Dr. Seal is currently the principal agentic AI scientist at Human Chemical. Dr. Srijit holds a PhD at the University of Cambridge in the area of chemoinformatics specializing in machine learning, image-based profiling and their applications in drug discovery and chemical safety. Before joining Human Chemical, he undertook a postdoctoral fellowship at the Broad Institute of MIT and Harvard where he focused on defining best practices for AI-driven toxicity prediction from Cell-Painting. Here, he also played an active role within the HESI funded OASIS consortium whose goal is to assess the combination of Cell Painting, transcriptomics and proteomics in a variety of cell models for safety assessment. Next, he served as Senior AI/ML Scientist at Merck in the US. Dr. Seal is also the developer of open-source tools such as PKSmart for pharmacokinetic prediction and DILIPredictor for liver toxicity detection. Beyond his research, Dr. Seal serves on the Board of Directors of the American Society for Cellular and Computational Toxicology (ASCCT) and on the Editorial Board of the Journal of Cheminformatics. He also holds academic affiliations at the University of Cambridge, Babeș-Bolyai University in Romania and the Uppsala University in Sweden.
OpenTox Summer School 2026
Machine learning for toxicity prediction is often limited by data and tools that reduce complex weight-of-evidence judgments to binary labels. Drug discovery decisions are rarely a single prediction but a choice about which assay is decisive, requiring retrieval, reasoning, and hypothesis generation rather than a probability alone. This talk presents how agentic systems equipped with perception, computation, action, and memory tools are well suited to this decision layer. We present three case studies across drug-induced renal injury, ICH M7 mutagenicity assessment, and the curation of in vivo liver findings from repeat-dose toxicity reports. Complementary work on Cell Painting and multi-omic profiling shows that biological descriptors augment chemistry-only models, while also exposing persistent data confounders. The overarching conclusion is that agentic reasoning amplifies whatever the underlying tools encode, making tool quality the binding constraint on trustworthy in silico toxicology.