AI drug discovery deals pile up, but the dossier's freshest evidence is a quantum code paper

The newest dated item is an arXiv preprint submitted on 30 October by Jae-Hyun Baek, Byung-Sun Won and Jon-Lark Kim. It is titled Researcher-in-the-Loop Evolutionary Discovery of Entanglement-Assisted Quantum Code Families. It has nothing to do with drug discovery.

That gap is the story. Readers arriving from a query about AI drug discovery partnerships will find plenty of headlines and very little verifiable text. Two sources on that topic, both from Endpoints News and both dated 30 October, are truncated at the paywall. One is an interview with Nathan Frey, Anthropic's drug R&D chief, conducted on stage at an Endpoints AI day. The visible text says only that Frey has stayed busy since Anthropic acquired his biotech startup Coefficient Bio this spring. It also says he was formerly a machine learning scientist at Genentech. The rest is behind a subscription prompt. The second, also 30 October, reports that Sanofi may seek higher European drug prices rather than withhold launches under a pricing deal associated with President Donald Trump. Its visible text stops before any figure appears.

What the paywalled pieces do and do not say

This matters for how the story gets written. A reporter cannot lift the substance of a subscription interview and present it as a finding.

What can be said is narrow and attributable. Endpoints News published a Q&A with Frey on 30 October, and the visible portion confirms the Coefficient Bio acquisition closed this spring. No deal value, no pipeline count and no partner names are visible. The same restraint applies to the Sanofi piece. It is a pricing story rather than a partnership story, and its content beyond the headline and standfirst is not available. So the usable evidence base for an analysis of AI drug discovery partnerships is thinner than the headline list suggests. The file contains 11 sources. Only one was published inside the last 72 hours, and it is the quantum codes paper. Of the rest, four are ScienceDaily releases dated 27 and 28 October. Their subjects range from oxygen stored in magma to fluorescence in fire salamanders. One is an NVIDIA blog post from 21 October about cloud partners in Europe. One is an NVIDIA blog post from 27 October about dementia research tooling. Two are the Endpoints items. None of them is a deal announcement with terms attached.

The visible text says only that Frey has stayed busy since Anthropic acquired his biotech startup Coefficient Bio this spring.

The nearest thing to a partnership document

The closest thing to primary material on AI and biomedical research collaboration is the NVIDIA blog post of 27 October on C-BRAIN, the Consortium for Biomedical Research and AI in Neurodegeneration. According to that post, C-BRAIN is led by Dr Randall Bateman of Washington University School of Medicine in St Louis. It comprises 18 member institutions across academic, philanthropic, pharma and research organisations. It launched in November 2025. The post states that C-BRAIN runs 64 NVIDIA data center GPUs on premises alongside cloud based nodes on Microsoft Azure, with Blackwell systems described as on the horizon.

The post also carries figures worth checking against primary sources before anyone repeats them as fact. It says 57 million people worldwide live with dementia. It says more than 10 million new dementia cases are diagnosed each year. It says National Institutes of Health funding rose sevenfold over 15 years to $3.9 billion annually. Those are the blog's numbers, not figures this piece has independently verified.

The post quotes Bateman describing the consortium's work as "AI tools built by scientists, for scientists". It quotes C-BRAIN chief technology officer Ade Ojewole on life sciences data that never reaches frontier models. Both quotes are attributed to the NVIDIA blog, which is a company publishing about a collaboration it is part of.

Why the partnership headlines are hard to report

A working reporter covering this beat would normally anchor a piece like this on filings, trial registries or company statements. The file offers none of those for any AI drug discovery deal. What it offers instead is a pattern of announcement driven coverage, with terms either undisclosed or locked behind subscriptions. That is not a reason to skip the story. It is a reason to say plainly what is known and what is not. It is also a reason to avoid treating a partnership press release as evidence that anything has been discovered.

The same caution applies to the 30 October Endpoints interview with Frey. The standfirst mentions catching Prof. Oak, the limits of Claude's taste, and cures versus QALYs. Those are the interviewer's framing, not statements of fact about drug programmes. Without the body text, none of it can be reported as a claim about Anthropic's pipeline. The acquisition of Coefficient Bio is the one durable fact in the visible portion. It is a corporate transaction, not a research result.

A different kind of AI discovery paper, with numbers attached

By contrast, the quantum codes preprint publishes its methods and its evidence in the open. The authors describe a researcher in the loop pipeline. It combines program evolution, exact evaluation, mathematical generalisation, certified nonexistence checks and formal proof checking, applied to entanglement assisted quantum error correcting codes. Evolved programs produced finite constructions that the researchers then generalised into a family written as [[n,1,n-1;n-q-1]]_q for prime powers q at or above 3 and n at or above 2q+1, and for q=2 at odd n at or above 5. The abstract states that Lean checks the algebraic certificate underlying the family. It states that the family closes 54 gaps in archived EAQECC tables.

The released evidence, per the abstract, includes 115 verified code records. It includes ten nonexistence results with checked DRAT proofs, family implementations and the preserved search programs for exact checks and seeded replay. In a controlled cohort of eight paired generation blocks, validation selected iterative programs reached success in 46.1% and 31.6% of held out executions at unseen lengths n=13 and n=15. Independently generated programs reached 19.5% and 3.1%. Those are the kind of figures a reader can audit. The paper is 25 pages with six figures, and the authors have posted code, data and Lean proofs alongside it.

What the rest of the file actually covers

The remaining ScienceDaily items are science releases, not industry news. One, dated 28 October and sourced to Science China Press, reports that magma deep inside Earth can store more oxygen than previously believed. It is based on high pressure experiments up to 7 gigapascals and temperatures of 1200 to 1800 degrees Celsius. The work was led by Yanhao Lin and Ho-kwang Mao of HPSTAR and published in National Science Review. Another, also 28 October, from the Max Planck Institute for Chemical Ecology, describes green and cyan fluorescence in fire salamanders. Their secretions stay fluorescent for more than 24 hours. It quotes first author Bernat Burriel and director Martin Kaltenpoth. A third, dated 28 October from Istanbul Technical University, describes a continental magmatic arc in northwestern Iran active roughly 101 to 97 million years ago. It was published in Geoscience Frontiers, with a quote from corresponding author Ali Mohammadi.

Two more carry medical relevance. A McGill University release dated 27 October reports in Nature Immunology that T cells can sense tissue stiffness. It says stiffer collagen gels push them toward a tissue resident memory profile, quoting lead author Judith Mandl. An NTNU release dated 27 October reports in NAR Molecular Medicine that JAK inhibitors including baricitinib can suppress antiviral gene activity. This could potentially ease spread of Rift Valley fever virus, influenza A, adenovirus and SARS-CoV-2. It quotes Denis Kainov and first author Erlend Ravlo. A Netherlands Institute for Neuroscience release dated 27 October describes immature neurons in brain tissue from the Netherlands Brain Bank. Senior author Evgenia Salta is quoted saying around 30 percent of older adults who develop Alzheimer's never experience its symptoms. None of these is a partnership.

How to read the beat

The honest conclusion is structural. AI drug discovery partnership coverage here is thin on primary documents and heavy on events. The one substantive consortium description comes from a vendor blog. The two trade pieces are paywalled. The most recent publication is a mathematics preprint. That does not mean the partnerships are not happening. It means the evidence cannot support claims about their terms, their pipelines or their results. Any analysis that pretended otherwise would be inventing detail.

For readers tracking the space, the practical advice is to wait for the documents. Deal values, milestone structures and target lists surface in filings and in the partners' own statements. Those are the sources worth building on. Until then, the most defensible thing a reporter can do is describe the material accurately, note where it stops, and resist a headline that promises more than the text delivers.