Agentic biology, checked at the bench
I do wet-lab genomics in the Skene Lab at Imperial and build agents for it. Agents can now plan and run whole analyses, and the hard part has become knowing whether to believe them. This is where I keep what I have tried and read, and the tools I am building.
Skills
- Wet lab
- iPSC microglia, prime editing, CETCh-seq, TIP-seq, scRNA-seq
- Code
- Python, R, TypeScript, Snakemake, FastAPI
- Agents
- Claude Code, OpenCode, MCP, agent skill libraries
- Infrastructure
- Imperial CX3 HPC, Docker, SQLite, Postgres
Projects
See all projectsBiology and agents
- ClawBioContributorA local-first bioinformatics skill library for AI agents, where many analyses write a reproducibility bundle.PythonOpenClawCPICscVIPyAging
- pegRNA pipelineAuthor, private repoAutomated pegRNA design for prime-editing C-terminal protein tagging.PRIDICT 2.0PythonSnakemakeImperial CX3 HPC
- flair-adAuthorA two-page companion site for the lab's IRF1 work, with a public explainer and a data dashboard.D3.jsGSAPFirebaseNetlifyPlaywright
- A day at the benchAuthorAn interactive page that squeezes weeks of waiting at the bench into one day you can play through.
Other agent work
- hearth-aiAuthor, private repoAn agentic London rentals search that crawls, filters, scores and alerts.FastAPISQLite WALCloudflare WorkersClaude HaikuClaude Sonnet 4.5
- Automated tradingAuthorA configurable CFD trading bot with a deterministic risk layer in front of every order.PythonOpenAI ResponsesTrading212IGyfinance
Open questions
Problems I am working on and would like company with. If one of these is yours too, email me.
How do you check an agent's analysis without redoing it by hand?
Unreviewed draftBioinformaticsWet lab
An agent can hand back a full analysis quickly, and the Cell Genomics review argues that checking it is now the hard part. Redoing it by hand defeats the point. Reading only the summary can miss quiet errors, such as the wrong reference or a filter that was dropped. I want to know which checks catch the most for the least effort. Help would look like a checklist you use, an example of an agent mistake you caught and how you caught it, or a paper that measures this.
What actually predicts prime-editing efficiency at a new site?
Unreviewed draftWet labML
Tools such as PRIDICT 2.0 predict prime-editing efficiency. I want to understand how well any efficiency predictor transfers to a new site, a new cell type or a different edit. Local chromatin, cell type and the edit itself could each matter, and I do not know which do. Finding out would take predictions set against measurements across sites and cell types. Papers that make that comparison would help, as would your own experience of which features held up and which did not.
Which proteins tolerate a C-terminal tag, and how would you know in advance?
Unreviewed draftWet lab
A C-terminal tag works for many proteins, but it can break function when the end of the protein carries something important, such as a localisation signal or an interaction motif. Finding out usually means making the edit and testing, which costs weeks. I would like a way to flag the risky ones before starting. Rules of thumb, sequence or structure based predictors, and published datasets that compare tagged and untagged protein function would all help.
What would a benchmark for agentic genomics have to measure?
Unreviewed draftMLBioinformatics
Papers on agentic systems report wins on benchmark datasets. A benchmark that scores only the final answer says little about whether the analysis was right for the right reasons. A better one might also score the plan, the code, the intermediate outputs and whether the agent flagged its own uncertainty. Help would look like your view on what belongs in such a benchmark, and pointers to existing ones that already do some of this.
Recently in the library
See all library items- paperAgentic genomics: from pipeline automation to autonomous validationA good first read on what agentic genomics means. Its argument is that the bottleneck has moved from building pipelines to validating what agents produce. It is a review, so it will not tell you how any one system performs on your data.Unreviewed draftNo background neededBioinformaticsMLAdded 21 October 2026
- toolAgentomicsThe claim to check is that it beat human expert solutions on a number of benchmark datasets. Performance on benchmarks does not always carry over to messier data, so I would want to see how it does on a new dataset before relying on it. Most interesting to people who build models from biomedical data and want an automated baseline.Unreviewed draftMLBioinformaticsAdded 21 October 2026
- agentAutoBAMinimal user input is the appeal and also the risk. Someone still has to check the plan and the output, and that checking is the validation problem the Cell Genomics review puts at the centre.Unreviewed draftBioinformaticsAdded 21 October 2026
- agentBio-CopilotThe interesting part is that human-agent interaction is built into the design. A person sits inside the analysis loop. Worth a look if you want an agent to help with large omics analyses without handing over the whole thing.Unreviewed draftBioinformaticsAdded 21 October 2026
- agentCellAtriaAn agentic framework from AstraZeneca that ingests and standardises single-cell RNA-seq data, with a chat interface in front of a pipeline called CellExpress. Worth a look if you work with single-cell data and want to see how an agent can handle the ingestion and standardisation step.Unreviewed draftBioinformaticsWet labAdded 21 October 2026
- agentClawBio skill libraryI contribute to this project, so weigh my view accordingly. It gives agents bioinformatics-specific skills to work from. The local-first design is the thing to check if your data is sensitive.Unreviewed draftBioinformaticsAdded 21 October 2026