ChemModernBERT: LD50 Predictions for 8,898 Compounds

8,898 compounds with median lethal dose (LD50) values form a curated dataset for toxicity prediction. Tanuj Sharma authored this research, which compares four molecular representation learning approaches, including the ChemModernBERT model. The dataset was last updated on March 19, 2026.

Use Cases

• Predict median lethal dose (LD50) values for chemical compounds based on molecular representations.
• Benchmark machine learning models for toxicity prediction using standardized internal and external test sets.
• Support computational pipelines for carcinogenicity assessment and early-stage chemical safety evaluation.
• Compare transformer architectures like ChemModernBERT against directed message passing neural networks and ensemble models.

Strengths

• Dataset is curated and contains 8,898 compounds.
• Model performance was evaluated using standardized internal and external test sets.
• The associated ChemModernBERT model achieved a low mean absolute error (MAE) of 0.390 on internal evaluation.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.
• The dataset is very small (7.5 KB), indicating limited scope.

Provenance

Source

Tanuj Sharma

Collection Method

Likely compiled for research comparing molecular representation learning approaches.

Freshness

Last updated 2026-03-19 06:47:58; freshness should be verified.