ConfRank: Improving GFN-FF Conformer Ranking with Pairwise Training

Conformer ranking is a crucial task for drug discovery, with methods for generating conformers often based on molecular (meta)dynamics or sophisticated sampling techniques. These methods are constrained by the underlying force computation regarding runtime and energy ranking accuracy, limiting their effectiveness for large-scale screening applications. To address these ranking limitations, we introduce ConfRank, a machine learning-based approach that enhances conformer ranking using pairwise training. We demonstrate its performance using GFN-FF-generated conformer ensembles, leveraging the DimeNet++ architecture trained on pairs of 159 760 uncharged organic compounds from the GEOM data set with r²SCAN-3c reference level. Instead of predicting only on single molecules, this approach captures relative energy differences between conformers, leading to a significant improvement of the overall conformational ranking, outperforming GFN-FF and GFN2-xTB. Thereby, the pairwise RMSD of the relative energy difference of two conformers can be reduced from 5.65 to 0.71 kcal mol^-1 on the test data set, allowing to correctly identify up to 81% of all lowest lying conformers correctly (GFN-FF: 10%, GFN2-xTB: 47%). The ConfRank approach is cost-effective, allowing for scalable deployment on both CPU and GPU, achieving runtime accelerations by up to 2 orders of magnitude compared to GFN2-xTB. Out-of-sample investigations on CREST-generated conformer ensembles from the QM9 data set and conformers taken from an extended GMTKN55 data set show promising results for the robustness of this approach. Thereby, ranking correlation coefficient such as Spearman can be improved to 0.90 (GFN-FF: 0.39, GFN2-xTB: 0.84) reducing the probability of an incorrect sign flip in pairwise energy comparison from 32 to 7%. On the extended GMTKN55 subsets the pairwise MAD (RMSD) could be reduced on almost all subsets by up to 62% (58%) with an average improvement of 30% (29%). Moreover, an exemplary case study on vancomycin shows similar performance, indicating applicability to larger (bio)molecular structures. Furthermore, we motivate the usage of the pairwise training approach from a theoretical perspective, highlighting that while pairwise training can lead to a decline in single sample prediction of absolute energies for ML models, it significantly enhances conformer ranking performance. The data and models used in this study are available at https://github.com/grimme-lab/confrank.