C2S-HBART: Toxicity Predictions with Uncertainty Estimates for Drug Screening

C2S-HBART, a novel conformalized heteroscedastic Bayesian framework, reduces the rate of confident but incorrect toxicity predictions from 12.02% to 1.8% compared to a baseline. The model was evaluated on the Tox21 dataset and provides statistically rigorous decision-support for preclinical screening. Author B. Özlüer Başer published the 17.4 KB text file on figshare in April 2026.

Use Cases

• Prioritizing safer drug candidates based on predicted toxicity and associated uncertainty.
• Flagging structurally complex molecules for experimental validation based on high predictive variance.
• Benchmarking new uncertainty quantification methods against the reported 26% improvement in information efficiency.
• Analyzing drivers of predictive uncertainty, such as molecular size and topological complexity.

Strengths

• Achieves near-nominal coverage of 0.952 for uncertainty intervals on the Tox21 dataset.
• Reduces the rate of silent failures (confident but incorrect predictions) to 1.8%.
• Provides a 26% improvement in information efficiency, yielding sharper predictions (Avg Width: 1.20).

Limitations

• Row count is unknown, which may limit suitability assessment.
• Column-level documentation is absent; field semantics must be inferred after download.
• The dataset is very small at 17.4 KB, indicating limited scope, likely containing model outputs or summary statistics rather than raw experimental data.

Provenance

Source

B. Özlüer Başer via figshare.

Collection Method

Model outputs from the C2S-HBART framework evaluated on the Tox21 dataset.

Freshness

Last updated 2026-04-20 10:13:53; freshness should be verified.