for ease of notation, and simply drop the
subscripts.

D.1 O(3)-equivariance of the prior probability

(L)
The isotropic Gaussian prior p(xT |x(P ) ) = N (µ(x(P ) ), σ 2 I) is equivariant to rotations and
reflections represented by an orthogonal matrix R ∈ R3×3 as long as µ(Rx(P ) ) = Rµ(x(P ) ) because:

(L) 1  1 (L)

p(RxT |Rx(P ) ) = p exp − 2 ||RxT − µ(Rx(P ) )||2
(2π)NL σ 2 2σ
1  1 (L)

=p exp − 2 ||RxT − Rµ(x(P ) )||2
(2π)NL σ 2 2σ
1  1 (L)

=p exp − 2 ||R xT − µ(x(P ) ) ||2
(2π)NL σ 2 2σ
1  1 (L)

=p exp − 2 ||xT − µ(x(P ) )||2
(2π)NL σ 2 2σ
(L)
= p(xT |x(P ) ).

Here we used ||Rx||2 = ||x||2 for orthogonal R.

D.2 O(3)-equivariance of the transition probabilities

The denoising transition probabilities from time step t to s < t are defined as isotropic normal
distributions:
(L) (L) (L) (L)
pθ (xt−1 |xt , x̂(L) , x(P ) ) = N (xt−1 |µt→s (xt , x̂(L) , x(P ) ), σt→s
2
I). (D4)

(L) (L)
Therefore, pθ (xt−1 |xt , x̂(L) , x(P ) ) is O(3)-equivariant by a similar argument to Section D.1 if µt→s
is computed equivariantly from the three-dimensional context.
2
αt|s σs2 (L) αs σt|s
Recalling the definition of µt→s = σt2
xt + σt2
x̂(L) , we can prove its equivariance as follows:

2
(L) αt|s σs2 (L)
αs σt|s (L)
µt→s (Rxt , Rx(P ) ) = 2 Rx t + 2 x̂(L) (Rxt , Rx(P ) )
σt σt
2
αt|s σs2 (L)
αs σt|s (L)
= Rx t + Rx̂(L) (xt , x(P ) ) (equivariance of x̂(L) )
σt2 σt2
 α σ2 2
t|s s (L) αs σt|s (L) (L) (P )

=R x t + x̂ (x t , x )
σt2 σt2
(L)
= Rµt→s (xt , x(P ) ),

1 (L) σt
where x̂(L) defined as x̂(L) = αt xt − αt ϵ̂ is equivariant because:

(L) 1 (L) σt (L)

x̂(L) (Rxt , Rx(P ) ) = Rxt − ϵ̂(Rxt , Rx(P ) , t)
αt αt

26
 1 (L) σt (L)
= Rxt − Rϵ̂(xt , x(P ) , t) (ϵ̂ predicted by equivariant neural network)
αt αt
1 σt 
(L) (L)
=R xt − ϵ̂(xt , x(P ) , t)
αt αt
(L)
= Rx̂(L) (xt , x(P ) ).

D.3 O(3)-equivariance of the learned likelihood

Let R ∈ R3×3 be an orthogonal matrix representing an element g from the general orthogonal group
O(3). We obtain the marginal probability density of the Markovian denoising process as follows
Z
(L) (L) (L) (L)
pθ (x0 |x(P ) ) = p(xT |x(P ) )pθ (x0:T −1 |xT , x(P ) )dx1:T
Z T
(L) (L) (L)
Y
= p(xT |x(P ) ) pθ (xt−1 |xt , x(P ) )dx1:T
t=1

and the sample’s likelihood is O(3)-equivariant:

Z T
(L) (L) (L) (L)
Y
pθ (Rx0 |Rx(P ) ) = p(RxT |Rx(P ) ) pθ (Rxt−1 |Rxt , Rx(P ) )dx1:T
t=1
Z T
(L) (L) (L)
Y
= p(xT |x(P ) ) pθ (Rxt−1 |Rxt , Rx(P ) )dx1:T (equivariant prior)
t=1
Z T
(L) (L) (L)
Y
= p(xT |x(P ) ) pθ (xt−1 |xt , x(P ) )dx1:T (equivariant transition probabilities)
t=1
(L)
= pθ (x0 |x(P ) )

Appendix E SE(3)-equivariant Graph Neural Network

Chiral molecules cannot be superimposed by any combination of rotations and translations. Instead
they are mirrored along a stereocenter, axis, or plane. As chirality can significantly alter a molecule’s
chemical properties, we use a variant of the E(3)-equivariant graph neural networks [54] presented
in Equations (3)-(5) that is sensitive to reflections and hence SE(3)-equivariant. We change the
coordinate update equation, Equ. (5), of standard EGNNs in the following way

X xli − xlj (xli − x̄l ) × (xlj − x̄l )

xl+1
i = xli + ϕdx (hli , hlj , d2ij , aij ) + ϕ× l l 2
x (hi , hj , dij , aij ), (E5)
j̸=i
dij + 1 ||(xli − x̄l ) × (xlj − x̄l )|| + 1

where x̄l denotes the center of mass of all nodes at layer l. This modification makes the EGNN layer
sensitive to reflections while staying close to the original formalism. Since the resulting graph neural
networks are only equivariant to the SE(3) group, we will hereafter call them SE(3)GNNs for short.

E.1 Discussion of Equivariance

Here we study how the suggested change in the coordinate update equation breaks reflection symme-
try while preserving equivariance to rotations. Messages and scalar feature updates (Equations (3)
and (4)) remain E(3)-invariant as in the original model and are therefore not considered in this
section. We analyze transformations composed of a translation by t ∈ R3 and a rotation/reflection

27
by an orthogonal matrix R ∈ R3×3 with RT R = I. The output at layer l + 1 given the transformed
input Rxli + t at layer l is calculated as:
X Rxli + t − (Rxlj + t) d (Rxli + t − (Rx̄l + t)) × (Rxlj + t − (Rx̄l + t)) ×
Rxli + t + ϕx (·) + × ϕx (·) (E6)
dij + 1 Zij +1
j̸=i
X R(xli − xlj ) d (Rxli − Rx̄l ) × (Rxlj − Rx̄l ) ×
= Rxli + t + ϕx (·) + × ϕx (·) (E7)
dij + 1 Zij +1
j̸=i
X R(xli − xlj ) d det(R)R (xli − x̄l ) × (xlj − x̄l ) ×


l
= Rxi + t + ϕx (·) + × ϕx (·) (E8)
dij + 1 Zij +1
j̸=i

X R (xli − x̄l ) × (xlj − x̄l )

l+1
= Rxi + t + det(R) − 1 × . (E9)
j̸=i
Zij +1

This result shows that the output coordinates are only equivariantly transformed if R is orientation
preserving, i.e. det(R) = 1. If R is a reflection (det(R) = −1), coordinates will be updated with an
additional summand that breaks the symmetry.

The learnable coefficients ϕdx (hli , hlj , d2ij , aij ) and ϕ× l l 2

x (hi , hj , dij , aij ) only depend on relative distances
and are therefore E(3)-invariant. Their arguments are represented with the “·” symbol for brevity.
×
Likewise, the normalization factor ||(xli − x̄l ) × (xlj − x̄l )|| is abbreviated as Zij . Already in the
first line we used the fact that the mean transforms equivariantly. Furthermore, we use Ra × Rb =
det(R)R(a × b) in the second step, which can be derived as follows:

xT (Ra × Rb) = det([x, Ra, Rb]) (E10)

| {z }
∈R3×3

= det(R[RT x, a, b]) (E11)

T
= det(R) det([R x, a, b]) (E12)
 
= det(R) xT R(a × b) (E13)
 
= xT det(R)R(a × b) (E14)

The stated property of the cross product follows because this derivation is true for all x ∈ R3 .

E.2 Empirical Results

To show the effectiveness of this architecture

on a simple toy example, we repeat the clas- Table E2: Accuracy on the R/S classification
sification experiment by Adams et al. [66] task. Results in the first section are taken from [66]
who train neural networks to classify tetra- and included for reference.
hedral chiral centers as right-handed (rectus,
Model R/S Accuracy (%)
‘R’) or left-handed (sinister, ‘S’). We closely
follow their data split and experimental set- ChIRo 98.5
up and only replace the classifier with EGNN SchNet 54.4
and SE(3)GNNs, respectively. The results in DimeNet++ 65.7
Table E2 clearly demonstrate that the SE(3)- SphereNet 98.2
equivariant EGNN is capable of solving this EGNN 50.4
task (without any hyperparameter optimiza- SE(3)GNN 83.4
tion) whereas the E(3)-equivariant version does
not do better than random guessing.

28
 1.0
8
0.8

6
QED score
0.6

SA score
0.4
4
0.2
2
0.0

Reference DiffSBDD-cond DiffSBDD-joint Pocket2Mol ResGen Reference DiffSBDD-cond DiffSBDD-joint Pocket2Mol ResGen
Model Model
15
6
10
5
5

Lipinski
LogP

0 4

−5
3
−10
2
Reference DiffSBDD-cond DiffSBDD-joint Pocket2Mol ResGen Reference DiffSBDD-cond DiffSBDD-joint Pocket2Mol ResGen
Model Model
5

0 8
CNN affinity
Vina score

−5 6

−10
4

−15
2

−20
Reference DiffSBDD-cond DiffSBDD-joint Pocket2Mol ResGen Reference DiffSBDD-cond DiffSBDD-joint Pocket2Mol ResGen
Model Model

Fig. F2: Distributions of computational scores for generated molecules and reference ligands from
the CrossDocked test set.

Appendix F Extended results

F.1 Sampling statistics

Table F3 summarizes the number of generated molecules available for the analyses in Section 3.

Table F3: Average number of molecules per

test set pocket.
Method CrossDocked Binding MOAD
Test set 1 1
Pocket2Mol [15] 98.01 132.14
ResGen [30] 114.41 109.83
DiffSBDD-cond 100 100
DiffSBDD-joint 100 100

F.2 Distributions of Molecular Properties

The distributions used to compute the Wasserstein distances in Table 1 are visualized in Figures F2
and F3.

29
 1.0
8
0.8
6
QED score

SA score
0.6

0.4 4

0.2
2
0.0
Reference DiffSBDD-cond DiffSBDD-joint Pocket2Mol ResGen Reference DiffSBDD-cond DiffSBDD-joint Pocket2Mol ResGen
Model Model

10 5

Lipinski
4
LogP

0
3
−5

2
Reference DiffSBDD-cond DiffSBDD-joint Pocket2Mol ResGen Reference DiffSBDD-cond DiffSBDD-joint Pocket2Mol ResGen
Model Model
5 10

0 8
CNN affinity
Vina score

−5
6

−10
4

−15
2
−20
Reference DiffSBDD-cond DiffSBDD-joint Pocket2Mol ResGen Reference DiffSBDD-cond DiffSBDD-joint Pocket2Mol ResGen
Model Model
Fig. F3: Distributions of computational scores for generated molecules and reference ligands from
the Binding MOAD test set.

A −4
B −2

−5 −4

−6
−6
−7
QuickVina score (kcal/mol)

QuickVina score (kcal/mol)

−8
−8

−10
−9

−10 −12

−11
−14

−12
−16
10 20 30 40 50 10 20 30 40 50
Ligand size Ligand size

Fig. F4: Correlation between ligand size and QuickVina score for reference molecules from the
Binding MOAD (A) and CrossDocked (B) test sets.

F.3 Dependence of Vina scores on molecule size

Figure F4 shows how strongly the empirical Vina score is correlated with the number of heavy atoms
in the ligands. For this analysis, we used Vina scores computed by the QuickVina2 [65] software
after re-docking. Since we want to match the distribution of scores of reference molecules as closely
as possible with our generated molecules, we expect that their sizes should roughly match as well.
However, our diffusion model operates on point clouds with fixed sizes, which we determine at the
beginning of sampling as explained in Section 7.5. By biasing the procedure, we match the sizes of
reference ligands more closely as shown in Table F4.

30
 Table F4: Average number of heavy atoms of
generated molecules.
Method CrossDocked Binding MOAD
Test set 23.9 28.0
Pocket2Mol [15] 19.0 16.8
ResGen [30] 16.2 18.4
DiffSBDD-cond 24.9 24.5
DiffSBDD-joint 24.5 25.2

F.4 Additional Molecular Metrics

In addition to the molecular properties discussed in Section 7.6 we assess the models’ ability to
produce novel and valid molecules using four simple metrics: validity, connectivity, uniqueness, and
novelty. Validity measures the proportion of generated molecules that pass basic tests by RDKit–
mostly ensuring correct valencies. Connectivity is the proportion of valid molecules that do not
contain any disconnected fragments. We convert every valid and connected molecule from a graph
into a canonical SMILES string representation, count the number unique occurrences in the set of
generated molecules and compare those to the training set SMILES to compute uniqueness and
novelty respectively.

Table F5 shows that only a small fraction of all generated molecules is invalid and must be discarded
for downstream processing. A much larger percentage of molecules is fragmented but, since we can
simply select and process the largest fragments in these cases, low connectivity does not necessarily
affect the efficiency of the generative process. Moreover, all models produce diverse sets of molecules
unseen in the training set.

Table F5: Basic molecular metrics for generated small molecules given a Cα and full atom repre-
sentation of the protein pocket.
Model Validity Connectivity Uniqueness Novelty
CrossDocked test set 100% 100% 96% 96.88%
DiffSBDD-cond (Cα ) 95.52% 79.52% 99.99% 99.97%
DiffSBDD-inpaint (Cα ) 99.18% 98.25% 99.52% 99.97%
DiffSBDD-cond 97.10% 78.27% 99.98% 99.99%
DiffSBDD-inpaint 92.99% 67.52% 100% 100%
Binding MOAD test set 97.69% 100% 38.58% 77.55%
DiffSBDD-cond (Cα ) 94.41% 77.38% 100% 100%
DiffSBDD-inpaint (Cα ) 98.36% 91.60% 99.99% 99.98%
DiffSBDD-cond 96.32% 63.37% 100% 100%
DiffSBDD-inpaint 93.88% 75.60% 100% 100%

F.5 Results with a coarse-grained pocket representation

Here we discuss the advantages and disadvantages of coarse-grained Cα protein representations as

context for the generative model. Table F6 shows that full-atom models outperform their coarse-
grained counterparts on the Vina metric, which is the only reported metric that captures interactions
with the protein. Ligand-centric metrics do not seem to depend on the protein representation as could
be expected. The main advantage of the Cα models is their significantly faster training and inference
time, a fact we made use of during model development for fine-tuning and preliminary analyses.

To further demonstrate the limitations of the coarse-grained models, we compare the generated raw
conformations to the best scoring QuickVina docking pose after re-docking and plot the distribution

31
Table F6: Evaluation of generated molecules for targets from the CrossDocked and Binding MOAD
test sets. We compare all-atom and coarse-grained (Cα ) pocket representations. Note that the SA
score has been normalized so that higher is better. Here, the SA scores were mapped to the unit
interval using SAnorm = (10 − SA)/9.
Vina (All) (↓) Vina (Top-10%) (↓) QED (↑) SAnorm (↑) Lipinski (↑) Diversity (↑) Time (s, ↓)
Test set −6.871 ± 2.32 — 0.476 ± 0.20 0.728 ± 0.14 4.340 ± 1.14 — —
DiffSBDD-cond (Cα ) −6.770 ± 2.73 −8.796 ± 1.75 0.475 ± 0.22 0.612 ± 0.12 4.536 ± 0.91 0.725 ± 0.06 49.651 ± 17.34
C.D.

DiffSBDD-inpaint (Cα ) −7.177 ± 3.28 −9.233 ± 1.82 0.556 ± 0.20 0.729 ± 0.12 4.742 ± 0.59 0.718 ± 0.07 94.481 ± 38.86
DiffSBDD-cond −6.950 ± 2.06 −9.120 ± 2.16 0.469 ± 0.21 0.578 ± 0.13 4.562 ± 0.89 0.728 ± 0.07 135.866 ± 51.66
DiffSBDD-inpaint −7.333 ± 2.56 −9.927 ± 2.59 0.467 ± 0.18 0.554 ± 0.12 4.702 ± 0.64 0.758 ± 0.05 160.314 ± 73.30
Test set −8.412 ± 2.03 — 0.522 ± 0.17 0.692 ± 0.12 4.669 ± 0.49 — —
DiffSBDD-cond (Cα ) −6.863 ± 1.59 −8.587 ± 1.34 0.480 ± 0.20 0.554 ± 0.11 4.662 ± 0.68 0.714 ± 0.05 36.285 ± 8.13
B.M.

DiffSBDD-inpaint (Cα ) −6.926 ± 3.39 −9.124 ± 1.35 0.548 ± 0.19 0.580 ± 0.13 4.757 ± 0.51 0.709 ± 0.05 58.305 ± 17.35
DiffSBDD-cond −7.171 ± 1.89 −9.184 ± 2.23 0.436 ± 0.20 0.568 ± 0.12 4.542 ± 0.79 0.714 ± 0.08 336.061 ± 85.02
DiffSBDD-inpaint −7.309 ± 4.03 −9.840 ± 2.18 0.542 ± 0.21 0.615 ± 0.12 4.777 ± 0.53 0.739 ± 0.05 369.873 ± 124.54

CrossDocked
A C ond. C α
B 0.200
Joint C α
Cond. full-atom Joint full-atom
0.20 RDKit conf. RDKit conf.
0.175

0.150
0.15
0.125
Density

Density

0.100
0.10

0.075

0.050
0.05

0.025

0.00 0.000
0 2 4 6 8 10 12 14 0 2 4 6 8 10 12 14 16
RMSD (Å) RMSD (Å)

Binding MOAD
C C ond. C α
D 0.200
Joint C α
0.200
Cond. full-atom Joint full-atom
0.175
RDKit conf. RDKit conf.
0.175
0.150
0.150
0.125
0.125
Density

Density

0.100
0.100

0.075
0.075

0.050 0.050

0.025 0.025

0.000 0.000
0 2 4 6 8 10 12 14 16 0.0 2.5 5.0 7.5 10.0 12.5 15.0 17.5
RMSD (Å) RMSD (Å)

Fig. F5: RMSD between generated and docked conformations for the CrossDocked (A, B) and
Binding MOAD (C, D) datasets. Full-atom models are compared to Cα models as well as a baseline
of random RDKit conformers of the molecules generated by the Cα -model. (A, C) DiffSBDD-cond.
(B) DiffSBDD-joint.

of resulting RMSD values in Figure F5. As a baseline, the procedure is repeated for RDKit conformers
of the same molecules with identical center of mass. For a large percentage of molecules generated
by the all-atom models, QuickVina agrees with the predicted bound conformations, leaving them
almost unchanged (RMSD below 2 Å). This demonstrates successful conditioning on the geometry
of the given protein pockets. For the Cα -only models results are less convincing. They produce poses
that barely improve upon conformers lacking pocket-context. Likely, this is caused by atomic clashes
with the proteins’ side chains that QuickVina needs to resolve.

32
 DiffSBDD-cond DiffSBDD-joint
CrossDocked
Binding MOAD

Fig. F6: Randomly selected samples of generated molecules.

F.6 Random generated molecules

Randomly selected molecules generated with different models are presented in Figure F6.

Appendix G Inpainting implementation

G.1 Implementation of fragment merging experiments

Fragment merging is the task of combining fragments with an overlapping binding site [67]. For
this example, instead of masking existing molecules, we instead take fragments from an experimen-
tal fragment screen against the non structural protein 3 (NSP3) from SARS-CoV-2 [45]. Using two
fragments as input (PDB entries 5rue and 5rsw) we successfully replicate the fragment merge per-
formed in Gahbauer et al. [68] which was accomplished using the chemoinformatics-based approach
Fragmenstein 5 . To accomplish this, instead of masking out and reinserting atoms, we instead choose

5
[Link]

33
 (a) Increasing resamplings (r)

Input r=1 r=2 r=5 r = 10

(b) (c) (d)

Fig. G7: (a) Importance of high resamplings. Effect of the number of resamplings on molecular
validity (b), connectivity (c) and uniqueness (d).

to fix all atoms during generation except the atom on each fragment cloest to the other. We need
perform t = 200 steps of the DiffSBDD-diversify procedure to allow the model to arrange the atom
positions as well as change the atom types. All PDB files were already structurally aligned.

G.2 Quantitative evaluation of inpainting for the whole Binding MOAD

test set

For all experiments across the whole test set, we perform automatic masking of atoms which are to be
fixed. For scaffold elaboration, we extract the Bemis-Murcko scaffold [69] using RDKit and compute
a binary mask to fix the scaffold, while functional groups are redesigned. For scaffold hopping, we
simply take the inverse of the mask used for scaffold elaboration. For linker design, we fragment each
molecule in the test set in multiple ways as in Igashov et al. [41], Imrie et al. [46]. To benchmark
against DiffLinker, we use the model weights and protocol as described in Igashov et al. [41] except
we give the ground-truth linker size as input, rather than predict it using the auxillary model, for
fairness. In small-scale experiments where finer control is desirable (e.g. as in the fragment merging
example described above), the binary mask can be defined manually.

Appendix H Optimization

We demonstrate the effect the number of noising/denoising steps (t) has on various molecular prop-
erties in Figure H8. We test all values of t at intervals of 10 steps and 200 molecules are sampled at
every timestep. Note this does not allow for explicit optimization of any particular property unless
combined with the evolutionary algorithm, as shown in Figure 1D (all plots are for PDB entry 5NDU
[43]).

34
 t t t t

Fig. H8: Effect of number of noising/denoising steps on molecule properties.

Appendix I Related Work

Diffusion Models for Molecules

Inspired by non-equilibrium thermodynamics, diffusion models have been proposed to learn data
distributions by modeling a denoising (reverse diffusion) process and have achieved remarkable success
in a variety of tasks such as image, audio synthesis and point cloud generation [52, 70, 71]. Recently,
efforts have been made to utilize diffusion models for molecule design [72]. Specifically, Hoogeboom
et al. [24] propose a diffusion model with an equivariant network that operates both on continuous
atomic coordinates and categorical atom types to generate new molecules in 3D space. Torsional
Diffusion [73] focuses on a conditional setting where molecular conformations (atomic coordinates)
are generated from molecular graphs (atom types and bonds). Similarly, 3D diffusion models have
been applied to generative design of larger biomolecular structures, such as antibodies [48] and other
proteins [58, 74].

Structure-based Drug Design

Structure-based Drug Design (SBDD) [1, 5] relies on the knowledge of the 3D structure of the
biological target obtained either through experimental methods or high-confidence predictions using
homology modelling [75]. Candidate molecules are then designed to bind with high affinity and
specificity to the target using interactive software [76] and often human-based intuition [5]. Recent
advances in deep generative models have brought a new wave of research that model the conditional
distribution of ligands given biological targets and thus enable de novo structure-based drug design.
Most of previous work consider this task as a sequential generation problem and design a variety of
generative methods including autoregressive models, reinforcement learning, etc., to generate ligands
inside protein pockets atom by atom [14–16, 60]. Most recent work explore the use of diffusion models
in structure-based drug design [18–20].

Geometric Deep Learning for Drug Discovery

Geometric deep learning refers to incorporating geometric priors in neural architecture design that
respects symmetry and invariance, thus reduces sample complexity and eliminates the need for data
augmentation [6]. It has been prevailing in a variety of drug discovery tasks from virtual screening
to de novo drug design as symmetry widely exists in the representation of drugs. One line of work
introduces graph and geometry priors and designs message passing neural networks and equivariant
neural networks that are permutation-, translation-, rotation-, and reflection-equivariant, respec-
tively [54, 77–80]. These architectures have been widely used in representing biomolecules from small
molecules to proteins [7] and solving downstream tasks such as molecular property prediction [81, 82],
binding pose prediction [11], transition state sampling [83], and molecular dynamics [84, 85]. Another
line of work focuses on generative design of new molecules [72]. More specifically, molecule design is

35
formulated as a graph or geometry generation problem, following either a one-shot generation strat-
egy that generates graphs (atom and bond features) in one step or attempting to generate atoms
and bonds sequentially.

36