Modelling compound cytotoxicity using conformal prediction and PubChem HTS data.

Area of Science:

• Computational chemistry
• Toxicology
• Drug discovery

Background:

• Compound cytotoxicity assessment is crucial for efficient drug discovery.
• High-throughput screening generates large datasets, necessitating robust predictive models.
• Imbalanced datasets, common in cytotoxicity screening, pose challenges for accurate modeling.

Purpose of the Study:

• To apply class conditional conformal prediction for modeling compound cytotoxicity.
• To develop accurate predictive models for both toxic and non-toxic compounds across multiple cell lines.
• To evaluate the performance of conformal prediction in large-scale, imbalanced cytotoxicity datasets.

Main Methods:

• Utilized 16 high-throughput cytotoxicity assays from PubChem, covering over 440,000 compounds and 16 cell lines.
• Trained individual classification models for each cell line using class conditional conformal prediction.
• Validated model performance on internal datasets and tested on external data from the same assay provider.

Main Results:

• Developed high-quality predictive models for cytotoxicity despite significant class imbalance (0.8% cytotoxic).
• Achieved balanced performance in predicting both toxic and non-toxic compounds.
• On external data, demonstrated 74% sensitivity and 65% specificity at an 80% confidence level for single-class predictions.

Conclusions:

• Class conditional conformal prediction offers a robust framework for large-scale cytotoxicity modeling.
• The approach provides reliable predictions with controllable error rates, enhancing decision-making in drug discovery.
• This method represents a balanced and accurate alternative to previous large-scale cytotoxicity modeling techniques.