Compact AI plans multi-step drug discovery workflows locally
Toward a Locally Deployable Agentic Co-Scientist: Small-Model Planning for Early-Stage Drug Discovery
Artificial Intelligence
Summary
Drug discovery involves using many different scientific tools in a sequence, which can be complicated to coordinate. The authors created a small AI model that can run on local computers and plan how to use these tools step-by-step. They trained the model to generate plans that follow a shared molecular data format and tested it to see how well it picks tools and sequences the steps. Their results show the model works well when handling known tasks but struggles with completely new step sequences, pointing to future improvements needed.
What this means in practice
- •For drug discovery software teams: Build locally deployable AI planners to coordinate multiple computational tools in early-stage drug development workflows.
- •For chemical engineering teams: Integrate and automate multi-tool workflows for molecular design using a small, customizable AI model.
Authors
Tian Liang, Jiayu Chang, Alejandro F. Frangi, Mobarak I. Hoque, Richard A. Bryce
Abstract
Early-stage computational drug discovery requires coordinating heterogeneous scientific tools across multi-step workflows. We present a lightweight, tool-augmented framework in which a locally deployable compact language model plans calls to 18 modular tools. A Unified Molecular Schema maintains shared molecular records, while a plug-in interface supports tool replacement and extension. We construct 1,263 manually refined query-plan pairs through workflow-graph path coverage and apply LoRA fine-tuning to three compact model families. Under the query-level split, all fine-tuned models generate fully parseable and schema-compliant plans on 47 held-out cross-group queries. Llama 3.2-3B achieves a tool-selection F1 of 0.998, sequence exact match of 0.979, and argument F1 of 0.960. Under the stricter workflow-grouped split, which excludes identical ordered tool sequences across partitions, sequence exact match reaches 0.452 to 0.548, highlighting the remaining difficulty for compact models in generating complete workflow paths unseen during training. These results demonstrate the feasibility of compact, locally deployable planning while identifying compositional generalization as an important direction for further improvement.