The World Health Organization classifies the emergence of pathogenic resistance in RNA viruses as a global public health threat.  While drug-resistance mutations (DRMs) in these viruses have been well cataloged, the pathways and mechanisms through which they arise remain largely unknown. Without proactive strategies, our public health responses remain reactive and “one step behind” emerging variants. Focusing on the human immuno-deficiency virus (HIV) as a model system, I will present my work that combines: (i) physics-based machine learning (ML) models built on patient‑derived multiple sequence alignments (MSAs) to learn the context-dependent relationships (epistasis) and pathways leading to drug resistance evolution; (ii) a novel kinetic co-evolutionary approach to generate sequence evolutionary trajectories that can capture the kinetics of drug resistance acquisition; and (iii) biophysical methods (including cryogenic electron microscopy or cryo-EM)  to reveal the structural and thermodynamic mechanisms leading to drug resistance. The overarching aim is a constraint‑aware theory of viral evolution.

I will first discuss the role of epistasis in evolutionarily trapping fitness detrimental DRMs in patient genetic sequence backgrounds [1] and highlight our novel methodology combining machine learnt fitness landscapes with kinetic Monte Carlo simulations that accurately predict the clinically reported acquisition times of >50+ DRMs across all HIV drug-target proteins [2].  We identify an ‘epistatic barrier’ as the dominant control knob on the resistance kinetics [2, 3]. To probe the  biophysical origins of the epistatic interactions, we then use ML-guided free‑energy perturbation molecular dynamics (FEP/MD) simulations revealing that epistatis arises primarily due to co-operative effects on protein stability that is intrinsic to the protein’s architecture [4]. This reframes resistance as navigation through pre‑wired corridors on the intrinsic protein fitness landscape enabling forecasting where resistance mutations will arise. Finally, I will discuss our current work integrating ML fitness landscape models with cryo-EM and multi-scale MD simulations to obtain the dominant pathways and molecular mechanisms leading to drug resistance in HIV patients.

Together, these results support a shift from current “reactive” variant tracking to mechanistic forecasting of evolutionary trajectories that can redefine therapeutic design and global surveillance by offering a blueprint for anticipating and pre-empting resistance before it manifests. I will also discuss how the different pieces can be unified within a Reinforcement Learning based framework into an end‑to‑end forecasting engine for emerging and endemic RNA viruses.