Gene Expression-Guided Molecule Generation With Joint VAE Training
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current deep learning-based generative models lack the ability to design molecules based on disease phenotype alone, leading to high failure rates in drug discovery due to insufficient understanding of drug mechanism of action and off-target effects.
Innovation Solution
A processor-implemented method and system using a joint training of a simplified molecular input line entry system (SMILES) variational autoencoder (s-VAE) and a profile variational autoencoder (p-VAE) to generate drug-like molecules from cell-specific gene expression profiles, employing a joint loss function of cross entropy and Kullback-Leibler divergence for conditional mapping.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If deep learning-based generative models are used to design molecules, then the ability to explore huge chemical space and design target-specific molecules is improved, but the ability to design molecules based on disease phenotype alone is lacking
Solution Approach 1:
The patent combines two separate VAE models (molecular VAE and gene expression VAE) into a unified joint training framework. The molecular VAE processes SMILES representations while the gene expression VAE processes transcriptomic data, and both are trained simultaneously with a joint loss function that includes cross-entropy loss and Kullback-Leibler divergence loss. This merging enables the system to design molecules based on disease phenotypes by establishing a direct mapping between gene expression profiles and molecular structures, thereby improving adaptability while maintaining reliability through the coordinated training of both models.
2Loss of information
If conventional drug discovery methods are used, then the understanding of drug mechanism of action is limited, but the process is simpler and more straightforward
Solution Approach 1:
The patent introduces gene expression profiles as an intermediary that bridges the gap between molecular structures and drug mechanisms of action. The gene expression VAE processes transcriptomic data to capture the biological response to drug treatment, while the molecular VAE generates molecular structures. The joint training framework uses these gene expression intermediaries to infer drug mechanisms of action, thereby reducing information loss about drug behavior while managing system complexity through the modular VAE architecture.
3Measurement precision
If deep learning models are trained without cell-specific gene expression profiles, then the training process is simpler, but the ability to predict off-target effects is insufficient
Solution Approach 1:
The patent applies local quality by using cell-specific gene expression profiles as input to the gene expression VAE. Instead of using generic or averaged gene expression data, the model processes transcriptomic data specific to particular cell types and conditions. This cell-specific approach enables more precise prediction of off-target effects by capturing the nuanced biological responses of different cell types. The joint training framework manages the increased complexity by efficiently processing these cell-specific profiles through the VAE architecture and joint loss function optimization.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
Drug induced gene expression provides information covering various aspects of drug discovery and development. Recent advances in accessibility of open-source drug-induced transcriptomic data along with ability of deep learning algorithms to understand hidden patterns have opened opportunity for designing drug molecules based on desired gene expression signatures. Embodiments herein provide method and system for cell specific model where gene expressions are processed via pretrained Simplified Molecular Input Line Entry System (SMILES) variational autoencoder (s-VAE) to produce new molecules. The model is trained with drug and drug induced gene expression data as input. Both pretrained s-VAE and profile variational autoencoder (p-VAE) are trained jointly. During joint training, difference between newly generated molecules and existing drug molecules is calculated as joint loss function composed of binary cross entropy loss and Kullback-Leibler divergence loss. This loss is backpropagated to decoder to learn conditional mapping of molecular space to transcriptomic space in cell-specific manner.