Machine Learning to Predict Teratogenicity: Theory and Practice.

Machine learning (ML) is a subfield of artificial intelligence (AI) that consists of developing algorithms that can automatically learn patterns and relationships from data, without being explicitly programmed. It continues to advance with the development of more sophisticated algorithms, increased computational power, and larger datasets, leading to significant advancements in AI technology. With the significant progress made in ML, the need to apply these systems in the area of teratogenicity is growing. It is sought as robust boosting methods to overcome many limitations and restrictions facing the experimental studies. By performing tasks such as classification, regression, clustering, anomaly detection, and decision systems, ML can be used to assess whether an agent is teratogen or not or to determine its teratogenic potential. It may also be used for the purpose of deciding on the use of medicinal products. In this chapter, we describe how ML can be used to investigate teratogenicity.

Artificial neural networks: non-linear QSAR studies of HEPT derivatives as HIV-1 reverse transcriptase inhibitors.

Structure-anti HIV activity relationships were established for a sample of 80 1-[2-hydroxyethoxy-methyl]-6-(phenylthio)thymine (HEPT) using a three-layer neural network (NN). Eight structural descriptors and physicochemical variables were used to characterize the HEPT derivatives under study. The network's architecture and parameters were optimized in order to obtain good results. All the NN architectures were able to establish a satisfactory relationship between the molecular descriptors and the anti-HIV activity. NN proved to give better results than other models in the literature. NN have been shown to be particularly successful in their ability to identify non-linear relationships.

Neural networks: Accurate nonlinear QSAR model for HEPT derivatives.

A nonlinear quantitative structure-anti-HIV-1-activity relationship (QSAR) study was investigated in a series of 1-[2-hydroxyethoxy-methyl]-6-(phenylthio) thymine] (HEPT) derivatives acting as nonnucleoside reverse transcriptase inhibitors (NNRTIs). This QSAR study has been undertaken by a three-layered neural network (NN) using molecular descriptors known to be responsible for the anti-HIV-1 activity. The usefulness of the model and the nonlinearity of the relationship between molecular descriptors and anti-HIV-1 activity have been clearly demonstrated. The obtained model outperforms those given in the literature in both the fitting and predictive stages. NN analysis yielded predicted activities in excellent agreement with the experimentally obtained values (R(2) = 0.977, predictive r(2) = 0.862). The effect of each molecular feature on the anti-HIV-1 activity variation has been clearly elucidated.