NRGRank: Coarse-grained structurally-informed ultra-massive virtual screening

NRGRank is a coarse-grained structurally-informed virtual screening Python package with accuracy comparable to docking-based methodologies but up to 100-fold speed increase. NRGRank is based on a coarse-grained evaluation of pairwise atom-type pseudo-energy interactions that implicitly accounts for compound and side-chain flexibility as well as limited backbone movements. We compare NRGRank to docking-based virtual screening software Glide, Autodock Vina and DOCK 3.7 on the DUD-E virtual screening benchmark using enrichment factors at 1% (EF1). We observe broad variations of EF1 values across targets, structural models and methods. For apo form or AlphaFold2 models, out of a subset of 37 targets from DUD-E, NRGRank has better EF1 values than Glide for 12 and 13 targets respectively. Even in holo form, where the accuracy of classical docking software increases, NRGRank has better EF1 values than 13, 10 and 5 targets out of 37 compared to AutoDock Vina, DOCK 3.7 and Glide respectively. Comparing the rank of true binders in Glide and NRGRank shows that true binders ranked in the top 1% are complementary between methods irrespective of the target form (AlphaFold, apo or holo). That is, utilizing NRGRank detects binders that are missed by Glide (and presumably other methods), whereas those found by Glide are missed by NRGRank. Furthermore, we observe that most hits found by NRGRank within the top 50 predictions (4.38 {+/-} 5.49 hits on average for AF2 targets) remain once the top 1% of predictions are re-scored with Glide, but the hit rate within the top 50 predictions increases. NRGRank can evaluate one molecule in 0.3 s on average, enabling a modern laptop with 8 cores to screen 1,000,000 molecules in 24 hours - up to two orders of magnitude faster than the reported speed of DOCK 3.7, AutoDock Vina running on GPUs and Glide. NRGRank occupies a unique niche among tools for virtual screening being insensitive to structural inaccuracies but with comparable accuracy as state-of-the-art docking methods and fast as AI-based methods but without the dangers of overfitting as it is based on 780 pseudo-energy parameters. Combined with the fact that NRGRank does not require extensive or expensive computational resources or expert pre-processing of targets, it is unique in making high-performance ultra-massive virtual screening accessible to all.