KANEL: Kolmogorov-Arnold Network Ensemble Learning Enables Early Hit Enrichment in High-Throughput Virtual Screening

A team from the University of North Carolina at Chapel Hill, including Pavel Koptev and Alexander Tropsha, has introduced KANEL (Kolmogorov-Arnold Network Ensemble Learning), a novel machine learning framework designed to revolutionize the early stages of drug discovery. The system addresses a critical bottleneck in high-throughput virtual screening: accurately prioritizing a tiny fraction of promising compounds from libraries containing millions. KANEL's innovation lies in its ensemble approach, which strategically combines the emerging, interpretable architecture of Kolmogorov-Arnold Networks (KANs) with established powerhouse models like XGBoost, random forests, and multilayer perceptrons. Each model is trained on complementary molecular representations—including LillyMol descriptors, RDKit-derived descriptors, and Morgan fingerprints—to capture diverse aspects of chemical structure and potential bioactivity.

The key performance shift championed by KANEL is its focus on "early hit enrichment" metrics, such as Positive Predicted Value calculated for the top N candidates (PPV@N). This is a more practical and actionable benchmark for drug hunters than traditional global metrics like Area Under the Curve (AUC). By optimizing for the accuracy of its very top predictions, the workflow ensures that the compounds selected for costly and time-consuming wet-lab experiments have the highest possible likelihood of being true "hits." This methodologically rigorous ensemble aims to reduce failure rates downstream, potentially accelerating the identification of lead compounds and saving significant R&D resources. The preprint, shared on arXiv, represents a meaningful step toward more efficient and interpretable AI-driven molecular design.