Systems and methods for identifying compounds in a combinatorial library having a particular molecular property

By employing variational autoencoders and normalized flow generative models, the computational challenges of screening compounds in large-scale combinatorial synthesis libraries were addressed, enabling efficient screening of synthetic compounds and improving the efficiency and scalability of drug discovery.

CN122374832APending Publication Date: 2026-07-10
View PDF 1 Cites 0 Cited by

Patent Information

Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Filing Date
2024-04-19
Publication Date
2026-07-10

AI Technical Summary

Technical Problem

Existing combinatorial synthesis libraries (CSLs) are large in scale, leading to computational challenges and wasted resources. Traditional methods struggle to efficiently screen compounds that meet a variety of molecular properties, especially in drug discovery. Generative AIs cannot generate synthetic compounds, and traditional methods are limited in scale and efficiency.

Method used

We employ variational autoencoders and normalized flow generative models. The encoder maps compounds to the latent space, the decoder reconstructs the compound structure, and the molecular properties of the compounds are optimized through a strategy model to ensure that the compounds can be easily manufactured in the laboratory. Combined with α-divergence series equilibrium exploration and development, we achieve efficient screening.

Benefits of technology

By screening out potent, selective, and diverse compounds from a library of over three trillion compounds, the system addresses the problem that generative AI cannot generate synthetic compounds, improving the efficiency and scalability of drug discovery while reducing the time and cost of experimental validation.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure FT_1
    Figure FT_1
  • Figure FT_2
    Figure FT_2
  • Figure FT_3
    Figure FT_3
Patent Text Reader

Abstract

Systems and methods are provided for identifying compounds in a combinatorial synthesis library (CSL) having particular molecular properties. The CSL is accessed using an autoencoder comprising an encoder and a decoder. The encoder maps compounds in the CSL into latent codes in a learned latent space. The decoder retrieves molecular structures of the compounds using such codes. A policy selects a batch of latent codes, and identifies a plurality of compounds from the batch of latent codes using the decoder. Molecular properties of these compounds are determined, and a reward function value for each compound is determined using the molecular properties. A probability density under a target distribution is determined using the reward function values, in which latent codes are sampled from the CSL by the autoencoder with probabilities proportional to the harmonic reward function values. Policy parameters are updated using at least one difference between the probability density and the target distribution using an alpha-divergence based objective function.
Need to check novelty before this filing date? Find Prior Art