| language: en | |
| license: mit | |
| tags: | |
| - drug-repurposing | |
| - biomedical-nlp | |
| - bert | |
| - text-classification | |
| - twas | |
| # TRACE Classifier | |
| This is the fine-tuned BiomedBERT classifier for [TRACE](https://github.com/otienoco/TRACE) — TWAS-driven Repurposing through AI-assisted Curation of Evidence. | |
| ## Model Description | |
| The classifier is built on [BiomedBERT](https://huggingface.co/microsoft/BiomedNLP-BiomedBERT-base-uncased-abstract-fulltext) and fine-tuned to classify PubMed abstracts for drug-gene relationships across three simultaneous tasks: | |
| - **Relationship detection** — does this abstract support a direct drug-gene interaction? | |
| - **Mechanism classification** — inhibitor, agonist, activator, antagonist, downregulator, upregulator, etc. | |
| - **Direction classification** — does the drug increase or decrease gene expression or activity? | |
| Held-out macro F1: **0.809** across all three tasks. | |
| ## Intended Use | |
| This model is used as part of the TRACE pipeline. It is not intended to be used as a standalone model outside of that context. See the [TRACE GitHub repository](https://github.com/otienoco/TRACE) for full usage instructions. | |
| ## Citation | |
| Otieno CO, Seagle HM, Akerele AT, Jaworski J, Guare L, Setia-Verma S, Velez Edwards DR, Edwards TL. TRACE: A fine-tuned biomedical language model for directionally informed drug repurposing from transcriptome-wide association studies. 2026. Preprint. | |