Accurately predicting HIV-1 sensitivity to broadly neutralizing antibodies (bNAbs) is a critical step in advancing the development of effective therapeutic and preventive strategies against HIV-1. Due to the virus’s extraordinary genetic diversity and rapid mutation rate, traditional experimental methods for assessing bNAb sensitivity are labor-intensive, time-consuming, and limited in scalability. However, current machine learning approaches prioritize feature significance while neglecting to address class imbalance and systematic biases in multi-laboratory datasets. To address these limitations, we develop the Statistical Distribution Sampling (SDS) method, a novel statistical framework that systematically addresses data imbalance and cross-institutional variance. Firstly, the amino acid sequence is converted into a numerical vector via the k-string method. Then, antibodies are classified as effective or ineffective based on IC50/IC80 thresholds, with a stratified random subset selected as the Sampling generator. Each vector dimension’s distribution is characterized through histogram analysis and kernel density estimation. Afterwards, 500 data points are sampled from the distribution of each column for each category. The assembled Statistical Distribution Sampling (SDS) matrix integrates column features, which combined with the Sampling generator, produces the final training dataset. Antibody efficacy predictions are ultimately generated through integration with the Random Forest algorithm. This approach enables comprehensive prediction of all database sequences while achieving strong predictive performance. Through 10 independent trials per antibody, our method achieved statistically superior AUC and ACC performance compared to the LBUM and SLAPNAP benchmarks. Moreover, our method demonstrates significantly lower standard deviation across 10 runs than both counterparts, proving its consistent robustness across diverse antibody scenarios.
He et al. (2026) studied this question.