This paper introduces PhenoProfiler, an innovative, end-to-end model designed for advancing phenotypic learning in image-based drug discovery. Unlike existing multi-step procedures, PhenoProfiler directly processes whole-slide multi-channel images into low-dimensional quantitative representations, enhancing efficiency and overcoming limitations like computational expense and error accumulation. The model utilizes a unique architecture, including a gradient encoder and a transformer encoder, and incorporates a multi-objective learning module combining classification, regression, and contrastive learning for robust feature extraction. Extensive evaluations on large-scale public datasets demonstrate that PhenoProfiler outperforms state-of-the-art methods in biological matching tasks and shows superior robustness against batch effects. Furthermore, the tool includes a phenotype correction strategy to emphasize relative phenotypic changes, making it a scalable and generalizable asset for high-throughput drug screening, available as both a Python package and a web server.

References:

• Li, Bo, et al. "PhenoProfiler: Advancing Phenotypic Learning for Image-based Drug Discovery." arXiv preprint arXiv:2502.19568 (2025).