HIV Drug Resistance Predictions with Protein Language Model Embeddings

Hayden Farquhar's 2026 figshare release contains raw genotype-phenotype data, processed datasets, and results for benchmarking protein language models on HIV drug resistance prediction. The 4.2 MB repository includes Stanford HIVDB datasets for protease, nucleoside, and non-nucleoside reverse transcriptase inhibitors, alongside comparative results for models like ESM-2. It provides all figures from the associated publication at high resolution.

Use Cases

• Benchmarking performance of protein language model embeddings (ESM-2, ESM-1v) against traditional features for predicting drug resistance phenotypes from HIV genotype sequences.
• Analyzing subtype-stratified performance to assess model generalizability across different HIV viral subtypes.
• Conducting temporal holdout validation to evaluate prediction robustness on future, unseen viral sequences.
• Training attention-based interpretability models to identify key mutations in protease (PI), nucleoside (NRTI), and non-nucleoside (NNRTI) reverse transcriptase inhibitor resistance.

Strengths

• Includes curated Stanford HIVDB genotype-phenotype datasets, a recognized standard for resistance research.
• Provides multi-model comparison results for three major protein language models (ESM-2, ESM C 600M, ESM-1v).
• Contains all 10 manuscript figures at publication-quality 400 DPI resolution.

Limitations

• The dataset is small at 4.2 MB, indicating limited raw sequence or sample volume.
• The exact number of genotype-phenotype records or rows is unspecified.
• Analysis code is archived separately, requiring an additional download for full reproducibility.

Provenance

Source

Stanford HIVDB (HIV Drug Resistance Database) genotype-phenotype datasets.

Collection Method

Data processed and results generated for a specific benchmarking study on protein language model embeddings.

Time Range

null

Freshness

Last updated in April 2026, indicating recent work.

Geography

null