Explorers used to sail toward the edge of the map and write “here be dragons” where knowledge ran out. Biology has its own blank spaces: the surface of a cell, the immune neighborhoods inside tumors, the mutation patterns that make one cancer respond and another resist. This issue has several attempts to replace dragons with coordinates in biology’s unmapped terrain.

Elsewhere today: OpenAI is giving vetted biodefense groups access to GPT-Rosalind. AutoScientists tests a decentralized setup for AI research agents in long-running computational experiments.

Also: if you’re new to BAIO (welcome 👋) or just want to catch up, I published a Top 10 recap of the month’s biggest AI × bio stories - world models for proteins, robot hands in the wet lab, AI scientists in Nature, Isomorphic’s $2.1B raise, and more.

Let’s dive in.

Deliverome launched with $5 million from Radial, Astera's life sciences division, to build an open atlas of the human surfaceome - cell-surface proteins drugs can use as addresses. It aims to increase actionable precision-delivery targets tenfold within five years.

AI can help design antibodies, gene editors, RNA drugs, and protein payloads. But getting those therapies to the right tissue, into the right cells, and away from the wrong ones is still hard. Outside the liver, it remains a brutal bottleneck.

Why is this still the case? More than anything, it’s a market failure. Pharma has little reason to publish a map competitors can use. Startups are pushed toward near-term drug candidates and defensible IP. Academia lacks both the resources and the incentives. “The field urgently needs a complete deliverome atlas, but no single player in the current system has both the capability and the incentive to create one,” Deliverome concludes.

So the industry keeps returning to the same small set of delivery targets - not necessarily because they are best, but because they are mapped well enough. As a nonprofit focused research organization, Deliverome is structured to do the thing everyone needs and no one is properly rewarded for doing: build the shared map.

The project will measure abundance and specificity, whether targets pull cargo into cells, and where that cargo goes after entry. As an example, Deliverome says more than 80% of antibody-drug conjugates in development target the same 11 surface proteins - fewer than 0.5% of the available surfaceome.

Why it matters: Delivery is the part that turns therapies into something that can actually help a patient. Better maps could make it possible to send RNA drugs, gene editors, and targeted cancer therapies to cell types the field struggles to reach today. Deliverome is not a therapy, but if it works, it could widen the set of medicines that are realistic to build.

Did you know? Deliverome plans to release protocols, reagents, raw data, and analytical outputs openly in AI-ready formats, with release notes roughly quarterly. Deliverome is also hiring a founding scientific team.

GPT-Rosalind, OpenAI's dedicated life-sciences model, is being routed into biosecurity. A new program called Rosalind Biodefense will give vetted developers and selected government partners access to the model for pandemic preparedness and biological threat response.

BAIO covered GPT-Rosalind in Issue 18, when OpenAI introduced it as its first dedicated life-sciences model and Novo Nordisk signed up. This is the next chapter: not just biology AI for faster drug discovery, but controlled access when the same capabilities touch biosecurity.

OpenAI says the program is aimed at defensive work: outbreak detection, diagnostic screening, epidemiological modeling, medical countermeasures, and preparedness planning. Access is restricted to trusted developers, selected U.S. government agencies, and allied partners.

In our last issue, RefusalBench showed the model-behavior side of AI biosecurity: refuse too much and you slow legitimate science; refuse too little and you create real risk. With Rosalind Biodefense, OpenAI is approaching the problem from the access side instead: give more powerful biology tools to vetted groups working on biodefense and pandemic preparedness.

Why it matters: Anthropic used this strategy for cybersecurity with Mythos. OpenAI is now trying something similar for biology with Rosalind: route frontier capabilities toward trusted defenders before those capabilities become easier to misuse.

Did you know? Institutions can request access here.

Routine cancer slides can now be used to map immune structures forming inside tumors, according to a new Science paper. The structures are called tertiary lymphoid structures, or TLSs - small immune-cell hubs where B cells, T cells and other immune players gather near the cancer. Their presence has been linked to better outcomes and stronger responses to immunotherapy, but simply asking whether a tumor has TLSs misses a lot.

Researchers at MD Anderson Cancer Center in Texas built two connected steps. First, they used detailed spatial biology data from 340 tumor samples across 12 cancer types to map how TLSs vary by maturity, location and cellular makeup. Then they developed an AI framework that can identify and classify TLSs from routine pathology images.

That second step allowed the team to scale the analysis to more than 25,000 TLSs from over 3,000 whole-slide images across 10 patient groups. Their TLS composition score - which captures how many TLSs are present and how mature they are - outperformed conventional TLS measures for stratifying patients by prognosis and treatment response.

Why it matters: BAIO has covered several attempts to squeeze more biology out of routine cancer slides. This one adds a specific layer: the organization of local immune hubs inside tumors. If prospective validation holds, TLS profiling could help explain why some immune-active tumors respond to treatment while others do not - not just by asking whether immune cells are present, but whether they are organized in the right way.

Did you know? The code used to generate the pan-cancer TLS atlas is available on GitHub.

Tumor sequencing can leave clinicians with a long list of mutations and a short list of useful answers. MutationProjector, a new AI model published in Cancer Discovery, tries to close that gap by predicting treatment response from the broader mutation pattern in a tumor.

The model was built by researchers at UC San Diego and trained on more than 30,000 tumors across 10 solid cancer types. It also uses molecular network data - information about which genes and proteins interact - so the model can interpret mutations in context, not one gene at a time.

That is important because cancer mutations rarely act alone. A mutation linked to sensitivity to one drug can mean something different when other relevant genes are altered too. The paper gives cisplatin (a chemotherapy medication) as an example: ERCC2 mutations have been linked to sensitivity, but the effect varies with other DNA-repair genes such as BRCA2, ATM, RB1 and FANCC. MutationProjector turns those combinations into a tumor representation that can be reused across clinical tasks.

In retrospective tests, the model predicted immunotherapy response and survival patterns in bladder cancer, lung cancer and melanoma, and chemotherapy response in bladder cancer. It compared favorably against existing biomarkers and standard machine-learning baselines.

Why it matters: Genetic testing is already routine in cancer care, but only about 8% of cancer cases are matched to an FDA-approved therapy on the basis of genetics. MutationProjector points to a way of making more of that sequencing useful: interpreting the patterns instead of hunting for one famous mutation. The next test is whether that holds up prospectively, in real treatment decisions.

Did you know? The code used in the study is available on GitHub.

Science rarely follows one clean path: hypotheses compete, evidence changes, and failed ideas need to be remembered. AutoScientists, a preprint from Harvard researchers, tries to emulate that. It gives AI agents something closer to a decentralized research group.

Many AI-scientist systems still follow one research thread or coordinate through an orchestrator. AutoScientists removes the central planner. Agents read a shared experimental state, propose directions, critique weak ideas before experiments run, and self-organize into teams. When progress stalls, teams can reorganize. Results go back into the shared record, successes and failures alike.

This sits next to earlier AI-scientist work BAIO covered: the May 19 Co-Scientist, Robin and ERA papers, and Science Beach's agent peer review.

Beyond AutoScientists’ decentralized setup, the other important piece is duration: it is built “for long-running computational scientific experimentation”.

The results are encouraging. On BioML-Bench, 24 tasks in imaging, drug discovery, protein engineering and single-cell omics, AutoScientists reached a 74.4% mean leaderboard percentile, versus 66.1% for Autoresearch, the strongest prior agent in their BioML-Bench comparison, under the same setup.

Why it matters: Long-running research does not move in a straight line. Promising directions stall, failed experiments need to be remembered, and new ideas often appear only after earlier results are understood. AutoScientists is an early sign that AI agents may be able to handle such work.

Did you know? AutoScientists code is available on GitHub.

Science Superpowers is an open-source methodology for AI-assisted research. It gives agents a structured workflow for doing science more carefully: define the question, pre-register the hypothesis and analysis plan, run the work reproducibly, check anomalies, verify claims, and archive the result. It works with Claude Code, Cursor, Codex, Gemini CLI, OpenCode and Antigravity.

☑️ Benchling Automation: Closing the lab-in-a-loop.

☑️ CortexMAE brings foundation models to open fMRI data.

☑️ Why genome foundation models are harder than protein models.

☑️ Bonito and DaltonTx team up on AI-driven delivery for oligotherapeutics.

☑️ Protuoso raises $9.5M for AI-designed multi-function biologics.