Time-Domain Analysis of Molecular Dynamics Trajectories Using Deep Neural Networks: Application to Activity Ranking of Tankyrase Inhibitors

Molecular dynamics simulations provide
valuable insights into the
behavior of molecular systems. Extending the recent trend of using
machine learning techniques to predict physicochemical properties
from molecular dynamics data, we propose to consider the trajectories
as multidimensional time series represented by 2D tensors containing
the ligand–protein interaction descriptor values for each time
step. Similar in structure to the time series encountered in modern
approaches for signal, speech, and natural language processing, these
time series can be directly analyzed using long short-term memory
(LSTM) recurrent neural networks or convolutional neural networks
(CNNs). The predictive regression models for the ligand–protein
affinity were built for a subset of the PDBbind v.2017 database and
applied to inhibitors of tankyrase, an enzyme of the poly­(ADP-ribose)-polymerase
(PARP) family that can be used in the treatment of colorectal cancer.
As an additional test set, a subset of the Community Structure–Activity
Resource (CSAR) data set was used. For comparison, the random forest
and simple neural network models based on the crystal pose or the
trajectory-averaged descriptors were used, as well as the commonly
employed docking and molecular mechanics Poisson–Boltzmann
surface area (MM-PBSA) scores. Convolutional neural networks based
on the 2D tensors of ligand–protein interaction descriptors
for short (2 ns) trajectories provide the best accuracy and predictive
power, reaching the Spearman rank correlation coefficient of 0.73
and Pearson correlation coefficient of 0.70 for the tankyrase test
set. Taking into account the recent increase in computational power
of modern GPUs and relatively low computational complexity of the
proposed approach, it can be used as an advanced virtual screening
filter for compound prioritization.