Quick note before we start: I’m heading to Vitalist Bay in Berkeley on Wednesday. BAIO will still publish twice a week, but send times may wobble a little while I’m in the US.

This issue accidentally revolves around Demis Hassabis’ corner of the universe. Bloomberg says Isomorphic Labs, the DeepMind spinout, is nearing a $2B-plus raise. Endpoints reports that young AI researchers are largely uninterested in pharma jobs - with Isomorphic as one of the rare drug developers they actually mention as attractive. And Google DeepMind’s AlphaEvolve is now improving the code underneath genomics and molecular simulation.

The one story not orbiting DeepMind/Isomorphic? Kanvas, which raised $48M to shine a light on the microbiome.

Let’s dive in.

Isomorphic Labs is in advanced talks to raise more than $2 billion, Bloomberg reports, with Thrive Capital leading and Alphabet participating. The round is not closed. But if confirmed, it would be one of the clearest signals yet that AI × bio is not fitting neatly into either “drug discovery startup” or “biotech company.”

The discussion on X has been telling. Biotech people have spent the weekend trying to make sense of the round through normal drug-development math: clinical programs are expensive, late-stage trials cost hundreds of millions, and a huge balance sheet lets a company take more shots before returning to market.

All true. Also incomplete.

These are not normal times. AI is not just another technological shift. The difference is not that Isomorphic uses AI in drug discovery. Pharma does that too. The difference is the starting point. In my interview with Max Jaderberg, Isomorphic’s president, earlier this year, he described the company almost in reverse order from pharma: first came the AI methods, compute and people; then AlphaFold as proof that deep learning could generalize across biology; then drug discovery as the application. The company name is in fact the thesis: a possible mapping between “the world of biology, cells” and “the world of AI and information science”: a sort of isomorphism.

That is why the biotech math feels incomplete. Jaderberg’s analogy, he told me, was not another biotech platform. It was chip design. Drug discovery, he argued, is still too artisanal: make or buy vast numbers of molecules, test what sticks, repeat. Isomorphic’s wager is that at least the preclinical design phase can move closer to engineering - more of the work done in silico, fewer blind shots in the lab, harder targets opened up by models that understand structure, binding and molecular behavior together.

Clinical trials remain clinical trials (or not). The FDA will not approve a molecule because AlphaFold’s cousins think it looks good. But this is more than Isomorphic buying extra shots on goal. It is trying to change how those shots are generated in the first place.

Why it matters: This is frontier AI economics entering biology without asking permission from the old guard.

Did you know? Isomorphic raised its first external round only last year: $600 million in March 2025, also led by Thrive Capital with GV and Alphabet participating. The company already has partnerships with Lilly, Novartis and Johnson & Johnson. Reuters reported in January that the company now expects its first clinical trials by the end of 2026. Isomorphic is hiring.

There’s another reason Isomorphic stands out from traditional pharma: young AI talent actually wants to work there.

Endpoints News reported from ICLR (International Conference on Learning Representations), one of the world’s top AI research conferences, and found a clear pattern among PhDs and postdocs working on AI for health and science: frontier AI labs and startups are attractive. Big pharma is not.

First, there is the salary chasm. Endpoints compared AI/ML life-science job postings and found many pharma roles sitting around $100K-$280K in base salary. Anthropic’s life-science AI role started at $280K. Its broader frontier AI roles can reach far higher.

But money is only part of it. Pharma barely showed up on the ICLR expo floor. Among about 60 organizations with booths, the biopharma presence was basically Isomorphic, ByteDance’s drug discovery unit, and Lila Sciences.

Why it matters: Pharma wants AI to reinvent drug discovery, but the talent market is voting with its feet. Sure, drug developers can always buy AI tools. But if the best AI scientists see pharma as slower, less ambitious, and less serious about research than frontier labs, it may struggle to buy the culture that makes top AI people want to build there.

Did you know? Alongside NeurIPS and ICML, ICLR is one of the three major machine-learning conferences - and a key recruiting ground for frontier AI labs.

The Gemini-powered coding agent, AlphaEvolve, writes code, tests it against automated evaluators, keeps what works, mutates it, and tries again. Google DeepMind introduced AlphaEvolve last year as an algorithm-discovery system, not a life-science tool. It had already found ways to save Google compute, speed up parts of Gemini training, and improve a 50-year-old math problem. Now the same system is showing up in genomics and molecular simulation.

The cleanest biology example comes from DNA sequencing. PacBio, the California company best known for long-read sequencing, uses DeepConsensus - a transformer model developed with Google AI Genomics - that corrects sequencing errors so researchers can more accurately identify genetic variants.

AlphaEvolve found an improvement to how DeepConsensus learns from sequence alignments. Google DeepMind says the change reduced variant-detection errors by 30%.

Schrödinger saw the same pattern from another angle. The New York company, whose software is used in drug discovery and molecular modeling, used AlphaEvolve to speed up machine-learned force fields (MLFF) - models that estimate the forces between atoms so researchers can simulate how molecules move and interact - by roughly 4x in training and inference.

“AlphaEvolve allows us to explore larger chemical spaces faster and more efficiently than ever before. Faster MLFF inference carries real business impact, shortening R&D cycles in drug discovery, catalyst design, and materials development, and enabling companies to screen molecular candidates in days rather than months,” says Gabriel Marques, Technical Lead of Machine Learning at Schrödinger, in the Google DeepMind article.

Why it matters: AlphaEvolve does not need to understand biology in the grand sense to be useful. It needs a problem with a measurable score: fewer sequencing errors, faster molecular simulation, better training code. Computational biology is full of those bottlenecks. If AlphaEvolve can keep finding small but real improvements in the algorithms underneath the field, the gains compound quietly across everything built on top.

Did you know? The refined DeepConsensus training approach is open source on GitHub, and Google DeepMind says AlphaEvolve is now being brought to enterprises through Google Cloud.

Kanvas Biosciences raised $48 million to turn the microbiome into something AI can model - not just sequence.

The Princeton, New Jersey company works on live microbiome drugs: medicines made from living bacteria. Its lead program, KAN-001, is aimed at cancer patients treated with immune checkpoint inhibitors, the drugs that release the brakes on immune cells so they can attack tumors.

Some patients respond better to checkpoint therapy when their microbiome has the right mix of bacteria. Fecal transplants from responders have shown promise, but they are messy: donor material varies, manufacturing is hard, and you do not always know which bacteria actually took hold.

Kanvas wants to replace that guesswork with maps. Its platform images bacteria and host cells in tissue, then uses the resulting spatial data to train AI models and design defined bacterial mixtures. The company has also built a new microscope to generate more of that data for what it calls a microbiome atlas.

Why it matters: The microbiome has produced plenty of correlations. Kanvas is trying to turn it into a controllable design problem: map where the bacteria are, learn which patterns matter, build the mixture, then test whether it works.

Did you know? Kanvas is hiring.

BenchOS is a new AI workspace for molecular and structural biology labs from Harvard’s Farnung Lab. Instead of keeping protocols, sample records, equipment bookings, plasmid maps and cryo-EM files in separate places, BenchOS pulls them into one shared system. Its chat agent, Benchmate, can use that lab context to design primers, plan PCR reactions, check cloning results and update records.

☑️ Phylo Bio turns UK Biobank rare-variant statistics into target hypotheses.

☑️ Open Source for Science fund launches to power AI-driven discovery.

☑️ American factories lag in adopting AI. This drugmaker is an exception.

☑️ Stanford Daily profiles James Zou’s AI scientist lab.