Table 4 Summary of DL models mentioned throughout the “Deep learning and molecular dynamics simulations” section of this review used to aid in performing MD workflows

DL-guided enhanced conformational sampling

 AE by Degiacomi [137]

AE model with a latent space encoded from the flattened Cartesian coordinate systems of MD simulation frames of a query protein, from which new protein conformations could be interpolated as starting structures for follow-up MD simulations

 DeepDriveMD [140]

Workflow for protein folding problems employing a CVAE model with a latent space encoded from contact map representations of the flattened Cartesian coordinates of MD simulation frames of a query protein. These conformations get clustered in the latent space in regions with biophysically relevant features, from which protein conformations could be identified as starting coordinates for follow-up MD simulations, in order to speed up the sampling of a protein folding process

 VDE workflow by Sultan et al. [141]

Workflow to sample the most important dynamical behavior of a protein, employing a VDE architecture with a latent space encoded from through-tICA-dimensionality-reduced conformational states of the query protein. The latent coordinate of the VDE was used as CV for well-tempered metadynamics simulations for the sampling of the most important dynamics of the system

Neural network potentials

 ANAKIN-ME [152, 171]

Accurate NeurAl network engine for Molecular Energies or ANI, a HDNNP trained to be a highly transferable architecture for the potential energy predictions of organic molecules

 NNP/MM [172]

A hybrid method combining NNPs and MM calculations, where specific regions of a system are simulated using NNPs and the other parts through faster traditional MM calculations, in order combine the accuracy and efficiency strengths of the two separate simulation methods

DL-guided analysis of MD trajectories

 CNN by Plante et al. [174]

CNN that helps uncovering conformational differences when different ligands are bound to a query protein. The model is trained on 2D scrambled pixel maps of protein conformations of protein–ligand complex MD simulations to predict what ligand that protein state is bound to. This architecture is coupled to an explanation technique highlighting the protein regions in each frame that the network paid attention to for its classification decision. This allows for the analysis of structural features possibly undergoing dynamical differences for each simulation system type

 GLOW [179]

CNN that helps uncovering conformational differences when different ligands are bound to a query protein. The model is trained on 2D residue contact maps of protein conformations of protein–ligand complex Gaussian accelerated MD simulations to predict what ligand that protein state is bound to. This architecture is coupled to an explanation technique highlighting the protein regions in each frame that the network paid attention to for its classification decision. This allows for the analysis of structural features possibly undergoing dynamical differences for each simulation system type, as well as map the free energy landscapes of these conformations

 DL-RP-MDS [183]

Workflow for the functional classification of genetic missense variants into benign or deleterious subgroups. MD simulations are run for a query protein and its missense variations of interest. Then, an AE is trained on the Ramachandran plots of the obtained simulation frames. The latent space of the AE is connected to a classification DFCNN to predict the missense variants to be either benign or deleterious