Active Learning Coverage Scoring for Compound Selection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing drug discovery processes are lengthy, expensive, and inefficient due to the synthesis and testing of numerous compounds, with machine learning models being limited by insufficient and unrepresentative training data, leading to suboptimal selection of compounds for synthesis.
Innovation Solution
A computational method using active learning and information theory to select compounds based on molecular properties, balancing Shannon entropy and feature coverage, ensuring maximum information gain and representative sampling to improve the training set for machine learning models.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If computational methods are used to evaluate numerous design parameters in parallel, then the level of analysis performed on synthesized compounds is greatly increased, but the accuracy of ML model predictions deteriorates due to insufficient and unrepresentative training data
Solution Approach 1:
The patent implements feedback by using the ML model to predict properties of compounds, then synthesizing and testing those compounds to obtain actual measurements. These measurements are fed back to retrain and improve the model, creating an iterative active learning cycle that continuously enhances prediction accuracy while maintaining high productivity through computational evaluation
Solution Approach 2:
The patent applies preliminary action by using the ML model to pre-select and prioritize compounds for synthesis based on predicted properties. This preliminary computational screening identifies the most promising candidates before actual synthesis, ensuring that limited experimental resources are dedicated to compounds most likely to provide valuable training data
2Reliability
If a significant number of compounds are synthesized and tested during optimization, then the likelihood of obtaining compounds with improved properties increases, but the time and cost of the drug discovery process increases
Solution Approach 1:
The patent extracts the most valuable information from compound testing by using the ML model to identify and focus on specific molecular properties and features that are most predictive of desired outcomes. This extraction of key informative features allows the system to achieve reliable predictions with fewer compounds tested, reducing time while maintaining reliability
Solution Approach 2:
The patent applies partial action by synthesizing and testing only a subset of compounds that are most likely to provide valuable training data, rather than exhaustively testing all possible compounds. The ML model guides this partial sampling to ensure maximum information gain with minimum experimental effort
3Productivity
If compounds are synthesized and tested in sets at each design cycle, then efficiency is improved compared to single compound testing, but the number of compounds that can be tested is limited by available resources
Solution Approach 1:
The patent applies dynamics by making the compound selection process adaptive and dynamic through active learning. The ML model continuously updates its predictions based on new experimental data, dynamically adjusting which compounds to synthesize and test in each cycle. This dynamic approach ensures that limited resources are always allocated to the most informative compounds, maximizing the effective number of compounds that can be tested
Data Source
AI summary
A method for computational drug design includes defining a population of a plurality of compounds. Each compound includes one or more molecular properties. The method includes defining a training set of compounds from the population for which one or more biological properties are known. The method includes selecting, from the population, a subset of one or more compounds that are not in the training set. The method includes determining a subset score of the selected subset based on molecular properties of the one or more compounds in the selected subset, and evaluating the selected subset based on the determined subset score. The subset score is determined based on a frequency of the molecular properties in the population and on a frequency of the molecular properties in a sampled set comprising the training set and the selected subset.