Agentic Orchestration of Drug Discovery ML Tools Under Partial Observability

Sara Masarone, Katie Victoria Beckwith, Matthew Mason, Thomas Clelford, Arran Willmott, Layla Hosseini-Gerami

Keywords: Agentic workflow, LLM, drug discovery, ML

TL;DR: To address the fragmentation of drug discovery tools, we introduce an orchestrating agent that dynamically coordinates specialised AI modules to provide adaptive, real-time safety analysis even when data is incomplete.

Abstract:

Over 90% of drugs fail in clinical trials, often due to unanticipated safety issues, resulting in substantial financial losses and missed therapeutic opportunities. We develop models that uncover mechanistic bases of observed safety liabilities and support rational drug modification, including an asset-sourcing agent, a cheminformatics module predicting off-target interactions, and a bioinformatics module linking off-targets to toxicity pathways.

Individually, these tools are performant but outputs are fragmented, and not all input data are always available, limiting rapid decision-making. To address this, we introduce an orchestrating agent that dynamically coordinates tool execution based on data availability, task context, and uncertainty. The agent selectively invokes, sequences, or defers modules to enable adaptive analysis under partial information. We present its architecture and early testing, illustrating a framework to unify a fragmented AI ecosystem into a coherent, agent-driven system.

Track: Tiny paper track (up to 4 pages)

AI Policy Confirmation: I confirm that this submission clearly discloses the role of AI systems and human contributors and complies with the ICLR 2026 Policies on Large Language Model Usage and the ICLR Code of Ethics.

Submission Number: 92