It has been discovered that machine learning-based scoring functions (MLBSFs) perform inconsistently across various benchmarks and are susceptible to bias in learning datasets. A more thorough knowledge of their performance is necessary if the field wants to produce MLBSFs that acquire a generalizable understanding of physics.
The performance of a variety of well-known MLBSFs (RFScore, SIGN, OnionNet-2, Pafnucy, and PointVS) was contrasted in this work with the authors suggested baseline models, which are limited to learning dataset biases on a variety of benchmarks. In practically all of the suggested benchmarks, they discovered that these baseline models were as accurate as these MLBSFs, suggesting that these models simply pick up on biases in the dataset. Researchers will be able to thoroughly examine MLBSF performance and ascertain how dataset biases impact their predictions thanks to the authors tests and the platform they provide, ToolBoxSF.
The readers may find this link useful: https://github.com/guydurant/toolboxsf
Reference:
Durant G. et. al. (2025) Robustly interrogating machine learning-based scoring functions: what are they learning? Bioinformatics 41 (2): btaf040