I don’t know if Lila Sciences deserves an $8.5 billion valuation. But I do think some people in the field are still underrating the exponential. If not Lila, then someone else. “Scientific superintelligence” will make some readers lean forward and others roll their eyes, but the project underneath it is real enough: automate more of science itself.

That is the optimistic side of this issue. The darker side is biosecurity. DNA synthesis screening - or rather, the lack of screening - has been a huge problem for years, and it is even harder to defend now that frontier AI models are getting better at biology. If digital tools make dangerous designs easier to produce, the handoff from digital instructions to physical genetic material cannot keep relying on voluntary norms. This week, Sam Altman, Dario Amodei, Demis Hassabis and others signed an open letter asking Congress to require safeguards for synthetic DNA and RNA orders.

Today we also have Pfizer adding Chai to its AI drug discovery stack, Nature taking stock of virtual cells, Claude trying chemistry’s detective work, and APB-Display, a new method for turning protein libraries into better training data.

Let’s dive in.

Pfizer has licensed Chai Discovery's AI platform, giving its scientists early access to Chai-3, as well as a custom model built around Pfizer's own data and workflows.

Chai is a San Francisco startup founded in March 2024. Within a few months, Chai had launched Chai-1, its first protein-folding model.

The company is pushing against one of AI drug discovery’s default assumptions. Most AI biotechs eventually want their own drug pipeline, because the upside of owning a blockbuster is too big to ignore. Chai instead only sells access to its models. Jack Dent, cofounder and president, tells Forbes that building assets was the “dogma we had to challenge.” Mikael Dolsten, Pfizer's former head of R&D and now a Chai board member, puts the tradeoff more bluntly: if you want pharma to trust you as a partner, you probably should not also look like a future competitor.

Chai already signed Lilly in January. Forbes reports it is now talking to more than 15 additional pharma companies and discussing a $400 million raise at a $3.4 billion valuation.

Chai says new model Chai-3 is what convinced Pfizer to sign. It supposedly doubles the antibody-design success rate of Chai-2, produces antibodies that bind 100 times more tightly to their targets, and in roughly half of cases generates molecules that bind as tightly as approved drugs. Those are company claims by the way, not independent validation.

Forbes cites PitchBook data showing $11.4 billion invested in AI drug discovery companies globally in 2025, up from $5.6 billion the year before, with $5.5 billion already invested this year.

We’ve been tracking that shift through Lilly's Insilico deal, Isomorphic's $2.1 billion raise, and Genesis Therapeutics' expanded Incyte collaboration.

Why it matters: Big pharma is no longer waiting for a settled playbook. As investor Sylvain Gariel noted on X, Pfizer + Chai lands five months after Pfizer + Boltz, and “both PRs make it read as a similar scope: a custom model fine-tuned on Pfizer's own data.” His question: is Pfizer running these models in parallel, competitively, complementarily, or assigning each a lane? Nobody outside Pfizer knows. What we do know is that the industry has moved from asking whether AI drug discovery will be adopted to figuring out, in public and in real time, how to wire it into pharma R&D.

Did you know? Chai-1 is open source on GitHub. Chai Discovery is hiring.

Bloomberg reports that Lila Sciences is in talks to raise about $2 billion at an expected $8.5 billion pre-money valuation. The round is not closed. Reported anchor investors: CalPERS and Nvidia's NVentures.

The number might look bizarre if you treat Lila as a “normal” AI biotech. It makes more sense if you treat it as a bet on a new R&D operating system.

Lila was founded inside Flagship Pioneering (the venture-creation firm behind Moderna) in 2023 and emerged from stealth last year with $200 million in seed funding. Its pitch is not one model, one drug, or one lab. It is what the company calls AI Science Factories: AI systems connected to automated labs that generate hypotheses, design experiments, run them, learn from the results, and repeat across medicine, chemistry, materials, energy, and defense.

This might sound familiar. At BAIO, we have covered OpenAI running 36,000 experiments through Ginkgo, Ginkgo opening Cloud Lab, LIBRIS making 1,000 lipid nanoparticle formulations per hour, Adaptyv's wet-lab API, and Medra's autonomous lab. Different pieces of the same story: AI is making proposals faster than science can test them.

Lila, however, is claiming the full stack. Its own language is grand: an “operating system for science” and “scientific superintelligence.” Peter Diamandis, who discloses he is an investor, calls it a “Move 37 moment for science”. Maybe. Or maybe the vocabulary is racing ahead of the proof.

Why it matters: The market is now pricing closed-loop science itself - model, robot lab, experiment, data, repeat - as a platform category. Lila may be early to a real industrial layer, or early to a valuation reality check.

Did you know? Lila is hiring.

Nature has published a Technology Feature on virtual cells. The piece covers a lot of ground BAIO has already been tracking closely, as you’ll see below, but it also points to three areas we have touched on less.

1 Mechanistic models are not yet obsolete.

Bo Wang, the University of Toronto AI researcher who leads biomedical AI at Xaira, reminds Nature that “virtual cell 1.0” began more than 20 years ago with differential equations - equations that describe how biological systems change over time. Those older mechanistic models try to encode processes such as metabolism, signaling, movement, and cell-cell interaction directly into the model.

PhysiCell is an example of this older mechanistic tradition still doing useful work. It is a long-running open-source modeling framework from Paul Macklin’s group at Indiana University that simulates how cells and tissues respond to environmental signals. Nature notes that it has been used in cancer modeling, including tumor progression and response to immunotherapy. It does not replace AI foundation models, but it shows why equation-based models remain part of the virtual-cell toolbox.

2 Public embarrassment may be part of progress.

Eric Xing, chief scientist at GenBio - the Palo Alto startup whose VCHarness we covered in Issue 17 - is building AIDO, an AI-driven digital organism. The first prototype will take prompts such as gene edits or small-molecule interventions and predict outputs such as cell shape and protein localization.

Xing says GenBio plans to release it publicly, even if the results are bad. Virtual cells will not become trustworthy through polished demos. They need people outside the lab to find where they break.

3 The field is moving from cells to biological context.

A virtual cell would be a milestone, but it is a strange milestone. Even before the field gets there, it already has to work on the next layers. Cells talk to neighboring cells. Immune cells enter tissues. Tumors reshape their microenvironments. Signals travel within and across organ systems in ways a single-cell model will never see. Biology is nonlinear, dynamic, and full of feedback. You do not get to solve it neatly one layer at a time.

Zooming out, Nature’s feature also shows how much of this field BAIO has already been tracking. In past issues we’ve told you about Xaira’s X-Cell, PerturbGen, and Tripso. We have also covered the recurring problem that cell foundation models often struggle to beat simpler statistical baselines. And in Issue 25, Martin Borch Jensen made a related point: AlphaFold predicts protein structures, but current virtual cell models do not build up from those molecular predictions. They still depend on cell-level measurements such as RNA sequencing and imaging.

Anthropic wants Claude to help with a very ordinary chemistry problem: after you make a molecule, you still have to check whether you made the right one.

Chemists commonly check this with NMR spectroscopy: an instrument readout that gives clues about the atoms inside a molecule. The output is a pattern of peaks. A chemist then has to match those peaks to the molecule’s atoms. It is slow, expert work.

Anthropic tested Claude against two tools chemists already use: ChemDraw, standard software for drawing molecules and predicting spectra, and MestReNova, specialist software for analyzing NMR data.

First, Anthropic gave the models 20 novel compounds from recent ChemRxiv preprints and asked them to predict the NMR readout those molecules should produce. In practice, that means predicting where signals from hydrogen and carbon atoms should appear. Opus 4.7 had the lowest average error on hydrogen signals and was effectively tied with MestReNova on carbon.

Anthropic also tested the harder reverse task. Instead of starting with a known molecule and predicting its spectrum, Claude was given a formula and NMR data and asked to infer the molecule’s structure. In plain English: here is what the instrument saw - what molecule probably produced it? On eight simpler cases, Opus 4.7 got the right structure every time. For seven denser structures, Anthropic added a hint: the starting material used in the reaction. With that extra clue, Opus 4.7 returned the correct structure on most runs.

Anthropic is careful about the limits. The test used 20 molecules for the first task and 15 for the reverse task, so this is early evidence, not a general chemistry benchmark. The examples also leave out harder parts of real structure work: 2D NMR, which gives extra clues about how atoms connect; stereochemistry, which asks whether the same atoms are arranged differently in 3D; and broader solvent coverage, because spectra can shift depending on what the molecule is dissolved in.

That still leaves a clear signal. Claude is being tested on a real chemist’s translation problem: taking a molecule, predicting the instrument readout it should produce, and going the other way from readout plus formula to a proposed structure.

Why it matters: Labs lose time checking whether the molecule on paper matches the molecule in the tube. A model that can move reliably between structure drawings, formulas, and instrument readouts would make that confirmation work faster - especially when the result is ambiguous and a chemist needs a second pass.

Did you know? Anthropic is expanding its AI for Science program to support more chemistry research.

Sam Altman, Dario Amodei, Demis Hassabis, Mustafa Suleyman, Alexandr Wang and others have signed a letter urging Congress to require safeguards for synthetic DNA and RNA orders.

Researchers can order genetic material from synthesis companies. That is essential for vaccines, diagnostics, drug discovery, and basic biology. But the same supply chain could also be misused to make dangerous pathogens or toxins.

The letter calls for mandatory nucleic acid synthesis screening - checking DNA and RNA orders for dangerous sequences - plus recordkeeping and customer checks. Some synthesis providers already do this voluntarily. Many do not. The signers want it required across the industry.

This connects directly to two recent BAIO stories. In Issue 28, RefusalBench showed the model-behavior problem: some frontier models refuse ordinary biology too often, while others still help on prompts they should reject. In Issue 29, OpenAI’s Rosalind Biodefense showed the access problem: powerful biology models may need trusted-user programs for pandemic preparedness and biological threat response.

This letter points to the physical side. Even if AI makes dangerous biology easier to design, somebody still has to get the genetic material made. That handoff is still too loosely guarded.

Why it matters: Screening DNA and RNA orders has been a known biosecurity gap for years. Voluntary screening was already a weak defense when the main concern was trained insiders. It looks much more reckless now that AI models can help non-experts understand dangerous biology and potentially evade weak safeguards. The letter is a reminder that the risk does not end at the chatbot. At some point, digital instructions have to become biological material - and that handoff is still too loosely guarded.

Did you know? IBBIS’s DNAmap tracks 1,023 DNA synthesis companies across 81 countries. Only 69 are listed as screening orders, while 619 have unknown screening status.

APB-Display is a new way to make protein testing catch up with protein design. AI models can propose enormous numbers of proteins, but labs still have to measure which ones bind their target, and how strongly. Measuring one protein carefully is easy enough. Measuring tens of thousands carefully is a bit more difficult. APB-Display puts protein variants on tiny beads, keeps each protein linked to its DNA barcode, then uses sorting and sequencing to measure thousands at once. In the preprint, related workflows report detailed binding measurements for more than 18,000 variants in three days, and rougher measurements for more than 88,000. In other words: faster experimental data for training the next protein models.

☑️ On training data for bio AI models.

☑️ Self-revising AI scientists move from search to principled discovery.

☑️ Deep learning for H&E-based meningioma molecular classification and outcome prediction: a retrospective cohort study.