If there’s one thing I’ve been bullish on in biology for the better part of the last five years, it’s partial reprogramming: the idea that old cells can be nudged back toward a younger state without losing identity.

The biotech industry, broadly speaking, seems to agree. Altos Labs, Life Biosciences, NewLimit and others have made cellular reprogramming one of the biggest bets in aging biology. To the point where one starts to wonder what happens to longevity science if turning back the clock does not turn out to be the fountain of youth we all hope for.

But so far, so good. NewLimit just raised $435 million and says its first aging-reprogramming medicine is headed for human trials next year.

In this issue we also look at a cancer-genome model that could help choose between chemotherapy regimens, an AI-designed cell-therapy startup moving toward the clinic, and three new benchmarks asking whether bio agents can do more than produce plausible-looking work.

Let’s dive in.

If there’s a limit to NewLimit, the company has yet to reach it. The aging-reprogramming startup just raised $435 million and says its first medicine is headed for human trials next year.

The South San Francisco company was founded by Coinbase CEO Brian Armstrong, former GV partner Blake Byers, and computational biologist Jacob Kimmel. Its bet is that aging can be treated by changing the epigenome - the control layer that decides which genes a cell uses - so old cells regain youthful function.

From the very beginning NewLimit has used AI to discover aging reprogramming medicines: combinations of transcription factors, the proteins that switch gene programs on and off. The company tests thousands of designs in old human cells, reads the results with genomics, and feeds that data back into the next models.

The first program is liver aging. NewLimit says its prototype medicine reversed cell age in old human liver cells. In preclinical work, the company says its liver reprogramming therapy helped old livers heal faster after injury, avoid damage from dietary challenges, and recover faster from alcohol exposure. And pay attention to the timeline: when the company launched, it thought reaching human trials would take until at least 2031. Now it says the first program is headed for the clinic in 2027.

Why it matters: Cellular reprogramming has arguably become the biggest bet in aging biology, with Altos Labs, Life Biosciences, NewLimit and others chasing the idea that old cells can be pushed back toward youthful function. With Life Biosciences already moving into human testing and NewLimit now aiming for 2027, the field is about to get its first early verdict.

Did you know? NewLimit is hiring.

A Weill Cornell-led team has released TESSERA, an AI model that tries to read a tumor’s genome as a whole.

When a cancer patient gets tumor sequencing, the result is not one clean answer. It is a messy list of DNA changes. Some are tiny spelling changes in the tumor’s genetic code. Others are bigger changes, where cancer cells have extra copies of parts of the genome, or have lost parts altogether.

Doctors already use some of these changes to guide treatment. But much of cancer genomics still works like a checklist: does this tumor have mutation X or mutation Y?

TESSERA asks a different question: what if the whole pattern matters?

The clearest example is metastatic colorectal cancer, where doctors often choose between two standard chemotherapy combinations. In a retrospective analysis of 1,452 Memorial Sloan Kettering patients, TESSERA split them into two groups: those the model predicted would do better on one regimen, and those predicted to do better on the other. In both groups, patients who had received the model’s preferred treatment did better.

Why it matters: This is the difference between using tumor DNA to estimate risk and using it to choose between treatments. Last issue’s MutationProjector pointed in a similar direction: tumor mutations can carry more information when read as patterns, not isolated entries on a checklist. TESSERA pushes the idea wider: one reusable model for reading cancer genomes. This is still a preprint, and the treatment result needs proper validation before it changes care.

Did you know? Code is on GitHub, model weights are on Hugging Face, and TCGA cancer-genome representations are on Zenodo.

Waypoint Bio raised $20 million to move AI-designed cell therapies for solid tumors toward the clinic.

The New York startup is working on CAR T therapies, where immune cells are engineered to recognize a target on tumor cells. CAR T has transformed some blood cancers, but solid tumors are harder: the cells have to enter a dense, hostile tumor environment, survive there, and kill the right cells without damaging too much healthy tissue.

Waypoint’s pitch is that AI can design better cell therapies, but only if the tests capture that messy environment. Its platform tests thousands of candidate designs in vivo at the same time, then uses spatial biology - measurements of where cells are and how they interact inside tissue - to rank which designs actually work.

The lead program, WAY-103, targets gastric, gastroesophageal junction, and pancreatic cancers. Waypoint says it showed more than 15-fold improved potency in animal models compared with multiple clinical benchmarks, with reduced on-target/off-tumor toxicity. A trial is planned for late 2026.

Why it matters: AI can generate more therapeutic designs than labs can test one by one. Waypoint is trying to make the experiment itself more scalable - and more realistic - so the model learns from how therapies behave inside tumors, not just from simplified readouts like whether a tumor shrank.

Did you know? Waypoint is hiring.

Biology agents are being built to automate the unwieldy computational work behind modern biology: choosing tools, writing code, managing files, running analyses, and returning results a scientist can inspect.

But they are nowhere near fire-and-forget systems. Producing an output is not the same as producing the right output. So the field is building benchmarks to measure where agents are useful, where they break, and whether they can catch their own mistakes.

Just in the past month, three new tests have attacked that problem from different angles.

☑️ BioAgent Bench came in early May and tested models including Claude Opus 4.5, Gemini 3 Pro, GPT-5.2, Claude Sonnet 4.5 and several open-weight models on common bioinformatics workflows: RNA-seq, variant calling, metagenomics and single-cell analysis. The frontier agents often completed the pipelines. The problem, though, was reliability. When the benchmark added corrupted input files, irrelevant decoy files, or long distracting prompts, agents sometimes continued with bad data, used the wrong files, or failed to finish.

☑️ PromptBio-Bench adds 244 expert-curated bioinformatics and data-science tasks, each packaged with input files and human reference answers. It tested three bioinformatics agents - Biomni, STELLA and ToolsGenie - on whether they could complete real analysis jobs and return the right output. Biomni and ToolsGenie completed nearly all tasks, while STELLA finished fewer. But across all three, accuracy dropped as tasks got harder. In other words: getting to the end of the analysis is not the same as doing the analysis well.

☑️ AutoMedBench moves the same question into medical AI. Instead of asking agents to answer a single medical question, it tested whether six frontier models - including Claude Opus 4.6, GPT-5.4, Gemini 3.1 Pro and open-weight models - could complete full medical-AI workflows across segmentation, image enhancement, visual question answering, report generation and lesion detection. Each agent had to plan, set up code, validate a pilot run, run the full job and submit the final output. The weak spot was validation: agents were better at getting pipelines to run than checking whether those pipelines were reliable before trusting the result.

Why it matters: These results are strikingly similar, and maybe a little deflating. Bio agents can increasingly finish the work, but they still struggle with reliability, correctness and validation. That may sound like a letdown, but it is also how a field moves from cool demos to useful infrastructure. These benchmarks do more than grade today’s agents. They also give developers a blueprint for the next stage.

Did you know? All three papers link to the public benchmark resources if you want to poke around yourself.

Genomi is an open-source, local-first agent harness for personal genomics. Point Claude Code, Codex, OpenClaw, Hermes, or another MCP-capable agent at your raw genome file, and Genomi builds a local SQLite index the agent can query. It covers pharmacogenomics, nutrigenomics, polygenic risk, carrier status, ancestry, and sequence tools. Your raw genome file stays on your machine, though selected evidence snippets may be sent to the model you use. We’re also asked to remember that this is experimental - not clinical, not diagnostic.

☑️ OpenAI Foundation launches AI Resilience grants, including bio-resilience.

☑️ CMonge predicts single-cell drug responses with one shared model across drugs and doses.

☑️ NeuroXT and BIDMC expand AI-driven Alzheimer’s precision-treatment collaboration.

☑️ SwitchCraft: A programmatic framework for designing state-switching proteins.

☑️ Converge Bio Secures $2.5M Grant to expand its fast-track AI gene discovery capabilities.

☑️ Tiny image changes can fool pathology foundation models.

☑️ HVM dataset offers expert-annotated CT maps for AI liver vascular segmentation.

☑️ PepAnno launches structure-aware deep-learning annotation for bioactive peptides.

☑️ Rocklin Lab releases Megascale Open Protein Stability Dataset to advance biomolecular AI.