ConfRank: Improving\nGFN-FF Conformer Ranking with\nPairwise Training

Conformer ranking is a crucial task for drug discovery,\nwith methods\nfor generating conformers often based on molecular (meta)dynamics\nor sophisticated sampling techniques. These methods are constrained\nby the underlying force computation regarding runtime and energy ranking\naccuracy, limiting their effectiveness for large-scale screening applications.\nTo address these ranking limitations, we introduce ConfRank, a machine\nlearning-based approach that enhances conformer ranking using pairwise\ntraining. We demonstrate its performance using GFN-FF-generated conformer\nensembles, leveraging the DimeNet++ architecture trained on pairs\nof 159 760 uncharged organic compounds from the GEOM data set with\nr<sup>2</sup>SCAN-3c reference level. Instead of predicting only on\nsingle molecules, this approach captures relative energy differences\nbetween conformers, leading to a significant improvement of the overall\nconformational ranking, outperforming GFN-FF and GFN2-xTB. Thereby,\nthe pairwise RMSD of the relative energy difference of two conformers\ncan be reduced from 5.65 to 0.71 kcal mol<sup>–1</sup> on the\ntest data set, allowing to correctly identify up to 81% of all lowest\nlying conformers correctly (GFN-FF: 10%, GFN2-xTB: 47%). The ConfRank\napproach is cost-effective, allowing for scalable deployment on both\nCPU and GPU, achieving runtime accelerations by up to 2 orders of\nmagnitude compared to GFN2-xTB. Out-of-sample investigations on CREST-generated\nconformer ensembles from the QM9 data set and conformers taken from\nan extended GMTKN55 data set show promising results for the robustness\nof this approach. Thereby, ranking correlation coefficient such as\nSpearman can be improved to 0.90 (GFN-FF: 0.39, GFN2-xTB: 0.84) reducing\nthe probability of an incorrect sign flip in pairwise energy comparison\nfrom 32 to 7%. On the extended GMTKN55 subsets the pairwise MAD (RMSD)\ncould be reduced on almost all subsets by up to 62% (58%) with an\naverage improvement of 30% (29%). Moreover, an exemplary case study\non vancomycin shows similar performance, indicating applicability\nto larger (bio)molecular structures. Furthermore, we motivate the\nusage of the pairwise training approach from a theoretical perspective,\nhighlighting that while pairwise training can lead to a decline in\nsingle sample prediction of absolute energies for ML models, it significantly\nenhances conformer ranking performance. The data and models used in\nthis study are available at https://github.com/grimme-lab/confrank.