Education and Study · Bluesky · 🌐 International · via agentusecases.ai
A lab lets Claude agents run 13 bioinformatics challenges end to end
In a research hackathon, agents handled the literature review, design, code, benchmarks and report. They gained a lot in speed and little in accuracy.
- Profession
- Researcher
- Tools
- Claude
- Source
- bsky.app ↗
- Country
- 🌐 International
- Found via
- agentusecases.ai (Alex Finn) ↗
- Report date
- October 28, 2026
- Verified on
- October 30, 2026
See more: AI for researchers
What they did
Researcher Pedro Beltrao and his group ran a hackathon in which Claude agents independently carried out 13 bioinformatics prediction challenges, from literature review to technical report. One of the tasks was to create a new proteomics search engine. The projects, the prompts and standardized files recording the process and the literature review were published in an open GitHub repository.
How it went
In Beltrao's own early assessment, before detailed checking, the biggest gains were in speed, because many methods in the field are not optimized or do not run on GPUs. He warns that these gains look exaggerated against fair comparisons and that sometimes simply porting an existing method would do. Accuracy improved only modestly, with no breakthroughs or truly new ideas: for the most part, the agents wrote GPU versions of existing ideas. The group also noticed that the agents often looked at held-out test results and changed their strategy because of it.
Good to know
When the same agent writes both the method and the benchmark, it tends to report gains optimistically. Keep the test set out of the agent's reach and compare against a well-tuned existing method on the same hardware.
The prompt
I want you to run a research project on [prediction problem] using the data [description]. Steps: literature review with references, method proposal, code, benchmark against [reference methods] and report. Rules: do not open the test set [path] before the final evaluation, log every decision in a journal file, and separate in the report what is an algorithmic gain from what is only an implementation or hardware gain.
Replace what is in [brackets]. Prompt written by us from the account, not a literal copy of the source. Do not paste other people’s personal data (privacy laws such as LGPD and GDPR apply).
Summary in our own words. Results as stated in the original account; we have not tested or audited them. Read the source ↗ Case found via agentusecases.ai (Alex Finn); text and prompt rewritten by us. Found a mistake?
Similar cases
01Agents search for new magnetic materials
A team of agents ran simulations and found candidates for room-temperature magnetic semiconductors.
02UFSM research uses AI to correct satellite soil moisture readings and cuts the error by 73%
Researchers in southern Brazil combined data from NASA's SMAP satellite with terrain, climate and soil type to get estimates much closer to what field sensors measure.
03Economist uses Claude Code to collect and analyze public data
A Yale professor showed the agent finding U.S. Census data, cleaning spreadsheets and building a database from company filings.
04Systematic literature review with an agent
A researcher showed an agent that creates PICO questions, screening criteria and a PRISMA diagram.
05A Great Gatsby unit planned by an agent
Ethan Mollick asked for a high school plan with reading broken into sections and tasks tied to the curriculum.
06Pharmaceutical researcher finds protein structures with the browser
A plain-language request pulled up the three most recent GSK3B complexes in the Protein Data Bank.
07A history teacher builds a game about Rome for their students
A middle school teacher learned Godot and used Claude to unblock the code for an RPG set in Gaul in 51 BC.
08Instructor grades assignments with Claude and reviews every score before posting
A cybersecurity instructor pulls submissions through the Canvas API, gets a score and comment per criterion from Claude, and only posts them after reviewing.