Molecular Fingerprint Based and Machine Learning Driven QSAR for Bioconcentration Pathways Determination

Quantitative structure-activity relationship associates molecules’ structural characteristics to their bio-activity, and it can be performed via machine learning to find risky chemicals that accumulate in living organisms. The present analysis focused on investigating how structural information of molecules encoded by molecular fingerprints can be used to predict bioaccumulation pathways. Numerical experiments involved extreme gradient boosting, support vector machines, and neural networks, including spiking neural networks. This investigation might be the first attempt to apply this particular kind of biologically inspired neural network for predicting molecules’ functions from their fingerprints. The computational models forecasted three possible bioaccumulation processes, with support vector machines obtaining the mean peak accuracy and the spiking neural network architectures achieving satisfactory results: the leaky neuron spiking neural network range of outcomes was not statistically different from the accuracies of the support vector machine algorithms. In addition, an algorithm broadly found in chemoinformatics literature as the extreme gradient boosting algorithm fulfilled compatible accuracies with spiking neural networks and support vector machines. This three-class machine learning-driven bioconcentration modeling of chemicals established a foundation for future analysis pipelines focusing on predicting bioaccumulation in biological tissues by investigating molecules’ structures.