GenoMAS
A code-driven multi-agent framework for automated gene expression analysis workflows.
- Source repo
- Liu-Hy/GenoMAS
- Stars
- ★ 130
- Last updated
- 15d ago
- License
- MIT
- Primary language
- Python
- FA score
- 50/100 · Major gaps
At a glance
- How it runs
- Works with
- Universal · cross-platformOpenAI API · Claude API
- Cost
- Free software; you pay for model usage
- Setup effort
- Medium · a few setup steps
- You'll need
- Typical use
- A computational genomics researcher needs repeatable preprocessing and regression runs across many GenoTEX trait-condition pairs.
- Not a fit if
- Teams without GPUs: local large models need multiple high-end GPUs
- Users wanting a quick full-workflow test: full benchmark needs ~42 GB, 3-5 days, $300+
- Teams without transcriptomic data or bioinformatics background
- Source review
- 50/100 · Major gaps
What does this agent do, and when should you use it?
GenoMAS combines a minimal multi-agent framework for scientific workflow automation with a specialized implementation for gene expression analysis. Its typed messaging and notebook-style workflow let agents plan, write code, execute it, debug, and backtrack across multi-step tasks. The specialized system analyzes transcriptomic datasets from GEO and TCGA to identify trait-related significant genes while accounting for confounders. Experiments run from the command line and write GenoTEX-structured preprocessing data, regression results, and logs under output/. Adoption requires local input data, a Python environment, and at least one provider credential, although local models can be run through Ollama.
Experiments are launched with python main.py, with configuration such as --version, --model, --api, and --data-root. The system reads GenoTEX input data, performs cohort preprocessing and regression analysis, and writes preprocessed cohort datasets to output/preprocess/{trait_name}/ and regression results to output/regress/{trait_name}/. Separate model and API-index settings are available for the Code Review, Domain Expert, Data Engineer, Statistician, and planning roles, while --parallel-mode cohorts can process cohorts concurrently. With --generate-action-units, it generates Action Unit prompts from agent guidelines, pauses for optional manual editing by default, and can proceed automatically with --non-interactive. It records activity in ./output/log_{version}.txt and supports checkpoint resume when the same interrupted command is rerun.
- A computational genomics researcher needs repeatable preprocessing and regression runs across many GenoTEX trait-condition pairs.
- A bioinformatics team working with GEO or TCGA transcriptomic data wants to identify trait-associated genes while accounting for confounders.
- A research engineer wants to evaluate heterogeneous model assignments for code review, domain expertise, data engineering, statistics, and planning.
- A team wants to validate cohort preprocessing on a few datasets before committing to the full statistical workflow, using --quick-test.
- A user constrained by API rate limits wants to reduce wall-clock time by processing a controlled number of cohorts with --parallel-mode cohorts.
How do you install or deploy this agent?
Create the documented environment and install dependencies:
conda create -n genomas python=3.10
conda activate genomas
pip install -r requirements.txtCopy the configuration template and provide at least one model-provider API key:
cp env.example .envOpenAI models also require an organization ID. Download the input data separately and place it under the data root; the documented default is ../data.
How do you use this agent?
Validate downloaded data first:
cd download
python validator.py --data-dir /path/to/data --validateThen run a basic experiment from the project root:
python main.py --version exp1 --model gpt-5-mini-2025-08-07 --api 1Add --data-root /path/to/data when the data is elsewhere. Use --quick-test for a small preprocessing-focused run. A local Ollama model can be selected with, for example, --model llama3.1.
What are this agent's strengths and limitations?
- It pairs a general multi-agent scientific-workflow framework with a concrete implementation for GEO and TCGA transcriptomic analysis.
- Its notebook-style execution model explicitly includes planning, code writing, execution, debugging, and backtracking instead of only producing recommendations.
- It supports both global model selection and role-level overrides for Code Review, Domain Expert, Data Engineer, Statistician, and planning.
- Its documented output layout, human-readable logs, and checkpoint resume make lengthy benchmark runs easier to inspect and continue.
- The full GenoTEX benchmark requires roughly 42 GB of input data; the documented estimate is 3–5 continuous days and more than $300 for all pairs.
- Cloud use requires at least one provider API key, OpenAI use additionally requires an organization ID, and parallel runs can stress API rate limits.
- Local deployment of large models has substantial GPU-memory and latency demands; the README states that DeepSeek-R1 671B requires multiple high-end GPUs.
- Full regression requires all related datasets to be preprocessed, so the documented quick-test path validates preprocessing rather than the complete analysis workflow.
How does this agent compare with similar options?
Key facts side by side with the most closely related agents.
| Agent | Source review | Form / cost | Stars | Updated | Language | Full support on |
|---|---|---|---|---|---|---|
| GenoMAS This agent | 50 · Major gaps | CLIFree + model costs | ★ 130 | 15d ago | Python | OpenAI API · Claude API |
| GenoTEX Gene Expression Benchmark | 54 · Major gaps | CLIFree | ★ 65 | 5mo ago | Jupyter Notebook | — |
| BambooAI Data Analyst | 70 · Some gaps | Web appFree + model costs | ★ 842 | today | Python | OpenAI API · Claude API |
| MiroFish Offline | 40 · Major gaps | Self-hosted serviceFree | ★ 2.7k | 7mo ago | Python | OpenAI API · Claude API |
How does FollowAgents rate this agent?
Why each dimension lost points
The code shows role-specific clients, an input data root, output directory, and log path; the README says API keys must be configured. The packet does not adequately explain what data is sent to external models, how keys are protected, or dependency risks, so least privilege, data flow, sensitive-data handling, and dependency security are only partial. Action Unit generation has interactive confirmation, while non-interactive mode adopts generated content automatically. Interruption recovery clears unfinished task outputs, with limited detail about recovery boundaries. The README names paper authors and provides citation details, supporting attribution.
The README claims checkpoint resume and gives timeout adjustment guidance, offering some failure-handling clues. The supplied main-program excerpt does not show checkpoint implementation or consistent error handling, and dependencies are unpinned; no other files corroborate these areas, so scores remain partial.
The README identifies gene-expression analysis and describes GEO/TCGA, model choices, role-specific configuration, and parallel cohort processing. However, generic framework support is mainly asserted, while the shown code remains tied to a specific scientific workflow; boundaries and invocation conditions are not precise.
The README has clear sections, setup steps, multiple examples, troubleshooting, cost estimates, and limitation notes; the MIT license text is complete. The dependency list is unpinned and no changelog is shown. Issues, pull requests, and author contact provide a maintenance path, though responsibility is described only generally.
The documentation describes usable preprocessing and regression outputs, logs, and checkpoint recovery, and gives full-run and small preprocessing examples. Benchmark and discovery claims are README assertions without independent support in this packet. A full run takes days and costs over $300, though a smaller path is offered.
The README points some claims to a paper, benchmark, output documentation, and configuration code, and provides runnable command examples. The packet does not include the paper, benchmark files, test evidence, or key implementation details to corroborate performance, recovery, or scientific-discovery claims. The distinction between documented facts and author claims is limited, so scores are not higher.
- A run requires about 42 GB of data and external model API keys; the full benchmark is estimated to take 3–5 days and cost over $300.
- Non-interactive mode automatically uses generated Action Unit prompts; review prompts and input data, especially when data may be sensitive.
- Performance, checkpoint recovery, and scientific-discovery claims are not independently corroborated in the supplied material; dependency versions are also unpinned.
Maintain this project? Add the FARS badge to your README
[](https://followagents.com/en/agents/genomas)<a href="https://followagents.com/en/agents/genomas"><img src="https://followagents.com/badge/genomas.svg" alt="FARS review"></a>FAQ
What does a full run cost and how long does it take?
Is a cloud API mandatory?
Can an interrupted experiment resume?
What data and outputs are in scope?
Thanks — we use this to improve the page.