Machine Learning for Toxicity Prediction in Low-Sample Molecular Classes
Deep learning models such as Chemprop have advanced quantitative molecular property prediction, but their reliance on large training sets limits use in data-scarce domains. We propose a framework that fine-tunes a general baseline model trained on publicly available data on small, class-specific datasets. The resulting models retain the baseline's generalization ability while gaining…
Machine learning models like Chemprop have revolutionized the prediction of molecular properties, yet their effectiveness is constrained by the availability of extensive training datasets. To address this limitation in data-scarce domains, researchers have developed a framework that fine-tunes a general baseline model using smaller, class-specific datasets.
This innovative approach allows the resulting models to maintain the generalization capabilities of the baseline model while significantly improving accuracy within specific chemical classes. Moreover, these models generate probabilistic outputs that reflect the uncertainty inherent in the training data.
The framework's efficacy was demonstrated in three distinct toxicity classes, each characterized by a common core structure, target, or mode of action. These classes include organophosphates, androgen receptor antagonists, and estrogen receptor beta antagonists. In each case, the fine-tuned model outperformed traditional machine-learning methods as well as the Environmental Protection Agency's TEST tool.
The probabilistic nature of the predictions offers valuable insight into the reliability of the model's output, facilitating the prioritization of compounds that warrant experimental validation. Furthermore, the framework's seamless integration with data streams of varying quality empowers decision-making processes in chemical safety assessments and drug discovery efforts, allowing for iterative refinement and improved outcomes.
Written by urgent.news from bioRxiv's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
