A systematic evaluation of deep learning-based protein structure prediction for HIV-1 enzymes

Accurate protein structure prediction is a prerequisite for proactive, structure-based assessment of antiviral drug resistance. In this work, we systematically benchmark five state-of-the-art deep learning protein structure prediction models (i.e., AlphaFold2, AlphaFold3, ESMFold, Ember3D, and ESM3) on three HIV-1 enzymes central to antiretroviral therapy. We evaluate predictions against a curated dataset of high-resolution experimental structures deposited after the training cutoff dates of the evaluated models, ensuring an unbiased assessment on unseen data. Structural accuracy is quantified using global and per-residue root mean square deviation and TM-score, complemented by an analysis of model-reported confidence scores (pLDDT) and the correlation between the pLDDT and local structural accuracy. Our results show that AlphaFold-based models achieve strong overall performance, although model rankings are enzyme-dependent, with ESMFold being competitive for specific targets, at a lower computational cost. Performance degrades markedly for RT, which presents the greatest modeling challenge due to its large size and multidomain structure. Ember3D and ESM3-Open show poor structural agreement across all targets. Confidence scores correlate meaningfully with local structural deviation for AlphaFold-based models and ESMFold, supporting their use as proxies for prediction reliability in the absence of experimental reference structures. Extended all-atom, side-chain, and functional-site analyses further show that prediction quality is model- and enzyme-dependent at functionally relevant regions, with AlphaFold3 and ESMFold best preserving functional-site geometry. Finally, a docking analysis using protease inhibitor Darunavir indicates that AlphaFold3 most closely reproduces reference docking affinities for HIV-1 protease. These findings provide practical guidance for selecting structure prediction models in virology applications, and highlight the current limitations of state-of-the-art models.