Summary

DiffDock confidence is inversely correlated with RMSD (1). Note that the training objective is different: simple cross-entropy training on whether a pose is within 2Å RMSD. A follow-up paper by Corso et al. (2) found that the confidence of these models is actually a more accurate predictor than the diffusion model itself.

See also

1.
Corso G, Stark H, Jing B, Barzilay R, Jaakkola TS. DiffDock: Diffusion Steps, Twists, and Turns for Molecular Docking. In: ICLR 2023. 2023. Available from: https://openreview.net/forum?id=kKF8_K-mBbS
2.
Corso G, Deng A, Fry B, Polizzi N, Barzilay R, Jaakkola T. Deep Confident Steps to New Pockets: Strategies for Docking Generalization. 2024; Available from: https://arxiv.org/abs/2402.18396