Neural Network Models for Prediction of Biological Activity using Molecular Dynamics Data: A Case of Photoswitchable Peptides

TL;DR

This paper shows that using molecular dynamics data improves neural network predictions of biological activity for complex, flexible molecules like photoswitchable peptides.

Contribution

The study introduces neural network models using molecular dynamics features to predict biological activity for conformationally flexible molecules.

Findings

MD-derived features outperform static 2D/3D descriptors in predicting peptide activity.

Two NN models accurately predict cytotoxicity and photoisomer activity differences.

The approach generalizes to molecules previously considered too complex for ligand-based models.

Abstract

Prediction of biological activities of chemical compounds by the machine learning techniques in general and the neural networks (NNs) in particular, is usually based on the analysis of their binding to the target of interest. If such affinity data is not available, the ligand‐based approaches can be used where the NN models are trained to assess similarity of compounds to those with known biological activity. Obviously, this approach only works well if the similarity between the training set and the evaluated molecules is sufficiently high. In the case of large and conformationally flexible organic compounds, the activity becomes dependent not only on chemical identity but also on the dynamics of molecular motions, which imposes significant challenges to existing approaches based on static structural 2D and 3D molecular descriptors. A prominent example of compounds, which are especially challenging for existing NN activity prediction techniques, are photoswitchable macrocyclic peptides containing a diarylethene “photoswitch” (DAE). These molecules exist in two isomeric forms with remarkably different biological activities, which are interconvertible by light of different wavelengths. Activity prediction models have to distinguish in this case not only between the different peptides but also between the photoisomers of the same peptide. In this work, we demonstrate that the features extracted from classical molecular dynamics (MD) trajectories are superior to conventional 2D or 3D descriptor‐based features when used in activity prediction NN models of DAE‐containing photoswitchable peptides. Using MD‐derived features, we successfully created two NN models that predict activities of photoswitchable peptidomimetics, analogs of the natural peptidic antibiotic gramicidin S. The first model precisely predicts the cytotoxic activity of similar peptide analogs. The second model reliably predicts the differences in the biological activities of DAE photoisomers of the same peptide, even if the type of its activity differs from one in the training dataset. Our results demonstrate that accounting for MD‐derived dynamic features allows generalizing the ligand‐based activity prediction NN models to the cases of large and conformationally flexible molecules, which were previously considered intractable by this class of models.

Two neural network models are reported that can reliably predict biological activities of complex molecules like cyclic peptides in target‐independent manner. Both models use the features extracted from molecular dynamics simulation data and contain information about the 3D structure and dynamics of the molecules.© 2025 WILEY‐VCH GmbH

Figures