One of today’s On Our Radar links came from Luigi Sigillo, first author of the paper introducing EMBL’s AI Librarian. Luigi reached out to me about it on X. I took a look and decided BAIO readers should see it.

BAIO is read by scientists, founders, engineers, investors and people working across biotech and pharma. You see parts of this field that I do not.

And I am not only asking for papers, product launches or funding news. What do you struggle with that rarely gets mentioned? Which supposedly solved problem still wastes your time? What tool is missing? What is everyone in your corner of AI x biology talking about that the rest of the field has not noticed?

Reply to this email or reach out on X. Your message may very well shape what BAIO reports on next.

In today’s issue, we look at Vilya’s attempt to make macrocycles easier to design, Biomni’s claim that it compressed months of computational biology work into five days, and GSK’s plan to use new cellular data for virtual experiments.

We also have NSF’s remotely programmable biomanufacturing lab, the breakup of DeepMind’s AlphaFold team and Benchling’s newly public Model Hub.

Let’s dive in.

Seattle-based biotech Vilya has released an AI model for predicting how a promising but difficult-to-model class of drug will fit against its target.

Called Vilya-2, it is built for chemically modified peptides, including ring-shaped macrocycles that do not fit neatly into conventional small-molecule or protein modeling systems.

Macrocycles, by the way, are the main reason Vilya exists. The biotech launched in 2022 from research at David Baker’s Institute for Protein Design. By 2024, it had raised $71 million in Series A funding.

Small molecules can often be taken as pills and enter cells, but may struggle to bind the broad, relatively flat surfaces where proteins interact. Antibodies can bind those surfaces, but are generally injected and usually cannot enter cells.

Macrocyclic peptides occupy a middle ground: larger and more structurally varied than ordinary small molecules, yet far smaller than antibodies. Some can be designed to cross cell membranes or be absorbed when taken orally.

Nature has already supplied successful macrocycles such as cyclosporine. Designing new ones deliberately is much harder. A useful molecule must adopt the right shape, bind its target, remain stable and cross the necessary biological barriers - properties that can conflict with one another.

Vilya uses computation to design and screen new macrocycles, then makes and tests the most promising candidates in the laboratory. It now lists five internal programs: two in oncology, two in immunology and one for inflammatory bowel disease.

“Traditional drug discovery searches through what already exists. We design what should exist,” the company says on its website.

In January, Vilya introduced Vilya-1, which predicts the shapes macrocycles can adopt and properties that help determine whether they could work as drugs.

Vilya-2 predicts how one of those candidates will fit against a protein target.

BAIO recently covered PeptiVerse, which predicts peptide properties including binding affinity. Vilya-2 predicts the bound structure itself: where the peptide sits against its target and what shape it adopts there.

Macrocycles can contain both ordinary and synthetic amino acids, along with extra chemical pieces that join parts of the chain and hold it in a ring. That variety makes them difficult for models designed specifically for either proteins or small molecules.

Vilya-2 represents every molecule in the same basic terms - which atoms are present and how they are connected - then predicts where those atoms sit in three-dimensional space. The aim is to let patterns learned from proteins, peptides and small molecules carry over to unfamiliar macrocycles.

In a late-July technical report, Vilya tested the model on 88 protein-peptide complexes whose experimentally measured structures were published after Vilya-2’s training cutoff. With 100 attempts per complex, Vilya-2 reproduced the peptide’s bound structure to within two angstroms in 54.1% of cases, compared with 40.9% for Boltz-2, an open biomolecular structure model. At 1,000 attempts, its result rose to 59.1%.

It bears mentioning that these are retrospective results on a benchmark built by Vilya itself. The preprint does not show Vilya-2 designing or selecting a new molecule that was then prospectively made and tested, so independent evaluation and prospective laboratory results are still needed.

Successful macrocycles such as cyclosporine show what this class of molecule can do, but many were discovered rather than deliberately designed. Vilya is trying to turn that biological luck into an engineerable drug class. If models can reliably predict how unfamiliar macrocycles fold and bind, researchers could explore this middle ground systematically instead of waiting for nature to supply the right molecule.

A computational biologist can spend months finding a good way to turn biological data into predictions. Phylo, a young South San Francisco company, says its AI research system Biomni used an integrated tool called TusoAI to test about 500 possible approaches in five days.

Biomni is an AI workspace for biological research. It can search papers, write and run code, use scientific tools and coordinate longer research tasks. One of Phylo’s core ideas is to help scientists spend less time making a thicket of tools work together - and more time doing the science itself.

TusoAI takes on a more specific job: developing the computational method itself.

Computational biology uses computers to answer biological questions involving too much data or too many possibilities for people to handle manually. But the computer still needs a recipe: which data to use, how to represent it, which model should look for patterns and how success will be measured. Much of the work is unglamorous - reconciling datasets that label the same genes differently, correcting for the batch a sample was processed in, deciding which measurements are real signal and which are artefacts. Finding a recipe that works can mean reading papers, writing and debugging code, testing alternatives and revising what fails.

Biomni and TusoAI automate much of that work. The researcher describes the task, and Biomni sets up the data, code and way of scoring the result. TusoAI searches the literature, proposes an approach, implements it, measures how well it works and uses the result to decide what to try next.

To demonstrate the system, Phylo asked it to build a model that predicts how a cell will respond when a gene is switched off.

The best-performing recipe combined nine sources describing what genes do, which pathways they belong to and how they interact. It then used several relatively simple statistical models to predict the effects of gene knockouts that had not been tested.

Phylo calls the resulting prediction system TusoPerturb. The company says it outperformed leading methods - including much larger deep-learning systems - on almost every metric and dataset across three benchmarks.

“We believe this is where bio AI model research is headed. Instead of spending months building models by hand, researchers can describe a problem in natural language and let Biomni explore hundreds of ideas on their own data, while handling all the ML infra,” Phylo co-founder and CEO Kexin Huang writes on X.

The results currently appear only in Phylo’s launch material; the paper with the full methods and evaluation is still forthcoming. The public repository contains the model code and several biological reference datasets, but not everything needed to reproduce the reported scores. The system also improved its approach against a small validation set, making independent testing important: did it find a generally useful recipe, or one unusually well suited to these benchmarks?

BAIO has been following Biomni since its launch. In February, in Issue 1, we covered Phylo’s $13.5 million raise and its pitch for Biomni Lab as an “integrated biology environment” where scientists could orchestrate AI agents. Two months later, in April, Issue 18 returned to Biomni when it added GPU-as-a-tool, allowing researchers to create, fine-tune and train biology models on their own data through natural-language instructions.

Phylo is not emerging from nowhere. Its scientific co-founders include Stanford AI researcher Jure Leskovec and Le Cong, whose lab BAIO has followed through LabOS, LabClaw, MedOS and LabWorld - systems for guiding scientists through experiments, equipping research agents with tools and training them in simulated laboratories. Across those projects, the recurring instinct is to identify what slows scientific work down and build AI around it.

Phylo already has early adoption and investor-reported praise from scientists at major pharma companies. If Tuso’s results hold up independently, it would suggest the company is building momentum by solving the work around scientific discovery that powerful models alone do not remove.

British pharmaceutical company GSK has expanded its collaboration with London biotech Relation Therapeutics in a deal worth up to $110 million.

Relation will use automated laboratories to alter genes or add compounds in human cellular disease models, then measure the resulting changes across multiple molecular layers. The aim is to produce unusually large and consistent multi-omic datasets for training MORGAN, a new family of cellular foundation models unveiled alongside the deal.

Each MORGAN model will focus on a tissue or cell type and try to predict how its cells respond to genetic changes or drugs. Relation says this could help reveal disease mechanisms and identify possible drug targets.

The companies already have two 2024 deals in osteoarthritis and fibrotic diseases, backed by $45 million upfront. The new agreement combines an undisclosed upfront payment with success-based milestones, so the full $110 million is not guaranteed.

Public evidence for MORGAN is thin. Relation has released no technical paper, benchmarks or example of a target proposed by the model and then confirmed experimentally.

GSK is not merely applying AI to data it already has. Relation will create new multi-omic data by systematically perturbing human cells, use it to train models of how diseased cells behave, then run virtual experiments on those models. Those virtual experiments are meant to help GSK judge earlier which medicines are most likely to work, before development becomes far more expensive.

AI may have found which rectal cancer patients benefit from adding irinotecan - a chemotherapy drug that caused more serious side effects and failed to improve survival across a trial population as a whole.

The findings, published in eBioMedicine, come from a new analysis of ARISTOTLE, a UK phase III trial conducted between 2011 and 2018. The trial tested whether adding irinotecan to standard chemoradiotherapy before surgery helped 564 people with locally advanced rectal cancer. It did not overall.

Researchers have now used AI to reanalyse pretreatment biopsy slides from 414 participants. The system identified tumour and stroma regions, counted the cells within them, then divided the patients according to how densely their tumours were packed with cancer cells.

Among patients with cancer-cell-dense tumours, adding irinotecan was associated with a 43% lower risk of recurrence or death and a 50% lower risk of death over up to five years. Patients with proportionally more stroma showed no clear benefit.

Pathologists can estimate tumour-cell density manually, but doing it across hundreds of biopsies means selecting tissue regions and counting large numbers of cells slide by slide. That makes the measurement too slow and cumbersome for studies at this scale.

The authors say the AI processed each slide in about two minutes. A manual method cited in the paper takes around 20.

“This demonstrates how AI can reveal tumour biology that is difficult to measure consistently by conventional means, and hopefully can be used to reveal additional insights that could lead to future treatments,” lead author Dr Zhuoyan Shen said in a University College London article.

Note that this is a retrospective second look at the trial. The tumour-cell-density measurement now needs to show in an independent dataset that it can identify who benefits from irinotecan, and eventually guide treatment prospectively.

BAIO has followed a run of cancer AIs that extract more from material hospitals already collect: GigaTIME inferred expensive tumour-immune imaging from routine slides, another model estimated a breast cancer genomic test, TARIO-2 predicted immunotherapy responders, MutationProjector interpreted treatment response from mutation patterns, and Oncoformer looked for cancer before diagnosis.

We have seen this pattern repeatedly, and we will keep seeing it. Medicine has accumulated enormous stores of trial data, pathology slides, scans and patient records that were analysed only through the tools available at the time. Applying increasingly capable AI to that existing material will continue to uncover biomarkers, responder groups and biological signals that were there all along. There are riches buried in medicine’s backlog.

The US National Science Foundation (NSF) has awarded the University of Maryland $17.3 million to build a physical lab that researchers will be able to program remotely.

At the CRAB Lab, users will submit biomanufacturing workflows through a cloud interface, and automated equipment will carry them out. Biomanufacturing uses living cells to make medicines and other biological products.

The lab is meant to automate more than the physical steps. Developing a biomanufacturing process means repeatedly testing different conditions and measuring how they affect product quality and yield. But conventional molecular measurements can be too slow to provide the rapid feedback needed to optimize each experiment and scale the process.

CRAB Lab will combine those molecular tests with a new electronic measurement tool that can collect much larger streams of data much faster. AI models will search those signals for patterns linked to product quality, yield and other important process measures, helping researchers analyse experiments and improve the next ones.

The four-year project will draw on Ginkgo Bioworks’ bioprocess-automation expertise and initially serve a consortium of 30 biopharmaceutical companies. It is one of 20 sites in a national programmable cloud-lab network backed by $380 million from NSF and more than $20 million in philanthropic support from the Astera Institute.

“AI paired with autonomous labs is how discoveries get made and scaled at the same time,” said Jason Kelly, co-founder and CEO of Ginkgo Bioworks, in a press release. “The CRAB Lab gives researchers and manufacturers a shared, real-time feedback loop that could move biomanufacturing innovation from years to months.”

The award funds the creation of a test bed, not an autonomous lab that has already demonstrated these capabilities. UMD has not said when researchers outside the project will be able to begin using it.

AI may help discover a biological product without solving how to manufacture it consistently at commercial scale. CRAB Lab brings autonomous experimentation to that later problem - and, as a shared remote facility, could give smaller biotechs access to process-development tools they could not build themselves.

Google DeepMind has dispersed the dedicated team behind AlphaFold, according to the Financial Times. Most researchers who authored the original papers have been reassigned over the past year to Gemini-powered science systems, enzyme design, genomics and nuclear fusion. Others moved to Alphabet’s drug-discovery company Isomorphic Labs. Nearly a quarter of the full-time DeepMind authors have left the company.

BAIO covered AlphaFold co-creator John Jumper’s move to Anthropic in Issue 35. The wider reshuffle does not signal a retreat from biology: Isomorphic Labs is extending AlphaFold’s drug-discovery lineage through IsoDDE, and DeepMind remains active across biology and AI-powered science.

DeepMind research vice-president Pushmeet Kohli disputed the idea that the company has pivoted away from hard biological problems. He said the science team had expanded its work and that AlphaFold researchers had agency over which new challenges to pursue.

AlphaFold came from concentrating an interdisciplinary team on one grand challenge. Many of those researchers are now spreading into other scientific problems and more general AI-for-science systems. That may be a natural evolution after AlphaFold’s success rather than a retreat - but it also tests whether DeepMind can reproduce that success without the same concentrated team.

Benchling has opened its Model Hub to anyone, not just Benchling customers. The web tool lets researchers run protein-structure prediction models without installing software or arranging access to expensive computing hardware.

Upload a protein sequence in FASTA - the standard plain-text format for storing biological sequences - and choose from models including AlphaFold2, Boltz-2, Chai-1 and OpenFold3. This makes it easy to compare how several models predict the same protein.

Personal accounts get 10 free predictions a month, while verified life-science work emails get 250. Benchling says it will not train AI models on uploaded data without permission, although it may use inputs and outputs to operate, evaluate and improve the service. Researchers handling proprietary sequences should still review the terms and their institution’s policies.

☑️ EMBL AI Librarian: Life-sciences knowledge layer for AI agents.

☑️ A prospective test puts an ADMET foundation model against real compounds.

☑️ Virtual cells need proteomics, not RNA alone.

☑️ A 3D human anatomy application built with Three.js using GPT 5.6 Sol.

☑️ A new benchmark asks whether antimicrobial-peptide models predict useful drugs.

☑️ ICON and Anthropic bring Claude into clinical-trial workflows.