BRAVE: a highly accurate method for predicting HIV-1 antibody resistance using large language models for proteins

Area of Science:

• Immunology
• Virology
• Computational Biology

Background:

• Broadly neutralizing antibodies (bNAbs) are crucial for HIV-1 prevention and treatment.
• Rapid HIV-1 mutation leads to neutralization resistance, limiting bNAb efficacy.
• Predictive computational methods are needed to anticipate HIV-1 resistance to bNAbs.

Purpose of the Study:

• To develop and validate a novel machine learning approach, BRAVE (Bnab Resistance Analysis Via Evolutionary scale modeling 2), for predicting HIV-1 resistance to 33 bNAbs.
• To assess BRAVE's performance against existing state-of-the-art resistance prediction tools.

Main Methods:

• Utilized a Random Forests classifier integrated with a protein language model to capture essential protein features.
• Trained and evaluated BRAVE on HIV-1 isolates against 33 known bNAbs.
• Performed rigorous statistical comparisons to benchmark BRAVE against leading prediction tools.

Main Results:

• BRAVE demonstrated superior performance across key metrics including accuracy, area under the curve, logarithmic loss, and F1-score.
• Statistical analysis confirmed BRAVE's significantly higher accuracy (p<0.001) compared to current state-of-the-art tools.
• The model effectively captures protein features relevant to bNAb resistance.

Conclusions:

• BRAVE represents a significant advancement in predicting HIV-1 neutralization resistance.
• This tool can inform clinical decisions regarding bNAb selection and usage.
• BRAVE facilitates sequence-based monitoring of viral escape in HIV-1 infected individuals.