CLaSP: A Contrastive Learning-Guided Latent Scoring Platform for Comprehensive Drug-Likeness Evaluation

Efficient drug-likeness evaluation is critical for accelerating drug discovery and reducing the costs of early stage compound screening. However, existing approaches either rely on rigid empirical rules or supervised classification models, which lack generalizability and interpretability. Here we introduce contrastive learning-guided latent scoring platform (CLaSP), a novel framework that integrates variational autoencoding with triplet contrastive learning for drug-likeness assessment. CLaSP constructs a structured latent space informed by both physicochemical and ADMET-related features, enabling a continuous, interpretable CLaSP_Score that reflects molecular developability. The feature set was curated from ADMETlab 3.0 and admetSAR 3.0 and refined via feature selection. Benchmark evaluations demonstrated that CLaSP outperformed QED and DBPP-Predictor across multiple data sets and real-world case studies. Furthermore, CLaSP effectively captured drug optimization trajectories, as shown in a case study of Wee1 inhibitors. A user-friendly web portal (https://lmmd.ecust.edu.cn/CLaSP) supports single and batch analyses for early drug design.