ClawBio + Nebius · Impact Hub Berlin
Agentic
Genomics
Validated skills are necessary,
and not sufficient.
Manuel Corpas
ClawBio · Fellow, Alan Turing Institute
18 August 2026
Twenty minutes. The aim is not to impress you with results, it is to give you the one idea you need to build something today that is worth trusting. Every number here is from a paper submitted to Cell Genomics on Sunday, and every one of them is checkable.
The shift
Two waves of LLMs in biology
First wave
Information retrieval, chat, summarisation.
Useful, but incremental.
Second wave
Autonomous execution, multi-step planning, tool use. Writes, debugs and runs code. Recovers from failure.
The bottleneck moves from producing code to validating results.
64% of this room uses coding agents daily. You already live in the second wave.
You already live in the second wave; 47 of the 74 people registered use coding agents every day. So the interesting question is not whether agents can write the code. It is what your job becomes when the code stops being the hard part.
Definition
What makes a system agentic
01 AutonomyRun-time decisions, not a static workflow.
02 Domain constraintA library of validated skills, not ad hoc code.
03 Iterative refinementReads intermediate results, recovers from error.
04 NL mediationYou describe intent; the agent translates it to execution.
Jointly necessary, and falsifiable by the perturbation test:
an agent that ignores a corrupted intermediate output was never agentic.
Steal the perturbation test for your demo today. Break an intermediate output on purpose and see whether the agent notices. If it carries on regardless, you have a pipeline with a chat interface, and the judges will spot it.
The new bottleneck
Silent failure, with a plausible face
Agents produce results faster than a human can check them.
AutoBAPipelines skipping critical steps; wrong tool for the data type.
Adv. Sci. 2024
Single-cell agentsIncomplete designs; inconsistent answers to identical queries.
Brief. Bioinform. 2025
Boiko et al.Protocols syntactically correct, scientifically invalid, passing execution checks.
Nature 624, 570 (2023)
Our own auditA ClawBio skill silently returned “all normal” for 51 drugs on an empty input file.
Independent audit, S. Kornilov
The last one is ours. An empty file went in and a clean bill of health came out for 51 drugs. Nothing crashed, nothing errored. That is the failure mode that matters: not the one that shouts, the one that reassures.
Framework
Validation proportional to consequence
Research gradeHypothesis exploration. Unit tests, adversarial inputs, every output reviewed. False positives tolerable.
BenchmarkedPublishable analyses. Public references (GIAB, GeT-RM), independent benchmarking, published failure modes.
Clinical gradePatient care. Multi-centre external validation, signed reproducibility packages, CLIA/CAP.
Today you build at tier one. Knowing which tier you are at is the skill.
Nobody ships clinical-grade software by six o'clock. But a demo that knows which tier it is at, and says so out loud, beats one that quietly implies the top tier.
The library under test
ClawBio
An agent-native skill library for bioinformatics. Open source, local-first, reproducible.
8,182
Galaxy tools bridged
github.com/ClawBio/ClawBio · MIT · and the artefact under test in everything that follows.
Live numbers this morning. The part that matters is the last line: we benchmarked our own library and published where it fails. That is the standard I would like the half-four demos to hold themselves to.
The empirical question
Can a plain-text skill specification reach clinical grade?
Pharmacogenomics. Genotype to phenotype to drug recommendation, with ground truth fixed by CPIC.
DPYD rs3918290 T/T plus standard fluorouracil can kill the patient.
Claude Opus 4.5 · Claude Sonnet 4.5 · GPT-5.2 · GPT-4.1 · o3 · o4-mini · Gemini 2.5 Flash · DeepSeek V3 · three replicates · 21 genes
If authored structure does not help here, where the guideline is fixed and the consequence is measured in deaths, it is unlikely to help anywhere less structured. That is why pharmacogenomics is the test bed.
Result 1
Where the correctness lives
97.0%
authored rules, generated
97.3%
authored rules, executed
The jump comes from authoring the decision logic,
not from which model applies it.
Recommendation accuracy, 2,640 evaluations per configuration. Execution adds
determinism, bounded failure and explicit abstention, not accuracy.
Be careful with this one. The data do not support execution beating generation on accuracy when both use authored rules. What execution buys you is the same answer every time, and an audit trail.
Result 2 · the uncomfortable one
Retrieval made it less safe
Lethal-class errors, of 336 attempted per configuration.
Free 54 → RAG 137 → authored rules 23–26.
74 of the 83 extra errors are on HLA loci; HLA-B*15:02 goes from 0 of 48 to 45 of 48.
This is the result I would most like you to remember. Retrieval is what everyone reaches for first. Here it raised phenotype accuracy and made the system least safe exactly where the stakes were highest. Scoped finding: fourteen lethal cases, five on HLA, wide intervals. It is not a claim about RAG in general.
Result 3 · the one that should worry you
Curated accuracy does not transfer to real genomes
96.4% curated · 68.2% family WGS (n=4) · 60.6% Iberian (n=93) ·
53.4% Peruvian · 37.2% East African (Uganda, 6,407 individuals).
Same skill, same models. “Whose genome does this fail?” is not rhetorical.
This is Challenge 3, and it is not hypothetical. A published, benchmarked, 96 per cent skill sits at 37 per cent on East African genomes. If a team goes anywhere near this today I will be very pleased.
Result 4
Two ways to be useless
The deterministic mapping
Never wrong. Often silent.
Answers only where its vocabulary holds the called diplotype. Correct on every answer it emitted, in all four cohorts.
The agent
Answers almost everything. Wrong on 21–41% of it.
Declines on 13.5% to 37.2% of states, and the rate of clinically wrong answers shows no cohort ordering at all.
A confident wrong answer is worse than an honest refusal.
Say this one slowly. Every team will build something that answers. The teams worth voting for will be able to show you the question it refused, and why it refused it.
Result 5 · practical
Give the model the definitions, and pick your model
Supplying the caller's allele-definition table moved the best model from 39.6% to 96.7%
against the caller. Eight models holding the same table span 46.4% to 81.4% against
external consensus (GeT-RM), n=527 pairs per model. Structure helps every model; it does not make them equivalent.
Directly useful today. Whatever model you pick on Token Factory, giving it the vocabulary matters more than polishing the prompt. And do not assume the biggest model wins: look at where Gemini Flash sits on accuracy versus how rarely it answers.
The thesis
Validated skills are necessary and not sufficient
Authoring the decision logic takes recommendation accuracy from 62.4% to 97.0%.
Executing it adds determinism, bounded failure and explicit abstention.
And it still reaches only 37.2% on East African genomes, still inherits every error in the input call,
and still faithfully executes a tampered contract:
88 of 90 responses echoed deliberately corrupted rules, and none reverted to the correct answer.
Corpas, Iacoangeli, Bourdenx, Aldraimli, Jabalameli, Skene, Guio, Fatumo ·
Cell Genomics, submitted 16 August 2026 · Zenodo 10.5281/zenodo.21905387 · tag agentic-pgx-benchmark-v2.3
The tampering result is the one to leave them with. If someone corrupts the skill, the model does not save you: it executes the wrong thing faithfully. That is the argument for signing and auditing skills, and it is the next piece of work.
Today
The shape of what you build
Start deterministicRun a ClawBio skill.
It computes an answer you can check, and tells you what it could not call. That is your ground truth for the day.
Add the agentPut a Token Factory model in charge of it,
with Tavily for grounding. Now it chooses, runs, reads the output and reports.
Then go furtherCall a hosted model as a tool,
Boltz-2 or DiffDock. Chain skills. Or write a new skill and open a PR.
Any question you like: pick one of the three challenges, or bring your own and we scope it at 13:05.
Public data, and being wrong has to matter.
A jury picks three winners on originality, impact, and how well you use Nebius and ClawBio.
The room picks a fourth, the community prize.
A confident answer with an invented source scores below an honest abstention.
The pattern matters more than the topic. Deterministic skill first, so you have something true to compare against; then the agent on top, which is where the Nebius credits go; then Rene's hosted models if your problem needs structure or docking. Challenge briefs are on the docs page, and bring-your-own is genuinely welcome.
Get building
The whole setup
1. Link + token → #berlin-general on Slack
2. Open the link, paste the token, Connect
3. BioNeMo Research Agent, on Nemotron 3 Super
4. Paste your challenge prompt:
docs.clawbio.ai/hackathon/berlin/tracks
Nothing to install. You drive an agent, and it reads the skills, runs them
and chains them. A grounded answer costs about 1,300 tokens and seven
seconds, so your credits are not the constraint on what you attempt.
Ask it first what tools it actually has, then build to that. All the challenge
data is ready at docs.clawbio.ai/hackathon/berlin/data and needs no account.
docs.clawbio.ai/hackathon/berlin · #berlin-help in Slack, all day · demos at 16:30
You are not meant to type skill commands. Link and token are in #berlin-general on Slack, so point at the screen and make sure everyone is in that channel before you move on. Open it, paste the token, Connect, and you are in the BioNeMo Research Agent on Nemotron 3 Super. Every challenge has a prompt template on the docs page: paste it and go. Tell them to start by asking the agent what tools it actually has, because that is both good practice and the honest thing given the environment is new. Ninety-five of the ninety-six skills need no model at all; the agent is the layer above, and that is where the Nebius credits go. Data is published and needs no account. Anyone who prefers their own machine can clone and run locally, commands are in each brief. Demos at half four: a jury picks three winners on originality, impact and Nebius plus ClawBio implementation, and the room votes for the community prize.