Overview
STELLA is a biomedical research agent framework intended to connect scientific questions with literature searches, database queries, and computational analyses. Its research design addresses the limitations of fixed tool collections by combining an evolving template library with an expandable tool pool. The supplied sources describe runnable repository code and a hosted browser interface, rather than only a paper proposal. The research paper presents the underlying self-evolving approach, while the current README describes additional workflows and case-study scripts.
A user supplies a scientific objective. A manager decomposes it and retrieves relevant workflow templates; a development agent performs analyses and searches; a critic examines results and suggests improvements. An optional tool-creation agent supplies new tools when existing ones are insufficient. Templates are structured workflow instructions reused or created from successful runs, not separate autonomous agents. Documented outputs include ranked gene candidates with evidence summaries and rationales, and enzyme-variant proposals with scores and supporting explanations; case scripts save CSV and text files.
STELLA should be evaluated as research software, not treated as an independently validated discovery service. Local execution requires an OpenRouter API key, with additional search and literature services optional. Models are configurable, and supplementary biomedical resources are distributed separately. The README explicitly notes that gene-ranking outputs can change with model stochasticity and literature-index updates. Its case studies and benchmark materials provide starting points for reproduction, but reported discoveries and experimental outcomes remain source claims. The repository’s code license does not establish terms for hosted access, external models, or separately downloaded resources.
Key Features
- Manager, development, and critic agents coordinate objective decomposition, bioinformatics execution, evidence retrieval, and result refinement.
- A template system retrieves and applies structured workflows and creates new templates from successful runs.
- An optional tool-creation agent extends the tool pool when predefined tools do not cover a task.
- The documented tool pool supports literature searches, database queries, and virtual-screening workflows.
- The NK-cell/AML case study ranks gene candidates using literature novelty, mechanistic evidence, NK-cell expression, and AML context, with structured evidence fields.
- The Strictosidine Synthase case study uses prior HPLC screening feedback for ESM re-scoring, FoldX stability filtering, and prioritized variant proposals.
Use Cases
- Suggested evaluation: reproduce the NK-cell/AML candidate-ranking workflow and inspect whether each rationale is supported by the retrieved literature.
- Suggested evaluation: examine feedback-guided Strictosidine Synthase variant prioritization, comparing proposed candidates and explanations with the supplied case-study outputs.
- Suggested evaluation: compare biomedical question-answering workflows with template retrieval enabled or disabled using the repository’s benchmark materials.
- Suggested evaluation: investigate whether optional tool creation supplies useful missing functionality for a bounded literature or bioinformatics task.
How to Use
Read the official README and paper to distinguish the agent architecture from the repository’s later case-study descriptions. Choose a narrowly scoped research question and define what evidence would make its output useful.
Choose the hosted interface or the source deployment described in the repository. The README also links a Docker package; follow its documented setup rather than assuming equivalent environments.
For local use, follow the README’s dependency instructions and configure an OpenRouter key privately. Decide whether optional search keys and supplementary biomedical resources are needed for your task.
Select the model configuration and whether to use template retrieval or optional tool creation. Submit the scientific objective through the documented Gradio interface or programmatic entry point.
For an intended reproduction evaluation, use the README’s case-study or benchmark links. Inspect saved results, verify supporting literature, and record model settings and retrieval timing before comparing runs.