Molecular Toxicity Virtual Screening Applying a Quantized Computational SNN-Based Framework

Abstract

Spiking neural networks are biologically inspired machine learning algorithms attracting researchers' attention for their applicability to alternative energy-efficient hardware other than traditional computers. In the current work, spiking neural networks have been tested in a quantitative structure-activity analysis targeting the toxicity of molecules. Multiple public-domain databases of compounds have been evaluated with spiking neural networks, achieving accuracies compatible with high-quality frameworks presented in the previous literature. The numerical experiments also included an analysis of hyperparameters and tested the spiking neural networks on molecular fingerprints of different lengths. Proposing alternatives to traditional software and hardware for time- and resource-consuming tasks, such as those found in chemoinformatics, may open the door to new research and improvements in the field.

Keywords: in silico toxicity prediction; machine learning; molecular fingerprints; spiking neural networks.

Figure 1

Visual summary of meta-analysis on Tox21. Standard deviation was included if reported in the original papers. AUC values from [40,41,42,43,44,45,47,48].

Figure 2

Classification of side effects of chemicals from SIDER dataset in the present and other works. AUC included from [45,52].

Figure 3

Outcomes for BBBP and Clintox compared to the previous literature. AUC values as reported in [40,41,46,49,51,53,54,55,56].

Figure 4

Overview of the procedure during the numerical experiments. The input bit strings were created from SMILES, and a binary label accompanied each instance. The binary sequences were input directly to the SNN, which was evaluated by nested cross-validation with an inner loop for model selection and an outer loop for evaluating the quality of the outcomes.

Figure 5

Membrane potential in LIF neurons is modeled receiving a binary train as input and producing a binary output in response: only when membrane potential is over-threshold do LIF neurons fire a spike marked by ones in the output sequence.

Figure 6

LIF circuit equivalent. A parallel resistor and capacitor represent the neuron, a configuration that could easily describe membrane behavior and, at the same time, is effortlessly implementable on silicon chips.

Figure 7

Exemplification of the common SNN architectures. All SNNs employed were shallow with one hidden layer. The fingerprints’ bits were passed to the input layer, whereas the last layer firing counts determined the predicted class.