This issue is full of attempts to answer a deceptively simple question: how do we know when the model is actually helping?

☑️ OpenAI has a harder genomics benchmark for scientific agents.

☑️ Valence is using AI to speed up molecular physics without turning it into an uncheckable black box.

☑️ Xaira and Boltz are raising the bar for what counts as a real AI-designed binder.

☑️ COMPASS tries to predict immunotherapy response while pointing to the immune biology behind the score.

☑️ And Arc’s next Virtual Cell Challenge arrives after a first round that reminded everyone how stubbornly hard virtual cells still are.

Before we begin, a reminder that BAIO now has a searchable Tools & Resources Directory, collecting the models, datasets, benchmarks and platforms we surface in the newsletter. I also just published the Top 10 AI × bio stories from June. And subscribers can now download our free report, Top 10 Insights 2026, which pulls together the most important lessons from the first few months of BAIO.

Let’s dive in.

OpenAI says its strongest model, GPT-5.6 Sol, solves 31.5% of GeneBench-Pro - a new benchmark built to test a part of computational biology that is hard to write down as instructions.

GeneBench-Pro asks whether agents can look at messy data, decide what question the data can support, spot artifacts, choose an analysis path, and revise if the first path breaks.

The benchmark has 129 problems across 10 domains. Reviewers estimated a typical problem would take a human expert 20-40 hours.

The tasks are synthetic, but not toy problems. OpenAI builds each dataset from a known underlying process, so the benchmark can grade against ground truth while still allowing for, the company says, “reasonable differences in subjective analytical choices”.

The result is eye-opening. GPT-5 scored below 5% when OpenAI began building the original GeneBench. GPT-5.6 Sol now reaches 28.7%, or 31.5% with Pro mode enabled. OpenAI says it may be saturated by the end of the year.

BAIO has tracked a new wave of benchmarks that ask whether AI systems can do real biological work, not just score well on tidy tasks. A few examples:

☑️ Genentech/Roche’s CompBioBench, which BAIO covered in Issue 16, tested whether agents could solve 100 computational biology tasks with verifiable answers, such as detecting contaminated sequencing data or spotting swapped sample labels.

☑️ GeneBench, covered in Issue 21, moved closer to real genomics work: messy data, minimal guidance, and models that often noticed warning signs but kept following the wrong analysis path.

☑️ VCBench, which we told you about a week ago, aimed at virtual cell models instead, and found that simple baselines still beat leading single-cell foundation models on four of five tasks.

GeneBench-Pro raises the bar again by testing high-level scientific reasoning under quantitative uncertainty.

Why it matters: Sequencing costs have fallen fast. OpenAI says the bottleneck is shifting to interpretation. If agents can take on even part of that work, they could speed up target discovery, translational research, and clinical genomics.

Did you know? OpenAI put 10 representative GeneBench-Pro questions on Hugging Face so researchers can inspect the benchmark format. It is also giving a 50-question subset to Artificial Analysis, the independent model-ranking site.

Valence Labs, Recursion’s AI research arm in Montreal, has developed AquaGen, a generative model that creates full-atom snapshots of drug-like molecules surrounded by water.

The idea behind the complex work is simple enough: biology happens in water, and molecules do not sit still. They twist, flex, bump into solvent molecules and move through many possible states. The standard way to explore those states is molecular dynamics, or MD: a simulation that follows atoms step by tiny step. It is physically faithful, but also sequential, slow and expensive.

AquaGen learns from MD data, then generates many independent molecular states directly. Unlike many AI chemistry models, it does not simply take a molecule and return a black-box property score. It includes every atom of the compound and surrounding water, so researchers can inspect the output and run standard physics calculations on it. Valence calls this a gray-box approach: the AI generates the molecular states, and physics-based calculations turn those states into the final estimate.

For now, Valence shows AquaGen on a narrower but important quantity called absolute hydration free energy. In plain English, that asks how favorable it is for a molecule to move from vacuum into water. A drug molecule in the body is surrounded by water before it reaches a protein target, and its interaction with water affects properties chemists care about.

On 218 molecules held out from training, AquaGen’s median error was 0.93 kcal/mol compared with molecular dynamics. In practice, its estimates were close enough to MD to help chemists decide which molecules deserve more expensive simulation, and which are unlikely to be worth that extra compute.

Valence reports 4-10x speedups over GPU-based MD. And because AquaGen produces full atomic snapshots, researchers can use them as starting points for a short MD follow-up run - a cleanup step that reduced the error further without starting from scratch.

The gray-box framing also connects neatly to Valence’s VCR-Agent work, which BAIO covered a few months back. There, the goal was to make virtual-cell reasoning easier to verify rather than asking scientists to trust a fluent answer. AquaGen applies a similar instinct to molecular physics: make AI faster, but keep the output checkable.

Why it matters: AquaGen is still an early system for solvated drug-like compounds, not full protein-drug binding. But if it scales to harder quantities like permeability or binding free energy, chemists could get useful physical estimates before spending heavier compute or lab time. Valence explicitly frames those harder tasks as future directions, not current results.

Did you know? Valence Labs is hiring.

Boltz, the team behind the open-source Boltz-2 structure-and-affinity model, recently published a protocol for evaluating AI-designed protein binders. Its point was simple: messy early assay signals should not be counted as confirmed binders.

Building directly on that, Xaira Therapeutics, the South San Francisco AI drug discovery company behind X-Cell, is pushing the standard one step further: even a confirmed binder is not enough if the molecule cannot become a drug.

A binder is a designed molecule, often a protein or antibody fragment, that sticks to a target. Xaira calls the higher bar a “progressable binder.” The molecule has to be cleanly measured, stable, manufacturable, low-risk for immune reactions and tied to a real therapeutic program. Otherwise, a “hit” may be scientifically interesting but irrelevant as a drug candidate.

Why it matters: Two things can be true at once. AI-designed molecules can now stick to targets, and there is still a chasm between a binder and a drug candidate. Boltz and Xaira are both trying to define the standards for crossing it.

Did you know? Xaira is hiring.

Checkpoint inhibitors can be life-changing cancer drugs, but many patients do not respond. COMPASS, a pan-cancer foundation model from Harvard Medical School and collaborators including Roche scientists, tries to predict response from gene activity in a tumor.

The model was pretrained on 10,184 tumors across 33 cancer types, then evaluated on 1,133 patients from 16 immunotherapy cohorts. It beat 22 existing methods on average, improving accuracy by 8.5%. The findings are published in Nature Medicine.

COMPASS is built to show some of its work. It can point to immune patterns in the tumor that may help explain why a patient responds or resists treatment.

Why it matters: BAIO has covered a run of cancer AI stories pointing in the same direction. GigaTIME turns routine slides into inferred tumor-immune maps. TARIO-2 predicts immunotherapy response from pathology slides. A Technion-led model estimates a breast-cancer genomic test from tissue images. COMPASS adds tumor gene activity to the same pattern: cancer AI is starting to read treatment-relevant biology, not just detect cancer.

Did you know? COMPASS is available on GitHub, with an interactive prediction server, feature-extraction tool and response-map explorer on a project website.

Takeda has signed an AI drug-discovery deal with Insilico Medicine worth up to $600 million.

Insilico gets $60 million in project initiation fees and near-term payments, with the rest tied to milestones and royalties. Insilico will use its Pharma.AI platform to lead discovery. Takeda gets exclusive worldwide rights to develop, manufacture and sell any resulting drugs.

The news lands days after BAIO covered Insilico’s $2.5 billion SK Biopharmaceuticals deal, and follows its $2.75 billion Lilly agreement in March. Takeda has been assembling its own AI partner stack too: we told you about its $1.7 billion Iambic deal in Issue 2.

Why it matters: Takeda is describing this as an ecosystem strategy, not a one-off deal. “AI is moving quickly, and no single company can build every capability alone,” Christopher Arendt, Takeda’s chief scientific officer, told The Wall Street Journal.

Did you know? Insilico Medicine is hiring.

Arc Institute says the next Virtual Cell Challenge launches August 20, with a new problem, wider metrics and a $100,000 grand prize.

The question is whether this year’s challenge can move beyond the lesson of the first one. The 2025 VCC drew huge interest - 5,000 registrants across 114 countries and over 300 final submissions - but the results were sobering.

That lines up with some of BAIO’s coverage this year. In issue 10 we reported that single-cell foundation models often fail to beat simple baselines on perturbation prediction. Issue 37 covered ASI’s VCBench, where simple baselines beat leading single-cell foundation models on four of five virtual cell tasks.

Labless is a crowdsourced benchmark lab for small research challenges. Its first project, Nanopath, asks people to improve a compact pathology foundation model that trains in one hour on one GPU. Clone the repo, change the training run, submit it to Labless, and your result appears on the shared leaderboard - wins and failed attempts included, so others and agents can learn from the full trail.

☑️ “The AI antibody revolution is real - but the business model is much less obvious”.

☑️ How Claude Science can help virtual cell modeling.

☑️ NorthStrive and Yuva report early cell-model results for AI-picked longevity compounds.

☑️ Mankind Pharma partners with Denovo Sciences for AI-powered drug discovery.

☑️ New Market Pitch maps the AI drug-discovery startup landscape.