ClawBio presentationsUK DRI at UCL · 21 September 2026
CLAWBIO / TRUSTWORTHY AGENTIC GENOMICS 01 UK DRI / 21 SEPTEMBER 2026 Working with AI without giving up scientific judgement ClawBio: two practical exercises Manuel Corpas UK DRI Core Informatics, UCL
CLAWBIO / TRUSTWORTHY AGENTIC GENOMICS 02 THE EVERYDAY RESEARCH PROBLEM We connect the work. Keep your attention on the science. Data ›_ Pipelines Literature Database Figures ›_ Scripts
CLAWBIO / TRUSTWORTHY AGENTIC GENOMICS 03 WHAT AN AGENT DOES Give a goal. Check each step. COMMANDS A B C D OBJECTIVE Goal Plan Act Check Adapt Stop or ask when needed
CLAWBIO / TRUSTWORTHY AGENTIC GENOMICS 04 THE TOOL WE WILL USE ClawBio A skill library for agentic biology Agents reason. Skills constrain. Tools compute. RNA DNA QC Cells
CLAWBIO / TRUSTWORTHY AGENTIC GENOMICS Source: repository README 05 READ THE SKILL A method you can inspect. SKILL.md v1.0 Specification ›_ Code Example inputs Checks + provenance
CLAWBIO / TRUSTWORTHY AGENTIC GENOMICS 06 EXERCISE 1 / PUBMED Build a briefing. Open the papers. 01 Start with the handout, then try APOE Use the pubmed-summariser skill; save the report. 02 Open a cited abstract and check the claim Does the source support the sentence you would reuse? 03 Inspect the search and summary method An abstract excerpt and an AI summary are different. A gene search is a starting point, not a systematic review. Exercise 1: workshop resource page
CLAWBIO / TRUSTWORTHY AGENTIC GENOMICS 07 EXERCISE 2 / DIFFERENTIAL EXPRESSION Interpret the table you actually have. 01 Run the mouse tau example together Check ranked genes, fold changes and adjusted p-values. 02 Inspect the APOE3 versus APOE4 export Check the column names before accepting the result. 03 Compare the renamed-column version The data are the same. The input labels changed. Biological themes are suggestions, not pathway enrichment. Exercise 2: workshop handout
CLAWBIO / TRUSTWORTHY AGENTIC GENOMICS 08 A FAILURE TO LOOK FOR The input changes the meaning. Ask what was filtered before this table. THE HANDOUT'S EXAMPLE REPORTED Every listed gene is significant A Prefiltered table Known B Batch effect? Not shown ? Full analysis Check NEXT: verify the upstream filtering
CLAWBIO / TRUSTWORTHY AGENTIC GENOMICS Workshop practice: retain inputs and outputs 09 KEEP AN EVIDENCE TRAIL Keep enough to check the answer later. Inputs Method ›_ Execution Outputs EVIDENCE What ran What changed What failed Replayable ≠ scientifically valid
CLAWBIO / TRUSTWORTHY AGENTIC GENOMICS Source: data-handling.md; workshop guidance 10 BEFORE USING YOUR OWN DATA Keep the task and data bounded. Scope Data Budget Approval AGENT Stop / ask Use the supplied exercise data first.
CLAWBIO / TRUSTWORTHY AGENTIC GENOMICS 11 YOUR PRACTICAL CHECKLIST Run it. Challenge it. Keep the evidence. 01 Choose one claim in the output Point to the exact paper or table entry supporting it. 02 Find one assumption or failure Check missing columns, filtering and contrast direction. 03 Record what you would change Keep the prompt, output and your correction together. Online and in the room: share one useful result and one concern. Follow the two handouts on the workshop resource page.
CLAWBIO / TRUSTWORTHY AGENTIC GENOMICS 12 NOW TRY IT What would make this output trustworthy? One useful result. One claim to check. github.com/ClawBio/ClawBio