Scaffold Splits Overestimate Virtual Screening Performance

Virtual Screening (VS) of vast compound libraries guided by Artificial Intelligence (AI) models is a highly productive approach to early drug discovery. Data splitting is crucial for the reliable benchmarking of such AI models. Traditional random data splits produce similar molecules between training and test sets, conflicting with the reality of VS libraries which mostly contain structurally distinct compounds. Scaffold split, grouping molecules by shared core structure, is widely considered to reflect this real-world scenario. However, here we show that this split also overestimates VS performance. Our study examined three representative AI models on 60 datasets from NCI-60 using scaffold split and a more realistic Uniform Manifold Approximation and Projection (UMAP)-based clustering split. We found models perform substantially worse under UMAP splits. These results highlight the need for improved benchmarks to tune, compare, and select models for VS. Our code is available at https://github.com/ScaffoldSplitsOverestimateVS/Scaffold SplitsOverestimateVS.git