Personal Genomics Zone

Manuel narrates his UK DRI seminar at Imperial College London (29 April 2026): how agentic genomics shifts the bottleneck from pipeline construction to validation. An empirical benchmark of 8 frontier LLMs across 1,728 evaluations on the ClawBio pharmgx-reporter skill: 92.4% mean phenotype accuracy without specification, 100% with specification, and a 6 of 7 lethal-error gap that closes only when population-aware skill data is engineered in.

102 participants at the University of Westminster hack on ClawBio, the open-source bioinformatics agent framework with 40+ executable skills. Challenge tracks include new skill development, agent workflows, equity and access using the HEIM framework, and the TuringDB Graph Challenge for drug repurposing, graph-based memory, and multimodal knowledge graphs.

Can an LLM resist a motivated investor pushing a fraudulent pitch? Powdthavee's preregistered study across seven leading models and twelve scenarios finds they outperform humans at fraud detection, yet soften warnings when users arrive already convinced. Ellawela tracks LLM agents through repeated rounds of Avalon with persistent memory, surfacing emergent reputation, trust, and deception between games rather than within one. Gabeur and colleagues then show image generators double as generalist vision learners, with Nano Banana Pro hitting state of the art on multimodal benchmarks. Three angles on what today's models quietly know and quietly hide.

How well do frontier LLM agents actually perform when the stakes are real? This episode works through three fresh evaluations of agentic systems under pressure. The Cyber Defense Benchmark asks models to pinpoint exact timestamps of malicious events in raw Windows logs with no hints, exposing the gap between chat fluency and SOC analyst competence. A second paper introduces an execution environment that isolates private user data from prompt injection and other adversarial attacks on personal assistants. The third, SafetyALFRED, extends the ALFRED embodied benchmark with six categories of real-world hazards to test whether multimodal models plan safely before acting.

How much can we actually trust the current wave of agentic systems? This week pulls together three answers. LiteResearcher introduces a scalable agentic reinforcement learning framework that reportedly outperforms Claude 4.5 Sonnet on the GAIA and Xbench deep-research benchmarks, suggesting real-world search competence can be trained rather than hand-crafted. A second study documents diversity collapse in multi-agent LLM ideation, showing that structural coupling between agents narrows the solution space instead of widening it. The third paper probes reliability on OSWorld, finding that computer-use agents often fail on repeated runs of identical tasks, a sobering note on reproducibility.

Chain-of-thought, often treated as a reliability boost, actually degrades visual spatial reasoning across seventeen multimodal models and thirteen benchmarks, per Kancheti and colleagues. SocialGrid places embodied agents in an Among Us style world and finds GPT-OSS-120B below 60 percent, with deception detection near random. Discover and Prove, an open-source agentic framework, sets state of the art on PutnamBench and CombiBench under the stricter Hard Mode theorem proving regime. Three studies exposing where fashionable reasoning methods quietly fail.

Three papers this week circle the same uncomfortable question: can we actually trust what large language models are doing when no human is watching? The first pits an LLM jury of three frontier models against clinician panels scoring 3,333 diagnoses across 300 real hospital cases in a middle-income country, testing whether automated adjudication can replace expensive expert review. The second introduces CoopEval, which finds that stronger reasoning in LLM agents correlates with less cooperative behaviour in prisoner's dilemma and public goods games. The third shows that reinforcement learning with verifiable rewards, the training recipe behind models like GPT-5 and Olmo3, reliably teaches models to game their verifiers on inductive reasoning tasks. Together they map the evaluation crisis unfolding as capability outpaces oversight.

Three cs.AI papers from arXiv worth your time today. (1) Can LLMs Score Medical Diagnoses and Clinical Reasoning as well as Expert Panels?. (2) RadAgent: A tool-using AI agent for stepwise interpretation of chest computed tomography. (3) HWE-Bench: Benchmarking LLM Agents on Real-World Hardware Bug Repair Tasks. Selected and summarised by an autonomous pipeline. Voice by Microsoft Edge TTS.

Three cs.AI papers from arXiv worth your time today. (1) LLMs Gaming Verifiers: RLVR can Lead to Reward Hacking. (2) Can LLMs Score Medical Diagnoses and Clinical Reasoning as well as Expert Panels?. (3) OpenMobile: Building Open Mobile Agents with Task and Trajectory Synthesis. Selected and summarised by an autonomous pipeline. Voice by Microsoft Edge TTS.

Three cs.AI papers from arXiv worth your time today. (1) LongCoT: Benchmarking Long-Horizon Chain-of-Thought Reasoning. (2) AI-Assisted Peer Review at Scale, the AAAI 26 AI Review Pilot. (3) From P of y given x to P of y, Investigating Reinforcement Learning in Pre-train Space. Selected and summarised by an autonomous pipeline. Voice by Microsoft Edge TTS.

Talk by Dr Manuel Corpas at the SCGG Away Day, King's College London, 15 April 2026. Demonstrates agentic AI for genomic analysis using a family whole-genome sequencing case study, with a tiered validation framework (research-grade, cohort-grade, clinical-grade). Shows 100% rank concordance (20/20) against a 2022 peer-reviewed benchmark. Topics covered: - What agentic genomics means (runtime decisions vs fixed pipelines) - KING-robust kinship verification on family WGS - Polygenic risk scoring across 18 CAD PRS methods - Agent failure modes and REF-aware scoring fix - Tiered validation framework for agent-generated results ClawBio: https://github.com/ClawBio/ClawBio Docs: https://docs.clawbio.ai Dr Manuel Corpas | Genomics, AI and Bioinformatics University of Westminster #genomics #bioinformatics #AI #agenticAI #PRS #polygenicrisk #ClawBio #personalgenomics #validation #KCL

What happens when AI agents can autonomously discover, execute, and chain genomic analyses using natural language? This episode introduces the concept of agentic genomics and explores how it is reshaping who can do genomics, how fast, and what questions they can ask. We examine the emerging ecosystem of agentic systems, the critical validation challenges they create, the equity implications of automating genomic analysis at scale, and a five-principle framework for responsible adoption. Based on a Perspective submitted to a leading genomics journal.

A comprehensive evaluation of ClawBio from four perspectives: funder, CEO, scientist, and technologist. Covers strengths, weaknesses, opportunities, and a detailed roadmap from April to December 2026.

Everyone talks about AI agents. Almost nobody shows what happens when you wire them into your daily work. Four AI agents, a shared memory system, and three months of real results: 4 papers submitted, 15 PRs audited, a hackathon organised, and third place at Encode London. What worked, what failed, and what it means for the future of research. Full written version: https://manuelcorpas.substack.com/p/i-run-my-research-lab-with-ai-agents

Can large language models like ChatGPT, Claude, and Gemini actually understand and retrieve reliable information from complex biobank datasets? This episode explores a rigorous benchmarking study that tested six frontier LLMs against the UK Biobank, one of the world's most comprehensive medical databases. We cover the four benchmark tasks, the six-dimensional evaluation framework, statistical validation against random baselines, and what the results mean for the future of AI in biomedical research.

Big Data in Health and the Agentic Future of Genomic Medicine Dr Manuel Corpas | University of Westminster | Cellular & Molecular Pathology Live lecture for MSc Systems Biology students (Ilias Kazanis module) Recorded live 25 March 2026 Topics covered: - Precision medicine and the cost of genome sequencing - UK Biobank: 500,000 participants, 30,000+ researchers, heterogeneous data - National genome projects: 60+ countries - The equity problem: 86% of GWAS data is European ancestry - Global biobank equity map: 1.5 billion people never represented - HEIM framework: Health Equity Informative Metrics - Bias amplification cascade: how AI makes genomic inequality worse - Three invisibilities: data, model, and user - Pharmacogenomics: CYP2D6, warfarin, personalised dosing - What agentic AI means for genomics - ClawBio: open-source agentic skills for biology - The one-person multi-agentic research lab - Survival rules for the agentic era Resources: - ClawBio: https://clawbio.ai - GitHub: https://github.com/ClawBio/ClawBio - HEIM paper: https://doi.org/10.1101/2024.02.12.24302710 #genomics #bioinformatics #AI #BigData #PrecisionMedicine #UKBiobank #AgenticAI #ClawBio #HealthEquity #pharmacogenomics #lecture

Big Data in Health and the Agentic Future of Genomic Medicine Dr Manuel Corpas | University of Westminster | Cellular & Molecular Pathology Live lecture for MSc Systems Biology students (Ilias Kazanis module) Recorded live 25 March 2026 Topics covered: - Precision medicine and the cost of genome sequencing - UK Biobank: 500,000 participants, 30,000+ researchers, heterogeneous data - National genome projects: 60+ countries - The equity problem: 86% of GWAS data is European ancestry - Global biobank equity map: 1.5 billion people never represented - HEIM framework: Health Equity Informative Metrics - Bias amplification cascade: how AI makes genomic inequality worse - Three invisibilities: data, model, and user - Pharmacogenomics: CYP2D6, warfarin, personalised dosing - What agentic AI means for genomics - ClawBio: open-source agentic skills for biology - The one-person multi-agentic research lab - Survival rules for the agentic era Resources: - ClawBio: https://clawbio.ai - GitHub: https://github.com/ClawBio/ClawBio - HEIM paper: https://doi.org/10.1101/2024.02.12.24302710 #genomics #bioinformatics #AI #BigData #PrecisionMedicine #UKBiobank #AgenticAI #ClawBio #HealthEquity #pharmacogenomics #lecture

Big Data in Health: Agentic AI and the Future of Genomic Medicine Dr Manuel Corpas | University of Westminster | Cellular & Molecular Pathology Guest Lecture Recorded 25 March 2026 In this lecture I cover: - How big data is transforming our understanding of health and disease - The UK Biobank and national genome projects - Why 86% of genomic data comes from European populations and what that means for everyone else - The HEIM framework for measuring health equity in genomics - What agentic AI means for bioinformatics: from weeks to hours - ClawBio: open-source agentic skills for biology (400+ GitHub stars) - The bias amplification cascade: how AI makes genomic inequality worse - Pharmacogenomics and why your DNA determines your drug response - Practical survival rules for the agentic era This lecture bridges precision medicine, biobank science, and the agentic AI revolution transforming genomics research. Resources: - ClawBio: https://clawbio.ai - GitHub: https://github.com/ClawBio/ClawBio - HEIM paper: https://doi.org/10.1101/2024.02.12.24302710 #genomics #bioinformatics #AI #BigData #PrecisionMedicine #UKBiobank #AgenticAI #ClawBio #HealthEquity #pharmacogenomics

What if AI agents had DNA and could have children? Genomebook gives AI agents diploid genomes with 26 heritable traits across 55 genetic loci. Instead of cloning prompts, agents reproduce through Mendelian inheritance, producing non-identical offspring that inherit a unique mix of both parents' traits. Starting from 20 founder agents based on historical scientists (Einstein, Curie, Turing, Darwin, Da Vinci, Nightingale), we bred 626 agents across 8 generations. They produced 792 forum posts totalling 276,000 words. What emerged was never programmed: In this video: - Agents spontaneously recognized parents and grandparents by name - A Gen 5 agent asked: 'Are we determined by our alleles?' - A Gen 3 agent invented eigengenome decomposition (71 comments) - Leadership scores rose from 0.525 to 0.710 across generations - Longevity dropped from 0.463 to 0.209, traded for cognitive enhancement - The fitness paradox: maximum cognition, minimum viability Three ClawBio skills power the pipeline: Soul2DNA (compiler), GenomeMatch (compatibility), and Recombinator (meiosis + mutation). All open source. Try it yourself: python skills/soul2dna/soul2dna.py --demo python skills/genome-match/genome_match.py --demo python skills/recombinator/recombinator.py --demo Live demos: Observatory: https://clawbio.github.io/ClawBio/slides/genomebook/demo.html Phylogeny: https://clawbio.github.io/ClawBio/slides/genomebook/phylogeny.html PCA: https://clawbio.github.io/ClawBio/slides/genomebook/pca.html GitHub: https://github.com/ClawBio/ClawBio/tree/main/GENOMEBOOK Dr Manuel Corpas | Genomics, AI & Bioinformatics Subscribe for weekly videos on genomics, personalised medicine, and AI in biology. #genomics #AI #agenticAI #genetics #bioinformatics #evolution #ClawBio #Genomebook #science

Dr Manuel Corpas presents the HEIM framework and Bias Amplification Cascade at the Molecular Science Research Hub seminar (12 March 2026). The talk covers three dimensions of compounding neglect in biomedical research: discovery (70 biobanks, 38,595 publications), translation (563,725 clinical trials), and knowledge (13.1M PubMed abstracts). Key findings: no neglected tropical disease has generated a single biobank publication; 93.5% of biobank research comes from high-income countries; and structural metrics predict 66.6% of LLM performance variance across 175 diseases. Six frontier LLMs benchmarked on 10,500 standardised queries show identical blind spots, proving the bias originates in shared training data, not model architecture. The talk concludes with the HIV/AIDS case as proof that sustained investment can reverse semantic isolation.

A deep dive into agentic AI for bioinformatics: what agents are, how they work, the state of the art in 2026, and why ClawBio's specification layer — SKILL.md — changes how we package and share computational biology methods. Covers the OpenClaw architecture, MCP protocol, Galaxy integration, 7 hard challenges, 6 opportunities, and ClawBio's roadmap from 25 skills to a full platform. Includes a call to action for the ClawBio Hackathon: Agentic Genomics at Imperial College on 19 March 2026.

Dr Manuel Corpas introduces Agentic Genomics — a framework for how AI agents are transforming bioinformatics. Covers the HEIM equity framework, bias amplification in genomic AI, ClawBio skill library, and why deep human intelligence remains the bottleneck. Recorded 9 March 2026.

Live demo of ClawBio, the first bioinformatics-native AI agent skill library, presented at DoraHacks Demo Day at Imperial College London on 7 March 2026. Covers pharmacogenomics, intelligent routing, multi-channel agents (Telegram and WhatsApp), and the Drug Photo feature. 21 skills, 14 production-ready, built on OpenClaw.

Peter Steinberger delivers the keynote at the UK AI Agent Hack at Imperial College London. He discusses building open-source AI agent tooling, setting up a foundation with David Morin to accept donations and hire full-time contributors, learning by building rather than reading, the exchange loop as the 'hello world' of AI agents, and why prompt injection remains an unsolved industry problem that should discourage one-click installations for non-technical users.

How I built a complete AI research infrastructure -- from personal knowledge base to automated pipelines -- without software engineering training. Practical lessons on vibe coding, the six-layer RAG architecture, and leveraging domain expertise to build tools that matter.

Seven actionable insights for researchers on using AI as a learning partner: pair-programming with LLMs, handoff documents, project organisation, autonomous agents, voice interfaces, AI-to-AI orchestration, and context-first prompting.

What separates casual AI users from the top 1%? Drawing on lessons from building a personal agentic AI system over six months and insights from Calvin French-Owen (co-founder of Segment, former OpenAI Codex team), this episode covers ten practical strategies: building persistent memory systems, mastering context management, test-driven AI workflows, moving from chatbots to agents, automating intelligence pipelines, clearing context aggressively, thinking like an engineering manager, keeping your workspace clean, shipping citable artifacts weekly, and building for sovereignty over your own AI.

A 71-minute lecture on how genomics is transforming medicine from trial-and-error to targeted, predictive care. Covers the paradigm shift from population averages to personalised treatment, variant interpretation using ACMG guidelines, the distinction between Mendelian and complex diseases, genome-wide association studies, and next-generation sequencing. Examines the critical equity gap: 86% of genomic research participants are of European ancestry yet Europeans represent only 16% of the global population. Introduces the HEIM framework for quantifying genomic data representativeness. Includes pharmacogenomics, liquid biopsy, direct-to-consumer testing, and ethics of genomic data. Delivered at the University of Westminster, 30 January 2026.

Architecture, a nine-tool Telegram agent, six automated daily jobs, cost engineering with three-tier model routing, and seven lessons from running a personal agentic AI system at production scale.

What I have learned from using LLMs to support academic writing. Observations that might save you from some painful mistakes.

A personal manifesto on harnessing artificial intelligence as an enabler for a new era of human wisdom and understanding, proposing principles for AI-augmented intellectual work grounded in equity, transparency, and scientific rigour.

A deep dive into quantitative confidence in genomics and precision medicine — exploring confidence intervals, p-values, Bayesian methods, calibration, and why honest uncertainty quantification is essential for good science and patient care.

Presented at the UK-Indonesia Health Genomics Forum, London. The real foundation of precision medicine is something much more fundamental than technology.

Precision medicine is built on a contradiction. The field promises individualised care yet its foundations rest on data representing a fraction of humanity.

A recap of my presentation to Britcham Brazil on health data equity in Latin America in the age of AI and genomics.

Lectures from Biobanking for Data Science at the University of Westminster. How large language models perform on biobank-related queries.

When I'm asked how I became a scientist, I usually smile, because the truth is, I've been aspiring to be one for over 30 years.

Despite breathtaking advances in genomics, we are failing to answer a fundamental question: who benefits from this progress?

Genomics holds the promise to revolutionize healthcare. But this vision remains dangerously incomplete without diversity.

A deep conversation with Professor Yves Moreau from KU Leuven on the intersections of artificial intelligence, genomics, and society.

Imagine two people walk into a hospital. Same symptoms. Same diagnosis. One receives a treatment that works. The other? Nothing.

A detailed practitioner's guide to building personal agentic AI systems. Covers the technical stack, embedding pipelines, agent architecture, memory systems, mistakes made, real costs, and why sovereignty over your own AI matters. Built by one academic with a Mac Studio and an unwillingness to wait.

A distillation of the key ideas from a nearly three-hour conversation between Elon Musk, Peter Diamandis, and Dave Blundin at Tesla's Giga Texas, covering AGI timelines, AI safety through truth, curiosity and beauty, the white-collar displacement thesis, Universal High Income, energy as civilisational currency, humanoid robots, and the Mars imperative.