Abstract
A predictive model is only as useful as its accessibility. The most accurate resistance prediction algorithm in the world is worthless if it lives exclusively in a researcher’s Python notebook and cannot be applied routinely by virologists, clinicians, or pharmaceutical scientists working with new patient samples. Translating computational models into deployable tools is a bottleneck that the field recognizes but rarely addresses.
SePTeR (Sequence-based Prediction of Temsavir Resistance) was designed to close this gap. It is a software tool that automates the entire resistance estimation process from viral sequences: a user inputs a sequence, and the tool returns a resistance estimate along with statistical analysis, visualization, and comparative analysis against reference sequences. The goal is to make the predictive power of our AI-based resistance models available to anyone working with HIV-1 sequence data, without requiring expertise in machine learning or bioinformatics programming.
The first version of SePTeR is functional and in active use within our collaboration with ViiV Healthcare. A planned second version will incorporate additional functionality including sequence alignment, sensitivity analysis, and individualized analysis through SHAP (SHapley Additive exPlanations) values — enabling users to understand not just whether a given sequence predicts resistance, but which specific positions and mutations are driving the prediction for that particular patient’s virus. This level of interpretability is essential for clinical translation, where treatment decisions must be explainable.
We plan to extend the SePTeR framework to the N6 broadly neutralizing antibody, with a proposal to ViiV Healthcare in development.