System and method for accelerating FEP methods using a 3D-restricted variational autoencoder
Inventors
Bucher, Alwin • Prat, Alvaro • Bastas, Orestis • Kamuntavicius, Gintautas • YANG, Zeyu • Knuff, Charles Dazler • Jocys, Zygimantas • Tal, Roy • Aty, Hisham Abdel
Assignees
Interested in licensing this patent?
MTEC can help explore whether this patent might be available for licensing for your application.
Publication Number
US-11568961-B2
Publication Date
2023-01-31
Expiration Date
Abstract
A system and method for accelerating the calculations of free energy differences by automating FEP-path-decision-making and replacing the standard series of alchemical interpolations typically created by molecular dynamic (MD) simulations with voxelated interpolated states. A novel machine learning approach comprising a restricted variational autoencoder (ResVAE) is used which can reduce the computational-cost associated with interpolations by restricting the dimensions of a molecular latent space. The ResVAE generates a model based on flow-based transformations of a 3D-VAE latent point that is trained to maximize the log-likelihood of MD samples which enables the model to compute transformations more efficiently between molecules and also handle deletions of atoms more efficiently during iterative FEP calculation steps.
Core Innovation
The invention describes an ML-accelerated free-energy perturbation framework that automates FEP path decision-making and replaces MD-driven alchemical interpolation with voxelated interpolated states. The framework uses a restricted variational autoencoder with flow-based transformations of a 3D-VAE latent representation, trained to maximize log-likelihood of MD samples, to enable efficient transformations between ligand states and improved handling of atom deletions during iterative FEP.
The framework also outlines using a 3D convolutional neural network as a force-field surrogate to accelerate 94G estimation. It replicates the FEP thermodynamic cycle in latent space by sampling FEP interpolation steps using the generative model's voxelated latent space, and variables derived from the target ligand's force-field from the trained 3D-CNN are incorporated into the free-energy perturbation calculation steps.
Additionally, the approach performs FEP from multiple initial ligand-protein complexes to a target complex. It produces ensemble free-energy difference predictions for target molecules, generating ensemble prediction of free energy difference predictions between two target molecules.
Claims Coverage
The provided independent claims identify a system and a method that accelerate FEP calculations using a flow-based generative model based on a 3D variational autoencoder voxelated latent point, a 3D convolutional neural network trained on molecular force-fields as a force-field surrogate, and sampling FEP interpolation steps while replicating the thermodynamic cycle in the voxelated latent space.
Flow-based transformation generative model from 3D-VAE voxelated latent point
Train a generative model based on flow-based transformations of a three-dimensional variational autoencoder voxelated latent point.
3D convolutional neural network trained on molecular force-fields
Train a three-dimensional convolutional neural network on molecular force-fields.
Voxelated latent space sampling of FEP interpolation steps
Sample interpolations steps of free energy perturbation calculations using the generative model.
Replicating the FEP thermodynamic cycle in voxelated latent space
Replicate the thermodynamic cycle of free energy perturbation calculations with the generative model's voxelated latent space.
Using target-ligand force-field variables derived from the trained 3D-CNN in FEP steps
Use variables derived from the target ligand's force-field from the trained three-dimensional convolutional neural network in free energy perturbation calculation steps.
Running FEP from a plurality of initial ligand-protein complexes to a target complex
Perform free energy perturbation calculations from a plurality of initial ligand-protein complexes to a target complex.
Ensemble prediction of free energy differences between two target molecules
Generate an ensemble prediction of free energy difference predictions between two target molecules.
Across the independent system and method claims, the core claimed coverage centers on combining a flow-based generative model operating on a 3D-VAE voxelated latent representation with a 3D-CNN surrogate trained on molecular force-fields, then using both to sample voxelated latent FEP interpolation steps that replicate the FEP thermodynamic cycle, incorporating target-ligand force-field-derived variables, running FEP from multiple initial ligand-protein complexes, and generating ensemble predictions of free-energy differences between two target molecules.
Stated Advantages
Accelerate free-energy perturbation (FEP) calculations.
Generate ensemble free-energy difference predictions between two target molecules.
Accelerate 94G estimation using a 3D convolutional neural network as a force-field surrogate.
Replicate the thermodynamic cycle in voxelated latent space for free-energy perturbation calculations.
Documented Applications
FEP calculations from multiple initial ligand-protein complexes to a target complex to produce ensemble free-energy difference predictions for target molecules.
Generating ensemble prediction of free energy difference predictions between two target molecules.
Interested in licensing this patent?