I have known Derya Unutmaz, the immunology professor at The Jackson Laboratory with a huge following on X, for a couple of years now.

Back when I was still doing the longevity-focused media platform LEVITY, we had him on the podcast. A month ago, at the longevity conference Vitalist Bay in Berkeley, I got to meet him in person for the first time. He’s a lovely guy.

We also have a lot in common. Above all, we are both convinced that AI will change things for the better, especially in health. We both want aging to be solved, the biggest challenge there is in biology, and that can only be accelerated with more intelligence in the world.

Why am I telling you all of this?

Right after Vitalist Bay, we decided the conversation should continue. So that is what we have started doing. The simplest way to describe it is to call it a video podcast, but I don’t have time right now for some grand YouTube launch or to make it available on every podcast platform.

The first episode was published only on X, because that is where Derya and I spend our time anyway. Future episodes might become a BAIO subscriber perk. We’ll see.

Whatever the platform, the point is that we are going to talk AI × bio a few times a month.

Another reason I’m telling you this is that this issue, completely by coincidence, includes a news story featuring Derya. OpenAI published a case study about how GPT-5 Pro helped him revisit a T-cell mystery his lab had shelved.

I think it also captures a lot of the optimism Derya and I share. In broader circles, AI is too often treated as a threat to be endured. But week after week, BAIO is tracking the other story: scientists using AI to read hidden biology, design new medicines, and take on problems that once looked out of reach.

And this issue has plenty of them: Nabla using AI to design antibodies against cancer signals normally hidden inside cells, Proto trying to make biology’s AI tools work together, AI-CURA automating the slow evidence review behind genetic variant interpretation, NVIDIA building a toolbox for biology agents, and Absci moving an AI-designed antibody through early human testing.

That is the BAIO beat in one issue: not vague futurism, but the week-by-week evidence that biology is becoming more programmable, more computational, and maybe, if we are lucky, more solvable.

Antibody fragments with drug-like properties have been designed by AI to target KRAS mutation signals displayed on cancer cells. KRAS is one of cancer’s classic hard-to-drug proteins.

The model, JAM-2, was developed by Nabla Bio, a Cambridge, Massachusetts startup trying to make antibody design programmable. “It's the intracellular proteome, a holy grail for drug discovery. AI is letting us design more sophisticated medicine with greater logic and precision,” says CEO Surge Biswas, in a video accompanying a new technical report.

Antibodies normally work outside cells, but much of the biology driving cancer happens inside them. Cells do offer a workaround: they cut up internal proteins and display tiny fragments on their surface, almost like molecular status flags. Even so, it has been extremely difficult to design a drug precise enough to read the right flag without attacking healthy cells.

That is what Nabla says JAM-2 has done. Starting only from the target sequences - the cancer-linked peptide plus the MHC molecules that display it - the model designed compact antibody fragments against five cancer-related signals shown on the outside of cells. The team then turned selected designs into T-cell engagers: molecules with one part that reads the cancer signal and another that summons a killer T cell.

“Across the five targets about 1 in 100 designs bound on the first attempt. These binders had exquisite selectivity, each one locking on to its intended peptide and ignoring everything else. That selectivity is what separates a safe drug from a dangerous one,” says Katherine Warfel, protein engineer at Nabla, in the video.

Taking on KRAS is no mean feat. G12V, G12C and G12D are three cancer-linked KRAS mutations. G12V differs from normal KRAS by just that one amino acid, so a drug has to hit the mutant signal without touching the healthy one. One JAM-2 design recognized G12V and G12C, but spared normal KRAS and G12D. In a lab test using primary human immune cells, it killed G12V-presenting cells at extremely low concentration, 0.07 nM. Cryo-EM, a high-resolution imaging method, then showed that the real molecule landed almost exactly where the model predicted - within 0.93 Å, roughly the width of a single atom.

It’s important to remember this is a technical report from the company, not a peer-reviewed paper, and the results are still lab assays, not animal or human data. The next challenge is whether the designs stay selective across the full human protein landscape and remain safe in the body.

Why it matters: If this holds up, AI antibody design moves closer to individualized cancer drugs. A patient’s tumor has its own mutation pattern, and some of those mutations can appear as fragments displayed on that patient’s HLA molecules. A sequence-driven model could, in principle, design a drug-like antibody for that exact cancer signal instead of hoping a screening library already contains the right binder.

Did you know? Nabla Bio is hiring.

Proto, from Arc Institute and Stanford, is a design framework and infrastructure layer for generative biology. Its job is to let researchers describe a biological design goal, connect the AI models needed for the task, and search for DNA, RNA or protein sequences that best satisfy the goal.

Researchers trying to use more than one AI x bio tool quickly run into the same problem: one model predicts structure, another designs sequences, another scores gene regulation, and each arrives with its own dependencies and input formats. “The first week, sometimes two or three, of every biological design project I worked on went to setup,” says Ben Viggiano, one of Proto’s co-first authors, in Arc’s accompanying interview.

Proto turns that setup work into reusable infrastructure. In the web interface, a design campaign becomes a graph: what sequence do you want to build, which model or method should propose candidates, which scores should judge them, and how should the system search for better versions?

The point is not only convenience. A new preprint from Arc and Stanford reports early experimental evidence that Proto can make biological AI models more useful together than they are on their own.

Why it matters: AI biology is certainly not short of models (just look at the Edge section in our last issue). It is short of reliable ways to combine them. Proto could become the common language biological design needs - before it can become systematic.

Did you know? Proto is open source and wraps more than 120 tools.

AI-CURA, a new LLM workflow described in Science Translational Medicine, can automate much of the evidence review that tells geneticists whether a DNA variant is harmless, uncertain or likely to cause disease.

Developed by Wei Ma and colleagues at the Hong Kong Genome Institute, the system works at the interpretation step, after sequencing has surfaced a candidate variant that needs classification.

Some rules are already machine-readable, such as how rare the variant is or whether software predicts damage to a protein or splicing. What still slows human variant curators down is the evidence buried in papers: affected patients, family inheritance, functional experiments, and whether the patient’s symptoms fit the gene.

AI-CURA used LLMs to read papers and apply seven literature-dependent ACMG rules, part of the clinical genetics framework used to classify variants.

With AI model DeepSeek-R1, AI-CURA matched human curators’ exact final classification on 144 of 150 ClinGen variants in a selected test set. In clinical terms, it matched them on 149 of 150, because pathogenic and likely pathogenic variants often point to the same diagnostic answer.

Why it matters: If AI-CURA holds up in broader testing, more rare-disease variants could be classified faster, and old unsolved cases could be rechecked as new papers appear.

Did you know? Code and model-version details are archived on Zenodo.

AI agents in life sciences have a new way to call biology and chemistry tools.

NVIDIA launched BioNeMo Agent Toolkit at the BIO International Convention in San Diego. It packages parts of NVIDIA’s BioNeMo stack as agent-callable skills: structure prediction, molecular generation, docking, sequence analysis, genomics, protein design and biomarker discovery. The idea is that Claude, Codex or an in-house pharma agent can choose the right model, format inputs, run the computation and interpret the output.

“Frontier models are the brains. BioNeMo is the scientific toolbox. Together, they give AI agents the skills of a PhD research assistant and the speed of a supercomputer,” said Jensen Huang, founder and CEO of NVIDIA.

NVIDIA says more than 50 companies and research groups are using or integrating the toolkit, including Lilly, Natera, Benchling, Thermo Fisher, Anthropic and OpenAI.

Why it matters: NVIDIA does not just want to power AI biology with GPUs. It wants BioNeMo to become the software layer that AI biology runs through, the place where agents call models, run workflows and turn scientific prompts into computation.

Did you know? BioNeMo Agent Toolkit is available now through NVIDIA developer resources and GitHub.

GPT-5 Pro helped immunologist Derya Unutmaz revisit a T-cell experiment his lab had shelved, according to a new OpenAI case study.

Unutmaz, a professor at The Jackson Laboratory and the University of Connecticut, had been stuck on a result from 2022. His lab was asking how glucose metabolism shapes T-cell fate. Developing T cells can become different immune cell types, and metabolism helps steer that choice.

To probe that, the lab compared two interventions. One was low glucose, which gives cells less fuel. The other was deoxyglucose, a glucose-like molecule that enters the same pathway but blocks normal glucose metabolism. The expectation was that both would push the cells in a similar direction.

They did not. Deoxyglucose pushed developing T cells much more strongly toward Th17, an inflammatory T-cell state involved in infection, autoimmunity and cancer. That suggested the blocker was doing something more specific than simply starving the cells.

After GPT-5 Pro arrived in 2025, Unutmaz uploaded the data and asked for analysis. The model pointed to IL-2, an immune-signaling protein that can restrain Th17 development. Its hypothesis: deoxyglucose may interfere with IL-2 production, removing a brake on Th17 formation.

“GPT‑5 came up with this really remarkable insight that retrospectively, makes perfect sense,” Unutmaz told OpenAI. Neither Unutmaz nor anyone else in the lab had made the connection.

Unutmaz also says GPT-5 Pro correctly predicted an unpublished experiment in which CD8+ T cells gained a stronger ability to kill lymphoma cells.

Why it matters: This is one lab story from an OpenAI case study. But it shows where frontier models may help scientists most: connecting a strange result to a plausible mechanism just far enough outside the lab’s mental map to be missed.

Did you know? BAIO first covered Derya Unutmaz way back in Issue 1, when he used OpenAI Codex to build a 90,000-mutation cancer catalog.

Absci has reported early human safety data for an AI-designed antibody against androgenetic alopecia, the common hair-loss condition.

The Vancouver, Washington biotech, says its antibody ABS-201, which targets the prolactin receptor, was well tolerated in the first single-dose cohorts of its Phase 1/2a HEADLINE trial. The readout covers 32 healthy adults given IV doses or placebo.

As of June 8, Absci reported no serious adverse events. The antibody’s estimated half-life was at least 65 days, supporting the company’s hope that future dosing could mean two or three injections over six months.

The trial has now moved into multiple-dose testing in people with androgenetic alopecia. Interim proof-of-concept hair data are expected in the second half of 2026, with fuller data in early 2027.

Why it matters: Absci is trying to show that generative antibody design can produce a real clinical candidate. ABS-201 is still early, but it is now being tested where AI-designed drugs eventually have to prove themselves: in humans.

Did you know? Absci also raised $100 million to advance ABS-201 in hair loss and endometriosis, with Lilly among the investors.

BioMedArena is an open-source toolkit for testing biomedical deep-research agents without paying the “per-paper engineering tax” each time. It packages 155 canonical benchmarks and 76 tools across areas such as medicine, biology and chemistry, then lets researchers plug in models, tools or agent setups through small adapters. Use it when you want to compare agents on the same evaluation surface instead of trusting one-off leaderboard claims.

☑️ Sakana AI released Fugu, a multi-agent orchestration model.

☑️ GenBio AI advances agentic AI for building virtual cells – powered by NVIDIA.

☑️ Thermo Fisher showcased new AI-enabled research at BIO International 2026.

☑️ Gero raised $17 million to advance its AI-driven aging drug-discovery platform.