Compound-Protein Representations for Flexible Bioactivity Prediction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional machine learning systems for predicting biological interactions between genes, compounds, and proteins suffer from inaccuracies, inefficiencies, and operational inflexibility, particularly in generating accurate and flexible bioactivity predictions.
Innovation Solution
A protein interaction learning system utilizes a compound-protein interaction machine learning model to generate a machine learning representation, which trains target machine learning models to predict bioactivity results and identifies contributing proteins through explainability models, improving accuracy, flexibility, and efficiency.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If conventional machine learning systems utilize large volumes of training data to generate predictions, then the coverage of biological interactions is improved, but the accuracy and efficiency deteriorate
Solution Approach 1:
The patent introduces compound-protein machine learning representations as an intermediary layer between raw training data and final bioactivity predictions. These representations encode interaction patterns in a condensed format, serving as a mediator that transforms large-volume raw data into concentrated, informative features that improve prediction accuracy without requiring proportional increases in data volume.
Solution Approach 2:
The system combines multiple types of information (compound features, protein features, and interaction patterns) into composite compound-protein machine learning representations. This composite approach integrates diverse data sources into a unified representation that captures complex biological interactions more effectively than single-source features, thereby improving prediction accuracy.
2Quantity of substance
If conventional machine learning systems utilize large volumes of training data to generate predictions, then the coverage of biological interactions is improved, but the computational efficiency deteriorates
Solution Approach 1:
The patent extracts essential interaction patterns from large volumes of training data and consolidates them into compact compound-protein machine learning representations. This extraction process separates the critical information from the voluminous raw data, retaining only the most relevant features for prediction while discarding redundant information, thereby improving computational efficiency.
Solution Approach 2:
The system transforms the parameter representation from raw high-dimensional training data to condensed machine learning representations with optimized feature dimensions. This parameter transformation reduces the computational complexity by changing the representation format to one that requires fewer computational resources for processing while maintaining predictive power.
3Reliability
If conventional machine learning systems are designed for specific prediction tasks, then the specialization is improved, but the operational flexibility deteriorates
Solution Approach 1:
The patent creates compound-protein machine learning representations that serve multiple functions across different prediction tasks. These representations are universally applicable to various bioactivity prediction problems (ADMET predictions, biological perturbation programs, etc.), allowing a single representation framework to support multiple specialized tasks without requiring separate models for each application.
Solution Approach 2:
The system employs dynamic target machine learning models that can be trained and adapted to different specific tasks using the same compound-protein representations. The models can dynamically adjust their parameters and configurations based on the specific prediction task at hand, enabling both specialization for individual tasks and flexibility across multiple applications.
Data Source
AI summary
The present disclosure relates to systems, non-transitory computer-readable media, and methods that utilizing compound-protein machine learning representations to generate target results. For example, the disclosed systems can utilize a compound-protein interaction machine learning model to generate a compound-protein machine learning representation for compound protein pairs. The disclosed systems can utilize the compound-protein machine learning representation to train and utilize other target machine learning models in generating predicted bioactivity results. For example, the disclosed systems train a target machine learning model from compound-protein machine learning representations to generate ADMET predictions and/or biological perturbation program predictions. Furthermore, the disclosed systems can utilize one or more explainability models in conjunction with target machine learning models trained based on compound-protein machine learning representations to identify proteins that contribute to predicted bioactivity results.