Guidelines for evaluating structure-based methods for protein-ligand binding prediction

Scoring functions built with advanced deep learning algorithms have recently gained prominence in protein-ligand pose generation and binding affinity prediction for computational drug design. However, a number of shortcomings make it increasingly difficult to evaluate scoring functions for a specific task across studies. Transparent disclosure of training data, code and procedures is essential to permit reproducibility, reuse and independent scrutiny. Guarding against data leakage, sampling bias and flawed design is equally important. Furthermore, meaningful comparisons of scoring functions require common data splits, checks for run-to-run variability and performance baselines. Here we propose a set of guidelines to help the community identify the most promising methods for these important problems.