This issue is, it turned out, about barriers dropping.

☑️ A 24-year-old redesigns the best-selling cancer drug from a web browser.

☑️ An OpenAI engineer rewrites AlphaFold2 so that any ML engineer can read it on a laptop, arguing that the illegibility of protein AI code is holding back the whole field.

☑️ A breast cancer risk platform runs on hardware built with “Radio Shack parts”.

☑️ OpenAI hands its frontier model to every US clinician for free.

☑️ And a Stanford spinout goes after the characterization bottleneck that sits between AI-designed drug candidates and the lab.

We often say that the tools in AI x biology are getting more powerful. But as we’ll see below, they’re also getting into more hands.

Also in this issue: James Zou's research becomes a billion-dollar company, and Curve Biosciences teaches a DNA model to read the epigenome.

Let's dive in.

James Zou is raising roughly $100 million at a billion-dollar valuation for a startup called Human Intelligence, Bloomberg reports. The company will build a “physiology foundation model” - AI designed to model human biological processes. Zou did not comment.

The portfolio behind the raise is well-validated. EchoNet, his cardiac AI, outperformed human sonographers in a blinded trial and was cleared by the FDA. A Nature-published system called Virtual Lab put AI agents to work designing antibody fragments against COVID-19 - two of 92 candidates held up when tested in the wet lab. And Virtual Biotech turned 37,000 agents loose on the records of 56,000 clinical trials.

Why it matters: AI has models for proteins (AlphaFold, ESM) and, more recently, models that compress millions of patient records into clinical predictions (like APOLLO). A “physiology foundation model” sounds like it would sit in the gap between them - modeling how organs and biological systems actually behave. That middle layer is largely unmapped. Zou's research portfolio is among the strongest in the field, but “physiology foundation model” remains, for now, a phrase, not a product.

Did you know? Zou is one of BAIO's most-covered researchers - Virtual Biotech (Issue 4), Eubiota (Issue 5), the Generalist Biological AI review (Issue 10), and Cerebra (Issue 11), have all featured in previous issues. The Virtual Lab is open source on GitHub.

AI can now design protein drug candidates at a pace that would have been unthinkable five years ago. But before any of those candidates can advance to testing, they have to be physically characterized - molecular structure verified, chemical modifications mapped, behavior understood. That step relies on mass spectrometry (a technique that reveals those structures and modifications by measuring how molecule fragments behave in an electromagnetic field), scarce human experts, and weeks of manual analysis.

10x Science, spun out of Nobel laureate Carolyn Bertozzi's lab at Stanford in December 2025, raised $4.8 million in an oversubscribed seed round led by Initialized Capital with backing from Y Combinator. And the company name is the pitch: make characterization ten times faster, using AI.

“You can add as many candidates as you want to the top of the funnel, but they all have to pass through this characterization process,” CEO David Roberts told TechCrunch. “Everything needs to be measured.”

The platform pairs chemistry and biology algorithms with AI that interprets spectrometry data - delivering what the company says are automated, explainable molecular insights in minutes where current workflows take months. The “explainable” part is especially important. Drug characterization is as much about showing your work as getting the right answer, and a step towards regulatory approval. The 10x Science platform surfaces the reasoning behind each identification: which fragments matched, at what coverage, with what error.

“The biggest constraint I see across the field is the gap between the data we can generate and the insights we can extract,” said Bertozzi in a press release. “The field has outgrown its infrastructure.”

Matthew Crawford, a scientist at Rilas Technologies who has been using the platform, told TechCrunch it surprised him by identifying a protein from just the filename and autonomously pulling the right sequence data. “It makes reasonable assumptions,” he said - something he attributed to the deep domain expertise of its creators.

Why it matters: BAIO has covered the steep drop-off between computational predictions and lab results. Characterization is part of that bottleneck. Faster, auditable characterization hopefully means more candidates reaching the lab sooner.

Did you know? 10x Science is hiring.

A transparent sliver of silicone, bonded to a glass slide with simple electrodes. That's the device. Breast cells flow through it in liquid, hitting a section narrower than the cell itself. Electrodes measure the current as each cell squeezes through and how long it takes to spring back to its original shape. Stiffer cells transit more slowly. Older-acting cells recover more slowly. A machine learning algorithm trained on these responses assigns each cell a “mechanical age” - how old it acts physically, regardless of the person's actual age.

Researchers at UC Berkeley and City of Hope trained this system on cells from 18 women. On a separate group of women it had never seen, the model correctly ranked older cells above younger ones 91% of the time. It then correctly flagged elevated risk in cells from women with BRCA1/2 mutations, family history, and cancer in the other breast - without being trained on any high-risk samples. The findings are published in Lancet’s eBioMedicine.

Only 5-10% of breast cancer cases involve known genetic mutations. For the rest, no non-genetic test flags individual risk. And most biological aging measures work at the population level - useful for research, less so in a clinic. MechanoAge works per person, per cell.

The authors call this a proof of principle. The cohort is small, ancestral diversity is limited, and the cells were measured outside the body. Clinical validation on fresh samples is still ahead. But they also note that since all cells contain the structural filaments driving these mechanical changes, the approach could extend beyond breast tissue to other aging-related diseases.

Why it matters: MechanoAge reads risk from the physical properties of normal-looking cells, before a tumor forms. If it validates, it could bring individual-level breast cancer screening to the vast majority of women who currently have no way to assess their risk - using a cheap device built from what one of the researchers called “Radio Shack parts.”

Did you know? Engineers have measured mechanical aging in metals, concrete, and polymers for a long time. This is reportedly the first time it has been quantified in living cells.

OpenAI has released ChatGPT for Clinicians - a version of ChatGPT designed specifically for clinical work, available free to any verified physician, nurse practitioner, physician assistant, or pharmacist, starting in the United States. It includes access to OpenAI's current frontier models, a clinical search function with real-time cited answers from peer-reviewed sources, and the ability to turn repeatable workflows like referral letters and prior authorizations into reusable templates.

Before launch, physician advisors tested 6,924 conversations across clinical care, documentation, and research. OpenAI says 99.6% of responses were rated safe and accurate. On a new open benchmark called HealthBench Professional - 525 physician-authored tasks spanning care consults, writing, and medical research - GPT-5.4 in the clinician workspace scored 59.0. Human physicians given unlimited time and web access scored 43.7. No other model came close.

A caveat worth noting though: OpenAI built the benchmark and tested its own models on it.

Why it matters: This is the first time a frontier AI lab has made its most capable model free for an entire clinical profession. The American Medical Association reports physician AI use has more than doubled since 2023. OpenAI is betting that clinicians already using ChatGPT informally will upgrade to a version with citations, privacy controls, and optional HIPAA compliance - and that the resulting data and feedback will make the product better.

Did you know? Yes, you probably did know, but still: OpenAI also released GPT-5.5. The company says the new model shows meaningful gains in scientific research workflows, including drug discovery. It also launched a Bio Bug Bounty alongside GPT-5.5, offering $25,000 to anyone who can find a universal jailbreak that defeats the model's biological safety guardrails - a sign that OpenAI considers its models' biology capabilities both a selling point and a risk.

DNA language models learn patterns from sequences of A, T, C, and G. But chemical modifications to those letters - particularly methylation - regulate which genes are active in which cells. Most genomic AI misses this because it's trained on bare DNA.

Curve Biosciences, a Stanford spinout that raised $40 million last year, found a way to add epigenetic awareness to an existing DNA model without changing its architecture. The method exploits bisulfite sequencing, a lab technique that chemically converts unmethylated cytosines to thymines - effectively writing methylation status into the letter sequence itself. Train a DNA model on this converted data and methylation state becomes visible in the structure of its internal representations. In a paper presented at International Conference on Learning Representations (ICLR) on April 27, the team shows this separation sharpens when comparing tumor tissue to healthy cells.

Why it matters: The paper shows you can add an entirely new biological modality - the epigenome - to an existing DNA model by simply changing the training data. That's a practical result for the foundation model field: no new architecture needed, just a different pretraining diet.

Did you know? Curve also reported results from a 1,482-patient clinical study of its blood-based liver cirrhosis monitoring test. The company says its AI showed strong performance at identifying disease progression in a blinded validation of 597 patients, though specific numbers have not been disclosed. A peer-reviewed manuscript is in preparation.

Keytruda generated $31.7 billion in 2025 alone and is approved for 20 cancer types - the best-selling cancer drug in history. Its specificity comes from six flexible loops at the tip of the antibody. Change even one amino acid in those loops and you can destroy the drug's ability to bind its target.

Anindyadeep Sinha, a 24-year-old ML engineer from West Bengal, has built LiteFold - a web platform that packages advanced protein structure prediction into a plain-English interface. He gave its AI assistant Rosalind a challenge: redesign Keytruda's binding loops while keeping the rest of the antibody intact. Starting from Keytruda's crystal structure, Rosalind generated 100 redesigned variants in minutes. The top candidate showed significantly higher structural confidence than the original antibody run through the same pipeline, and held its shape across simulated physical conditions.

Why redesign a drug that already works? As patents on blockbuster biologics expire, there is intense commercial interest in “bio-betters” - molecules that hit the same target with different sequences, potentially with improved stability or easier manufacturing. When the drug you're trying to improve does $30+ billion a year, even a marginal improvement has enormous value.

LiteFold is transparent about the many, many limits. These are computational confidence scores, not experimental binding data. There is no paper. “Whether the AI-generated loops actually engage the target as predicted is a question for the wet lab,” the company writes.

Why it matters: What the case study demonstrates is that a small team can go from a crystal structure to scored, redesigned antibody variants in minutes - using a web interface and a natural language prompt. The tools that make this possible used to require serious compute infrastructure. LiteFold puts them in a browser. The result is a proof of concept. But the fact that a handful of people with modest funding can computationally redesign a $31.7 billion molecule might say something about where the barrier to entry is heading.

Did you know? LiteFold's Rosalind is unrelated to OpenAI's GPT-Rosalind, which we covered in Issue 18.

AlphaFold2 is the most important system in AI x biology - but if you're a machine learning engineer who wants to understand how it works, the code is either built for Google's infrastructure or wrapped in production complexity that's hard to learn from. Chris Hayduk, an engineer on OpenAI's life sciences team, argues this creates a bottleneck: skilled engineers stay in language models and image generation where the code is legible, and protein AI moves slower than it should. His fix is minAlphaFold2, a rewrite of AlphaFold2 designed to be read, modified, and run on a laptop. Every file maps to a section of the original paper's supplement. Clone it, get it running in under a minute, and start building.

☑️ AI scientists produce results without reasoning scientifically.

☑️ The double-edged sword of AI models and cancer treatment.

☑️ Boehringer Ingelheim opens new center for AI and machine learning in London.

☑️ Biomni Lab x GeneCards: Grounded biomedical agentic queries.

☑️ ConforNets: Latents-based conformational control in OpenFold3.

☑️ Why some brains with Alzheimer’s stay sharp.