Machine Learning of Allosteric Effects: The Analysis of Ligand-Induced Dynamics to Predict Functional Effects in TRAP1

Abstract

Allosteric molecules provide a powerful means to modulate protein function. However, the effect of such ligands on distal orthosteric sites cannot be easily described by classical docking methods. Here, we applied machine learning (ML) approaches to expose the links between local dynamic patterns and different degrees of allosteric inhibition of the ATPase function in the molecular chaperone TRAP1. We focused on 11 novel allosteric modulators with similar affinities to the target but with inhibitory efficacy between the 26.3 and 76%. Using a set of experimentally related local descriptors, ML enabled us to connect the molecular dynamics (MD) accessible to ligand-bound (perturbed) and unbound (unperturbed) systems to the degree of ATPase allosteric inhibition. The ML analysis of the comparative perturbed ensembles revealed a redistribution of dynamic states in the inhibitor-bound versus inhibitor-free systems following allosteric binding. Linear regression models were built to quantify the percentage of experimental variance explained by the predicted inhibitor-bound TRAP1 states. Our strategy provides a comparative MD–ML framework to infer allosteric ligand functionality. Alleviating the time scale issues which prevent the routine use of MD, a combination of MD and ML represents a promising strategy to support in silico mechanistic studies and drug design.

Ligand-induced protein dynamics differences correlate with protein-ligand binding affinities: An unsupervised deep learning approach

“…Furthermore, a similar approach could be used to study other protein-ligand binding events. For instance, it is interesting to evaluate the relationship between allosteric dynamics and ligand function, as has been done in a few previous studies [20,29]. Another potential application would be predicting the effects of protein mutations from the dynamics of the ligand.…”

“…In this study, we propose a novel method to predict binding energies from proteins' dynamics change upon ligand binding, based on a deep learning approach for MD data analysis [28]. In contrast to general approaches for MD trajectory-based supervised machine learning [29,30], our method uses raw MD trajectories of a ligand binding site with different kinds of ligands, and quantitatively measures differences in the dynamics using unsupervised learning. Overall, the workflow consists of simplification of MD trajectories as local dynamics ensemble (LDE), calculation of dynamics difference as Wasserstein distance, and extraction of variables and detection of the contributing residues.…”

From Data to Knowledge: Systematic Review of Tools for Automatic Analysis of Molecular Dynamics Output

“…On the other hand, attempts to classify ligands on the basis of separate features or chemometrics properties (here, molecular fingerprints) were far less efficient. In contrast, Ferraro et al (2021) aimed to predict allosteric ligand functionality quantitatively. A computational experiment was performed on the allosteric modulators of the molecular chaperone TRAP1, which had similar affinities, but inhibited ATPase function with different efficacy.…”

Application advances of deep learning methods for de novo drug design and molecular dynamics simulation

“…Apart from that, the models show a good agreement with infrared spectra prediction of a methanol molecule, n‐alkanes, and the protonated alanine tripeptide with the theoretical and experimental spectra. Another example involves Ferraro et al, 131 who employed machine learning to detect the dynamic patterns of allosteric inhibitor‐bound and inhibitor‐free TRAP1 based on the MD ensemble. It indicates that machine learning can maximize the information that is included in ns–μs time scale MD simulations.…”