System and method for molecular reconstruction and probability distributions using a 3D variational-conditioned generative adversarial network

Inventors

Prat, Alvaro • Bucher, Alwin • Jocys, Zygimantas • Tal, Roy • Knuff, Charles Dazler

Assignees

RO5 Inc

Abstract

A system and method that produces an accurate probability distribution representative of a target molecule that may be used in pharmacokinetics and analogous applications. A generator is seeded from a variational autoencoder during training and is then used after training in series with a second variational autoencoder to produce the probability distributions from molecular tensors.

Core Innovation

The invention provides a pharmaceutical research system in which a molecular reconstruction module reconstructs molecular representations and outputs molecular probability distributions. The module receives a true representation of a molecule and constructs a wave-like representation from the true representation, where the wave-like representation comprises tensors of molecular data well-suited for machine learning. The wave-like representation is passed to a first variational autoencoder to produce a reconstructed equivalent of the true representation.

The reconstructed equivalent is then passed to a trained generator, which produces a Gaussian-like probability distribution of the reconstructed equivalent. The system outputs the probability distribution of the molecular representation, with the approach centered on producing molecular probability distributions from reconstructed molecular representations. In the disclosed embodiments, the generator is trained adversarially using a generative adversarial network framework.

The disclosed pipeline is additionally described as producing an interpretable Gaussian-like probability distribution over molecular density, enabling a density-to-molecule transformation to yield valid molecular representations. A broader platform architecture is also described, including a knowledge graph and machine learning modules for bioactivity, ADMET, and clinical trials support, together with background on graph/sequence/protein-ligand modeling and docking or 3D-CNN bioactivity prediction.

Claims Coverage

The independent claims cover a system and a corresponding method for reconstructing molecular representations and outputting molecular probability distributions. Across the independent claims, the inventive sequence is the same: wave-like tensor construction from a true molecular representation, reconstruction using a first variational autoencoder, and probabilistic output using a trained generator that yields a Gaussian-like probability distribution over the reconstructed molecular representation (with dependent claims refining generator training and seeding).

The independent claims are directed to a molecular reconstruction pipeline that converts a true molecular representation into a machine-learning-suitable wave-like tensor representation, reconstructs an equivalent via a first variational autoencoder, and then uses a trained generator to output a Gaussian-like probability distribution of the reconstructed molecular representation.

Stated Advantages

Documented Applications