Predicting cytotoxicity from heterogeneous data sources with Bayesian learning

Area of Science:

• Computational chemistry
• Drug discovery
• Toxicology

Background:

• Utilized data from over 80 cytotoxicity assays from internal and public sources.
• Investigated feasibility of a general cytotoxicity model using diverse assay formats.
• Aimed to identify potentially cytotoxic compounds early in drug discovery.

Purpose of the Study:

• Develop a computational model for general cytotoxicity prediction.
• Highlight potentially cytotoxic compound series early in the drug discovery pipeline.
• Integrate diverse cytotoxicity assay data into a unified predictive model.

Main Methods:

• Developed Bayesian models using Scitegic FCFP_6 fingerprints and physical property descriptors.
• Identified mutually predictive assay pairs using ROC scores (>0.60).
• Merged data from 48 connected assays into a training set of 145,590 compounds.

Main Results:

• Visualized assay inter-predictivity in a network graph.
• Observed that mutual predictivity between assay pairs (A, B) and (B, C) did not guarantee predictivity for (A, C).
• Derived and validated a general cytotoxicity model using cross-validation and FDA-approved drugs.

Conclusions:

• Generated a predictive model to accelerate drug discovery by enabling early identification of cytotoxic compounds.
• Demonstrated that different assay formats can be mutually predictive, reducing redundant testing.
• Facilitated selection of optimal assay panels or corporate standards based on predictivity.
• Provided a comprehensive dataset of known cytotoxic compounds from the Pfizer collection.
• Aided in designing new compounds with desired cytotoxicity profiles by comparing model output with in vitro safety data.