Machine Learning Models for DNA-Encoded Library Signal Processing
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional DNA encoded library (DEL) experiments suffer from low signal-to-noise ratio due to experimental noise and biases, which negatively impact the performance of machine learning models used for drug discovery by complicating the identification of compounds that bind to protein targets.
Innovation Solution
The development of machine learning models, including classification and regression models, that selectively expand molecular representations and account for confounding sources of noise and biases, enabling improved accuracy in virtual compound screens, hit selection, and binding affinity predictions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If conventional DEL experiments are used to generate training data, then large-scale compound screening is enabled, but the signal-to-noise ratio decreases due to experimental noise and biases
Solution Approach 1:
The patent extracts and removes confounding factors (noise and biases) from the DEL experimental data through computational methods. Specifically, it separates the true binding signal from experimental artifacts by identifying and eliminating sources of noise such as non-specific binding signals and sequence composition biases, thereby improving the signal-to-noise ratio while preserving the large-scale screening capability
Solution Approach 2:
The patent introduces computational processing as an intermediary step between raw DEL experimental data and machine learning model training. This intermediary processing layer applies algorithms to correct biases and filter noise, transforming the raw low-quality data into cleaned, high-quality training data that maintains the original large-scale compound information while removing detrimental experimental artifacts
2Ease of manufacture
If DEL experimental data with noise and biases is used for training, then model training can proceed with available data, but model performance deteriorates
Solution Approach 1:
The patent converts the harmful effect of noise and biases in DEL data into a benefit by using these confounding factors as explicit training targets. The machine learning model is trained to recognize and account for these biases, learning to distinguish true binding signals from experimental artifacts. This approach transforms previously harmful noise into useful information that improves model robustness and performance
Solution Approach 2:
The patent changes the parameters and features used in model training by incorporating corrected and normalized DEL data parameters. It transforms raw sequencing counts into bias-corrected binding affinity estimates by adjusting for factors such as library composition, experimental conditions, and detection biases, thereby improving model reliability while maintaining training feasibility
Data Source
AI summary
Embodiments of the disclosure involve training machine learned models using DNA-encoded library experimental data outputs and for deploying the trained machine learned models for conducting a virtual compound screen, for performing a hit selection and analysis, or for predicting binding affinities between compounds and targets. Machine learned models are trained using one or more augmentations that selectively expand molecular representations of a training dataset. Furthermore, machine learned models are trained to account for confounding covariates, thereby improving the machine learned models' abilities to conduct a virtual screen, perform a hit selection, and to predict binding affinities.