Anthropic's BioMysteryBench sets Claude 99 bioinformatics questions on real data with checkable answers; the best model solved 30% of those no expert on the panel could
What happened: Experts wrote each question from a real dataset, with an answer fixed by the data or a validated record (which organ a cell sample came from, which gene was knocked out) and a notebook showing the signal is there. Claude worked in a container with standard tools and public databases, graded on its answer alone. Up to five experts tried each question and 76 were solved by at least one. On the 23 none solved, Claude Mythos Preview's solve rate, averaged over five attempts, was 30%.
Why it matters: Grading the answer rather than the method lets a model take any route, and fixing the answer in the data rather than in a scientist's conclusion lets the benchmark hold questions people cannot answer. Five attempts per question also show how firm a win is: Opus 4.6 solved 86% of the solvable questions it got right at least four times in five, but only 44% of its wins on the hard set.
Evidence and limits: Anthropic built the benchmark with its own experts and scored only its own models, and most per-model scores are shown only in charts. A question nobody solved may have no answer in the data, as the post says. The full set, now 90 questions, is gated for evaluation only; five are public.
Lilly will license recombinases Profluent is to design with AI, enzymes for writing long DNA into the genome: upfront undisclosed, up to $2.25B in milestones
What happened: Profluent will use its protein models to design site-specific recombinases, enzymes that insert long stretches of DNA at a chosen place in the genome, for several targets. Lilly takes an exclusive licence to carry the ones it selects through animal studies, clinical trials and sale. Profluent gets an upfront payment and committed research funding, neither sized, and is eligible for up to $2.25B in development and commercial milestones plus tiered royalties. Ali Madani, Profluent's co-founder and chief executive, speaks for the company.
Why it matters: Recombinases can write a whole gene into DNA, which today's CRISPR editors do poorly, and the usual way to get one is to search nature for an enzyme that happens to fit the site. Here a large drugmaker pays a model company to design them for chosen targets. Its other announced deals, with Revvity, Corteva and, five months later, BioMarin, give no value.
Evidence and limits: The targets, the number of programmes and the upfront are not disclosed, so the $2.25B is milestone money that depends on results nobody has shown yet. No data on a designed recombinase came with the deal.
Fathom Therapeutics, formerly Atommap, raises a $47M Series A led by Sutter Hill Ventures for physics-based simulation of protein motion to design small molecules
What happened: The New York company's engine, Microcosmos, simulates how proteins move at atomic resolution and uses those simulations to design small molecules; the company says it speeds up such modelling 10,000-fold without losing accuracy. Co-founder and CEO Huafeng Xu leads it with co-founders Jesús Izaguirre, from Roivant Discovery and Silicon Therapeutics, and Yujie Wu, from Roivant Discovery and Schrödinger. Sutter Hill Ventures, through managing director Keith Loebner, led the oversubscribed round, with Ethan Kurzweil's Chemistry, Alexandria and NY Ventures. The money scales its lab-in-the-loop work and internal programmes.
Why it matters: Fathom's CEO says drug discovery is held back by designing against static structures of isolated proteins. Its pitch is to design against proteins as they move, computed from physics rather than learned from structures alone, and a $47M Series A shows investors paying for that route.
Evidence and limits: The 10,000-fold speed-up and the degraders found in six weeks against an "undruggable" target are the company's claims, and the release cites no paper or data for them. Its lead programme has reached animal efficacy studies, not people.
David Silver's Ineffable Intelligence comes out of stealth with a $1.1B seed round co-led by Sequoia and Lightspeed, at a $5.1B valuation
What happened: The London lab, founded by David Silver after he led reinforcement learning at Google DeepMind, raised $1.1B at a $5.1B post-money valuation, per its counsel Cooley. Alfred Lin and Sonya Huang of Sequoia and Ravi Mhatre and Raviraj Jain of Lightspeed lead; NVIDIA, DST Global, Index Ventures, Google, EQT Ventures, Wellcome Trust, the British Business Bank and the UK's Sovereign AI Fund take part. Silver says he will give away all his gains from his Ineffable equity.
Why it matters: It is the largest seed round in Europe, put behind a plan to build an agent that learns from its own experience instead of from pre-training on human data. It shows capital will fund that bet on reinforcement learning, and that a senior DeepMind researcher can raise a frontier-lab round before showing a result. Six more co-founders followed in September, four of them from DeepMind.
Evidence and limits: A financing says where money is going, not that the method works: the company has shown no result, benchmark or paper. The valuation is the company's, stated through its counsel. The participant list and the pledge come from Tech.eu's report of the company's statement, and the 'strategic angels' are not named.
CuspAI open-sources kUPS, a GPU molecular simulation engine built so machine-learned force fields plug straight in
What happened: Nicholas Gao, Jonas Köhler, Felix Hanke and Anita Ramanan describe a JAX toolkit that puts molecular dynamics, Monte Carlo and geometry optimisation behind one interface and runs many systems at once on GPUs. Tuned for GPUs with NVIDIA, it ships with Tojax, which turns PyTorch functions into JAX. On one NVIDIA L4 GPU, with a metal-organic framework as the test case, it ran MACE molecular dynamics at 1.8 times the throughput of ASE and CO2 adsorption Monte Carlo at 49 times that of RASPA on one CPU core.
Why it matters: Screening materials such as carbon-capture frameworks means chaining simulation methods that live in separate CPU codes, and machine-learned force fields fit them badly. A batched GPU engine is what a company betting on AI-scale screening needs, and releasing it under Apache 2.0 lets others use it. The speed-ups are the company's own: two workloads on one material, against tools running on a CPU core or one system at a time, with no outside benchmark yet.
AuthorsNicholas Gao, Jonas Köhler, Felix Hanke, Anita Ramanan
Kevin Weil leaves OpenAI as OpenAI for Science is split into other research teams
What happened: Kevin Weil, who started OpenAI for Science after serving as chief product officer, said on X that 17 April was his last day, "as OpenAI for Science is being decentralized into other research teams". A company spokesperson told reporters that Prism, its writing workspace for scientists, moves to the Codex team.
Why it matters: OpenAI no longer runs science as a unit with its own head: the work now sits in the teams building its models and Codex. Its next science tools will come from there, not under the OpenAI for Science name.
Bioptimus reports its pathology model H-optimus-1 ranks first on two outside benchmarks, but 14th of 23 on tissue segmentation
What happened: In a late-breaking abstract at the AACR 2026 meeting, Marin Scalbert, Rodolphe Jenatton and colleagues describe H-optimus-1, a 1.1-billion-parameter vision model trained without labels on over 1 million slides from more than 800,000 patients at over 4,000 clinical centres. Across 13 tasks on 15 datasets, such as predicting mutations and gene activity from stained tissue, they report the best average score. Their poster adds outside rankings: first on PathBench and HEST, 3rd of 23 on THUNDER's few-shot tasks.
Why it matters: A stained tissue slide is the most common record of a cancer. A model that turns any slide into general-purpose features lets a hospital or drugmaker build a biomarker test from a smaller labelled set. H-optimus-1 is Bioptimus's main model, free for academic use since February 2025, and the poster's case is that benchmarks run by other groups, not only its own tests, put it at or near the top.
Evidence and limits: Company-reported: the abstract gives no per-task numbers, and the outside rankings are the company's reading of other groups' leaderboards. THUNDER ranks it 14th of 23 on segmentation. Nothing here tests the features in a clinical decision.
Anew Labs presents AI-designed oral small molecules against all three IL-17 dimers
What happened: At AAI in Boston the team reported orally bioavailable pan-IL-17 inhibitors from its generative platform, the first molecules it says engage IL-17F/F strongly. FOCIS in June added in vivo data.
Why it matters: IL-17 blockade is an injectable-antibody market; what is new is a small molecule reaching the F/F dimer. This is conference data: no paper, no structure, no dose.
An amateur posts a proof of an Erdős conjecture on primitive sets that GPT-5.4 Pro gave him from one prompt, which Terence Tao says takes a new route
What happened: Liam Price, 23 and without advanced maths training, posted to erdosproblems.com a proof GPT-5.4 Pro gave him: that for primitive sets, where no member divides another, the largest possible Erdős sum falls to one as the members grow. Jared Duker Lichtman, who proved a related Erdős conjecture in 2022, says the raw output was poor and needed an expert to read; he and Terence Tao have shortened it.
Why it matters: Many earlier AI solutions to Erdős problems were less original than they looked. Here, Tao says, people had worked on the problem and all began with the same standard step, while the model, Scientific American reports, used a formula well known in nearby areas that nobody had applied to it. Tao and Lichtman see other uses for the method.
Evidence and limits: The report names no Lean proof and no paper; the check so far is Tao's and Lichtman's reading. Tao says the jury is still out on its long-term significance.
Five models that predict how fast an enzyme works, Zelixir's own CataPro among them, picked xylanase mutants for the lab; under one in ten of each model's picks was 20% more active
What happened: Zhongyang Zhang, Weifeng Li, Liangzhen Zheng, Nanyu Han and colleagues at Yunnan Normal University, Shandong University and Zelixir tested DLKcat, UniKP, CataPro, EITLEM-Kinetics and TurNuP on newer data, unfamiliar enzymes and mutation scans. None predicted turnover rate (kcat) accurately, and their ranking fell apart on enzymes unlike their training data. Each model then picked its top 20 mutants of the xylanase PjxA; 98 were made and measured. Agreement with measured activity was weak: UniKP did best, just above 0.2 (Pearson) and 0.3 (Spearman); CataPro and DLKcat near zero, two negative.
Why it matters: These models are offered as a way to shortlist enzyme mutants and to fill gaps in metabolic models. Run partly by CataPro's own authors, this test says they cannot yet do the first: they react when the active site is broken but ignore the molecule being converted, unable to tell a substrate from its product or an inhibitor. The authors conclude they work at best as a first filter; CataPro's picks were among the better three, though few beat the starting enzyme.
Evidence and limits: One enzyme, one round of 20 picks per model, crude-enzyme activity measured under one set of conditions, and a preprint not yet peer-reviewed. Liangzhen Zheng of Zelixir is a corresponding author of both this test and CataPro, so the group graded its own model alongside four others.
AuthorsZhongyang Zhang, Weifeng Li, Liangzhen Zheng, Nanyu Han
Lee Cronin, Chemify's founder, argues in PNAS that a Chemputer with step-by-step error correction could in principle make any stable molecule, and re-reads 117 robot runs to back it
What happened: The paper defines a Chemical Synthesis Turing Machine: four primitive operations (add or remove matter, add or remove energy) as the instruction set, a compiler from reaction graph to hardware graph, and error correction at every step. Given enough fidelity per step, it argues, one reconfigurable machine can make any stable, isolable molecule. Its test is 117 routes the group had already run on its robots as χDL programs: unit operations grow roughly in line with reaction steps.
Why it matters: If synthesis is code, a route can be shared, compiled for different machines and checked automatically, the way software is. This is Cronin's formal case for programmable synthesis, and it names the obstacle: errors pile up over many steps, so the claim rests on catching and fixing them while each step runs.
Evidence and limits: Universality is argued, not shown: no new molecule was made for the paper, and the 117 routes are earlier one- to three-step runs re-counted, chosen from common medicinal, process and total-synthesis reactions. The error-correction scheme is described, not tested here, and the paper calls real-time adaptation to new chemistry a significant hurdle. Cronin founded and holds shares in Chemify.
AuthorsLeroy Cronin, Sebastian Pagel, Abhishek Sharma
Polyphron and Cellino grow aligned brain tissue from four donors' stem cells, predicting the conditions for two thinly tested donors from the other two
What happened: Cellino made four stem-cell lines from four donors' blood (both sexes, ages 24 to 54) on its automated, image-guided platform. Polyphron grew cortical neurons from them in 3D chips; one shared set of growth signals gave up to a 9-fold spread in how well the nerve fibres lined up. With two donors mapped fully and two thinly, its model picked five conditions per donor. Tested, 19 of the 20 best wells matched an optimised reference line, with 12 times fewer experiments than that line had needed.
Why it matters: Stem cells from different people respond differently to the same recipe, which is a main reason grown human tissue does not scale. Here the foundry carried what it learned on two donors over to two others with little data, the step a bank of patient-matched tissue would need.
Evidence and limits: A company white paper, not peer reviewed, with both firms reporting their own work. It targets one structural feature, fibre alignment, in four donors, and does not show the tissue predicts how drugs act in people.
An LLM agent from Cronin's group, whose founder leads Chemify, turns published procedures into robot code and runs six of them
What happened: ACRA extracts synthesis procedures from documents, writes them as XDL, the Chemputer's Chemical Descriptor Language, checks them against the text, simulates the run and then executes it. It pulled 717 procedures out of publications, 57 from a German practical script and 117 from a PhD thesis. On 150 ChemRnD procedures, 99.33% became valid XDL and 94.67% passed the discrepancy check and simulation too; on a separate 75, its data and memory stores raised success from 85.3% to 93.3%. Six syntheses ran on a Chemputer and an Opentrons OT-2.
Why it matters: The backlog the paper is aimed at is the published literature itself: procedures nobody has repeated, written ambiguously. If an agent can turn a paper's method into a script a robot executes, repeating a synthesis costs machine time rather than a chemist's week, and the one procedure here that would not work was found by running it.
Evidence and limits: Six procedures were run, by the authors, and one is the paper's own negative result: a published sugar synthesis gave no conversion, a chemist re-ran it by hand, and the paper classes that procedure not reproducible. The high percentages are software checks, not products in a flask, and sorting a procedure into executable, blueprint or incomplete was right 67.5% of the time on 39 cases. Chemify is not on the paper, whose authors are all at Glasgow.
AuthorsSebastian Pagel, Michael Jirasek, Leroy Cronin
LabGenius reports its machine-designed Nectin-4 x CD3 engager kills tumour cells at 0.7 pM and cuts tumour growth by over 90% in mice
What happened: LGTX-101, a trivalent VHH T-cell engager generated with LabGenius's discovery engine EVA, carries three Nectin-4 arms to one CD3 arm. On cancer lines whose Nectin-4 level is within twofold of a healthy-cell model, its mean EC50 was 0.712 plus or minus 0.132 pM, and the healthy model was not killed, a window the abstract puts above 1,000-fold. T cells fired below 5 pM in primary bladder cancer co-cultures, not with keratinocytes. In a humanised BT-474 xenograft, established tumours grew more than 90% less.
Why it matters: Nectin-4 is already a drug target, and what limits it is the healthy tissue that also carries it; LabGenius's claim is that selectivity can be engineered by how many arms have to bind at once (three to one here, so only dense target expression fires the molecule) rather than by finding a better epitope, and that its own discovery engine is what found the design. This is a conference abstract: mice and cell lines, the company's own assays, no paper, and the dosing interval it proposes rests on preliminary pharmacokinetics.
