Pattern Recognition Using Finite-Iteration Cellular Systems

Lecture Notes in Computer Science, 2018

Convolutional neural networks (CNN) have deeply impacted the field of machine learning. These networks, designed to process objects with a fixed topology, can readily be applied to images, videos and sounds but cannot be easily extended to structures with an arbitrary topology such as graphs. Examples of applications of machine learning to graphs include the prediction of the properties molecular graphs, or the classification of 3D meshes. Within the chemical graphs framework, we propose a method to extend networks based on a fixed topology to input graphs with an arbitrary topology. We also propose an enriched feature vector attached to each node of a chemical graph and a new layer interfacing graphs with arbitrary topologies with a full connected layer.

2020

The generation of graph-structured data is an emerging problem in the field of deep learning. Various solutions have been proposed in the last few years, yet the exploration of this branch is still in an early phase. In sequential approaches, the construction of a graph is the result of a sequence of decisions, in which, at each step, a node or a group of nodes is added to the graph, along with its connections. A very relevant application of graph generation methods is the discovery of new drug molecules, which are naturally represented as graphs. In this paper, we introduce a sequential molecular graph generator based on a set of graph neural network modules, which we call MG^2N^2. Its modular architecture simplifies the training procedure, also allowing an independent retraining of a single module. The use of graph neural networks maximizes the information in input at each generative step, which consists of the subgraph produced during the previous steps. Experiments of unconditio...

Natural Computing, 2009

Journal of Materials Informatics, 2023

Material molecular representation (MMR) plays an important role in material property or chemical reaction prediction. However, traditional expert-designed MMR methods face challenges in dealing with high dimensionality and heterogeneity of material data, leading to limited generalization capabilities and insufficient information representation. In recent years, graph neural networks (GNNs), a deep learning algorithm specifically designed for graph structures, have made inroads into the field of MMR. It will be instructive and inspiring to conduct a survey on various GNNs used for MMR. To achieve this objective, we compare GNNs with conventional MMR methods and illustrate the advantages of GNNs, such as their expressiveness and adaptability. In addition, we systematically classify and summarize the methods and applications of GNNs. Finally, we provide our insights into future research directions, taking into account the characteristics of molecular data and the inherent drawbacks of GNNs. This comprehensive survey is intended to present a holistic view of GNNs for MMR, focusing on the core concepts, the main techniques, and the future trends in this area.

arXiv (Cornell University), 2018

Functional groups (FGs) are molecular substructures that are served as a foundation for analyzing and predicting chemical properties of molecules. Automatic discovery of FGs will impact various fields of research, including medicinal chemistry and material sciences, by reducing the amount of lab experiments required for discovery or synthesis of new molecules. In this paper, we investigate methods based on graph convolutional neural networks (GCNNs) for localizing FGs that contribute to specific chemical properties of interest. In our framework, molecules are modeled as undirected relational graphs with atoms as nodes and bonds as edges. Using this relational graph structure, we trained GCNNs in a supervised way on experimentallyvalidated molecular training sets to predict specific chemical properties, e.g., toxicity. Upon learning a GCNN, we analyzed its activation patterns to automatically identify FGs using four different explainability methods that we have developed: gradient-based saliency maps, Class Activation Mapping (CAM), gradient-weighted CAM (Grad-CAM), and Excitation Back-Propagation. Although these methods are originally derived for convolutional neural networks (CNNs), we adapt them to develop the corresponding suitable versions for GCNNs. We evaluated the contrastive power of these methods with respect to the specificity of the identified molecular substructures and their relevance for chemical functions. Grad-CAM had the highest contrastive power and generated qualitatively the best FGs. This work paves the way for automatic analysis and design of new molecules.

2008

The Cellular Associative Neural Network (CANN) is a novel symbolic pattern matching system, currently used for both the identification of objects in noisy images and for graph similarity searching of chemical structures. Objects are defined by a set of symbolic rules, which iteratively combine low level features into higher level constructs, until object level definitions can be obtained. The flow of information follows a cellular automata based model to aid parallel implementation and rules are stored in AURA associative memories, which provide efficient storage, fast retrieval and the ability to identify partially matching rules in constant time.

bioRxiv (Cold Spring Harbor Laboratory), 2020

Cellular signaling pathways are responsible for decision making that sustains life. Most signaling pathways include post-translational modification cycles, that process multiple inputs and are tightly interconnected. Here we consider a model for phosphorylation/dephosphorylation cycles, and we show that under some assumptions they can operate as molecular neurons or perceptrons, that generate sigmoidal-like activation functions by processing sums of inputs with positive and negative weights. We carry out a steady-state and structural stability analysis for single molecular perceptrons as well as for feedforward interconnections, concluding that interconnected phosphorylation/dephosphorylation cycles may work as multilayer biomolecular neural networks (BNNs) with the capacity to perform a variety of computations. As an application, we design signaling networks that behave as linear and non-linear classifiers.

Lecture Notes in Computer Science, 2012

Learning and generalisation are fundamental behavioural traits of intelligent life. We present a synthetic biochemical circuit which can exhibit nontrivial learning and generalisation behaviours, which is a first step towards demonstrating that these behaviours may be realised at the molecular level. The aim of our system is to learn positive real-valued weights for a real-valued linear function of positive inputs. Mathematically, this can be viewed as solving a non-negative least-squares regression problem. Our design is based on deoxyribozymes, which are catalytic DNA strands. We present simulation results which demonstrate that the system can converge towards a desired set of weights after a number of training instances are provided.

The Journal of Machine Learning Research, 2003

We describe a general methodology for the design of large-scale recursive neural network architec-tures (DAG-RNNs) which comprises three fundamental steps: (1) representation of a given domain using suitable directed acyclic graphs (DAGs) to connect visible and ...

BMC Bioinformatics, 2020

Background Predicting of chemical compounds is one of the fundamental tasks in bioinformatics and chemoinformatics, because it contributes to various applications in metabolic engineering and drug discovery. The recent rapid growth of the amount of available data has enabled applications of computational approaches such as statistical modeling and machine learning method. Both a set of chemical interactions and chemical compound structures are represented as graphs, and various graph-based approaches including graph convolutional neural networks have been successfully applied to chemical network prediction. However, there was no efficient method that can consider the two different types of graphs in an end-to-end manner. Results We give a new formulation of the chemical network prediction problem as a link prediction problem in a graph of graphs (GoG) which can represent the hierarchical structure consisting of compound graphs and an inter-compound graph. We propose a new graph conv...

Proceedings of the 2001 Congress on Evolutionary Computation (IEEE Cat. No.01TH8546), 2001

The hypernetwork model is a hierarchical architecture that has a representation of the molecular, cellular, and organismic levels of biological organization. Influences flow within each level, and through levels, forming dynamic networks of molecular interactions. With its molecular variation-selection learning algorithm, the hypernetwork is able to solve fairly complex tasks such as the (4-10)-input parity task, and the tic-tac-toe endgame problem, with good results. These performance illustrates the learning capabilities of this model.

Cellular Neural Networks and their Applications, …, 1996

The template coefficients (weights) of a CNN, which d l give a d e s i d p e t f o t " e , can either be found by design OT by learning.. By designw meansl thut the dccrircdfunction to be performed could be translated into a set of local dynamic rules, while "ay ICorning' is based ezclwively on pairs of input and c o n q w d n g output signals, the nlcrtioMhip of which m y be by far too complicated for the cqlicit fonnulclrion of loml rules. An ov" of design and leaming methods applicrrbk to CNNs, which sometimes att not c M y distingllishcrbk, d l be given k. Both technological constmints imposed by spec$% hadwatt implementation and pmctical constraints caused by the SpCriFc application and q d e m embedding are influencing design and leanzing. 1 Introduction Since their introduction in 1988 [l] the d k g n of both continuous-time and discretetime cellular neural networks (CT-CNNs and DT-CNNs) hae been an interding research topic. The aim is to h d a set of parameters (coefficients, synaptic weighta), which in the case of locally connected translationally invariant CNNi are usually ulled templates, so that the network perform according to a given tark. The equation for each cell c of CT-CNN is M follows: f(+) := sgn(5).

Prediction of biological and chemical features of a given chemical structure is a challenging problem for the existing nonlinear mapping performed by neural networks. In combinatorial chemistry, computational approaches are capable to significantly decrease the necessary amounts of synthesis for the development of a specific chemical or biological drug. Therefore, the main goal is to distinguish appropriate descriptors from insignificant ones. The experimental design for the classical nonlinear neural network mapping for the approximation of five descriptors and the corresponding reaction of the immune system for the drug development are reported briefly. The results for the different descriptors are presented in comparison.

Journal of Chemical Information and Modeling, 2006

In this paper, we report on the potential of a recently developed neural network for structures applied to the prediction of physical chemical properties of compounds. The proposed recursive neural network (RecNN) model is able to directly take as input a structured representation of the molecule and to model a direct and adaptive relationship between the molecular structure and target property. Therefore, it combines in a learning system the flexibility and general advantages of a neural network model with the representational power of a structured domain. As a result, a completely new approach to quantitative structure-activity relationship/ quantitative structure-property relationship (QSPR/QSAR) analysis is obtained. An original representation of the molecular structures has been developed accounting for both the occurrence of specific atoms/groups and the topological relationships among them. Gibbs free energy of solvation in water, ∆ solv G°, has been chosen as a benchmark for the model. The different approaches proposed in the literature for the prediction of this property have been reconsidered from a general perspective. The advantages of RecNN as a suitable tool for the automatization of fundamental parts of the QSPR/QSAR analysis have been highlighted. The RecNN model has been applied to the analysis of the ∆ solv G°in water of 138 monofunctional acyclic organic compounds and tested on an external data set of 33 compounds. As a result of the statistical analysis, we obtained, for the predictive accuracy estimated on the test set, correlation coefficient R ) 0.9985, standard deviation S ) 0.68 kJ mol -1 , and mean absolute error MAE ) 0.46 kJ mol -1 . The inherent ability of RecNN to abstract chemical knowledge through the adaptive learning process has been investigated by principal components analysis of the internal representations computed by the network. It has been found that the model recognizes the chemical compounds on the basis of a nontrivial combination of their chemical structure and target property. PREDICTING PHYSICAL-CHEMICAL PROPERTIES J. Chem. Inf. Model. C D J. Chem. Inf. Model. BERNAZZANI ET AL. PREDICTING PHYSICAL-CHEMICAL PROPERTIES J. Chem. Inf. Model. E PREDICTING PHYSICAL-CHEMICAL PROPERTIES J. Chem. Inf. Model. I J J. Chem. Inf. Model. BERNAZZANI ET AL. L J. Chem. Inf. Model. BERNAZZANI ET AL.

Three-Dimensional Quantitative Structure Activity Relationships, 2002