Hi everyone, I have a quick platform update before today’s very full issue: we’ve launched the AI × Bio Tools & Resources Directory.

It’s a searchable, filterable index of the tools, models, datasets, benchmarks and platforms we’ve surfaced in BAIO - from literature agents and virtual cell benchmarks to protein design tools and lab automation platforms. We’ll update it periodically. For now it should cover every issue except this one. If you spot any errors or want to nominate a candidate to include - feel free to reach out!

In today’s issue, we have AI turning ECGs into richer heart readouts, Penn using LLMs to help choose a CAR T target, Aleph imaging blood flow in the living brain through the skull, a garage drug-discovery story that blew up on X, and another rude benchmark awakening for the virtual cell field.

Let’s dive in.

Two AI systems are pushing the humble ECG beyond what doctors usually read from it.

First, Pathway Labs’ EchoNext has received FDA clearance to flag six high-risk heart conditions from a standard 12-lead ECG - the cheap, routine test that records the heart’s electrical activity. Patients flagged by the model can then be sent for an echocardiogram, the ultrasound test used to diagnose problems such as valve disease, cardiomyopathy, pulmonary hypertension and severe thickening of the heart.

The Columbia and NewYork-Presbyterian team behind EchoNext trained the model on more than 700,000 ECG-echocardiogram pairs, and Pathway says validation studies now span more than 20 hospitals and 500,000 patients.

The Nature paper pushes the ECG idea further. A team from UC Berkeley, MIT, the University of Chicago and Swedish health institutions trained a deep learning model on ECGs from one Swedish healthcare region, linking the waveforms to health records and death certificates to ask whether a cheap electrical readout could identify patients at risk of sudden cardiac death.

Today’s main risk marker is LVEF - left ventricular ejection fraction, a measure from ultrasound of how strongly the heart’s main pumping chamber squeezes. It is the best widely used tool clinicians have, but a limited one. The ECG model identified a high-risk group with a 7.0% annual sudden cardiac death rate, compared with 4.6% among patients with reduced LVEF. Crucially, 86.1% of the model’s high-risk patients were not flagged by LVEF.

Then the team used a generative ECG model to ask what would have to change in a low-risk ECG for the risk model to see it as high-risk. That produced synthetic high-risk ECG beats and pointed to lead aVL, one of the ECG’s recording angles. In that view, the model highlighted a slurred shape that the authors say had not previously been described. The team quantified the shape and found that it carried independent predictive signal in the main dataset and an external US cohort, with a similar but less certain pattern in the smaller Taiwan registry. It still needs prospective testing before it can guide treatment.

Why it matters: BAIO covered EchoJEPA way back in Issue 2: a model trained on 18 million heart ultrasounds. This story starts with an even more routine measurement. ECGs are already taken across healthcare, often before anyone orders cardiac imaging or specialist follow-up. EchoNext uses that signal to decide who should get an ultrasound. The Nature paper goes further, showing that the same kind of trace may contain a sudden-death signal clinicians have not been using. It is part of a pattern we keep returning to: AI in medicine may move fastest where the data already exists, and the signal was hiding in plain sight.

Did you know? The team behind the Nature paper released code for the pipeline it used to turn ECG beats into synthetic higher-risk versions.

Last issue, we told you how Nabla Bio used AI to design a molecule that could recognize a difficult cancer signal. Now, Penn Medicine researchers are moving one step upstream.

In a Cell paper, a team from the Perelman School of Medicine at the University of Pennsylvania and Penn’s Abramson Cancer Center reports using large language models to help choose which cancer signal a CAR T cell therapy should attack in the first place.

A short recap: CAR T cells need something to grab. That is not simple. The target has to sit on the surface of cancer cells, show up strongly enough to matter, and be absent or limited enough in healthy tissue that engineered immune cells do not attack the wrong place.

The Penn-led team, with CAR T pioneer Carl June as senior author, used an LLM-assisted pipeline to search for that kind of target. The team started with public single-cell RNA sequencing data from human skin cancers and healthy tissue, then added information from public databases to score more than 10,000 candidate genes on features that matter for CAR T therapy: tumor enrichment, surface localization, tissue specificity and clinical feasibility.

Several frontier LLMs were then asked to nominate targets from the filtered list. The simulations were repeated 1,000 times, and scientists reviewed the resulting shortlist.

The top hit was GPNMB, a protein found on the cell surface. The team checked that it appeared across both blood cancers and solid tumors, then built a human CAR T cell against it. In mouse models of leukemia, melanoma and colorectal cancer, the paper reports that the engineered cells eliminated tumors.

As for the pipeline itself, the authors argue that the same approach could be used beyond one cancer type, and potentially beyond cancer, because the framework is not tied to a single disease. It asks a more general question: which diseased cells can be targeted safely, and which surface marker makes that attack possible?

Why it matters: BAIO recently covered Waypoint Bio, which uses AI and spatial biology to test how engineered cell therapies behave inside tumors. Penn is working on a different part of the same problem: before you optimize a CAR T cell, you need to know what it should recognize. That target choice has kept CAR T strongest in blood cancers and harder to expand into solid tumors. If LLMs can help sift cell atlases, public databases and translational constraints into better target shortlists, they could make CAR T discovery less artisanal - and easier to apply across more diseases.

Did you know? The source code for the target-discovery framework is on GitHub, and the integrated skin-cancer dataset is on Zenodo.

Medra has launched AI Experimentalist, the “scientific reasoning layer” for the autonomous robot lab BAIO covered back in April. The company also announced a DARPA-funded project to develop the system inside ML001, its autonomous lab in San Francisco.

The idea is to close more of the experimental loop. A scientist sets a goal and constraints in chat. AI Experimentalist reviews the context, designs a workflow, coordinates the wet-lab run, analyzes the results and proposes the next protocol.

Medra chose to illustrate the system with a practical wet-lab task: testing whether an antibody binds its target.

Before that test can happen, scientists first need to make the antibody protein. In Medra’s plasmid-based workflow, that meant cloning: putting the DNA sequence that encodes the antibody into a circular DNA carrier, called a plasmid, so the protein-production machinery can read it and make the antibody. That step takes time.

AI Experimentalist identified cloning as the bottleneck and redesigned the workflow around linear DNA templates, which can be generated and used more directly in cell-free protein production. That removed two days. But it created a second problem: linear DNA templates are usually less stable than plasmids and often produce less protein. Medra says the system then tested multiple template designs and protocol conditions in parallel until expression recovered enough for binding measurements. But “that compounding effect,” Medra says, “is only possible when scientific reasoning is directly connected to wet-lab execution.”

The result, according to the company: an antibody-screening workflow cut from about three days to roughly 14 hours.

Why it matters: Medra is not alone here - self-driving labs are emerging across companies and academic groups in the US, China and elsewhere - but the company has a knack for making its launches feel big. And when DARPA is on board, you pay attention. The thing to watch now is whether Medra, or one of its competitors, can turn the promise into a real learning system: an AI scientist whose next experiment improves because of what happened in the last one.

Did you know? Partners can access AI Experimentalist through Medra labs deployed on site, or remotely through ML001 in San Francisco.

Just a week ago, we told you about the backlash to the “Midjourney Scanner,” including the confident objection that ultrasound cannot simply look through bone.

Aleph’s new result is an almost comically well-timed, “well, about that…”

Aleph is a San Francisco neurotechnology lab, building ultrasound-based brain interfaces. Its long-term aim is a “general-purpose mind interface”; the current result is a through-skull vascular-imaging milestone.

There’s actually a direct connection to the Midjourney Scanner. The Midjourney prototype uses tech from the ultrasound-on-chip company Butterfly Network. Aleph is part of Butterfly’s Embedded program and is using Butterfly’s chip-based ultrasound platform.

“Researchers like Aleph Neuro are using Butterfly’s Ultrasound-on-Chip to reimagine the capabilities of ultrasound and challenge the status quo,” said Joseph DeVivo, president, CEO and chair of Butterfly Network, reported by MassDevice.

The lab says it has produced the most detailed vascular image yet of a living human brain captured with ultrasound through the intact skull. The result is a 3D map of blood vessels and blood flow, made with ultrasound localization microscopy (I encourage you to visit the lab’s website to see it in action).

The method uses microbubbles: tiny gas-filled contrast agents infused into the bloodstream. Blood carries them through the brain’s vessels, where they reflect ultrasound much more strongly than surrounding tissue.

Each bubble is smaller than ultrasound can resolve directly. On a normal ultrasound image, a single bubble appears as a blurry spot. But one of the (many) big ideas here is that if the bubbles are sparse enough, those blurry spots do not overlap. Software can then estimate the center of each spot with much higher precision than the spot itself.

As the bubbles flow, the system repeats that process frame by frame: locate a bubble, track where it moves, then add that position to the map. Over a four-minute scan, millions of localized bubble positions can be stacked into a detailed 3D reconstruction.

And what about AI? Well, Aleph argues that current ultrasound pipelines throw away too much raw signal: a probe can receive terabytes of data per hour, while the standard processing pipeline compresses that down to 0.1% of the original. The company’s convinced that end-to-end machine learning, trained on much larger neurovascular ultrasound datasets, can recover more of that signal - and eventually move toward contrast-free brain imaging.

Why it matters: Aleph is not shy about the end goal: brain interfaces for what it calls a “telepathic future.” But the here-and-now is already sci-fi enough. The company has shown a way to map blood flow in a living human brain through the intact skull, using ultrasound, contrast bubbles and computation. If the approach holds up - this is a technical blog, not a peer-reviewed paper, mind you - the payoff is not just brain-computer interfaces. Aleph argues that diseases such as stroke, Alzheimer’s and traumatic brain injury leave vascular signatures that today’s imaging can miss.

Did you know? Aleph released its processing pipeline on GitHub, including a command that downloads a public 98GB ultrasound sample dataset and builds an interactive 3D track-flow viewer.

A chemistry-lab-in-a-garage story just blew up on X.

Douglas Yao, founder of Pace Pharmaceuticals, says he designed and synthesized PAC-832, an Alzheimer’s drug candidate, with help from modern AI tools, a liquid-handling robot and a garage chemistry setup. “I wouldn’t have been able to complete this project without them,” Yao writes.

PAC-832 is meant to block GalR1, a galanin receptor tied to acetylcholine release and memory. In a non-peer-reviewed manuscript, Yao reports cell assays, brain exposure in mice and improved performance in two mouse memory tests where memory impairment had been temporarily induced.

There are caveats galore here: the mouse tests modeled temporary memory impairment, not Alzheimer’s disease itself, and the drug has not been tested in people. Yao says the drug candidate is “currently undergoing IND-enabling studies.”

Why it matters: BAIO has covered the dog cancer vaccine, home genome sequencing and browser-based protein redesign. This is the same falling-barrier story, but closer to drug development itself. The question right now is less whether PAC-832 works (even though we’re rooting for Yao). The question is what happens when AI agents, cheap robotics, contract research and garage labs let more people try. And the answer is likely going to be interesting.

Did you know? Yao says a liquid-handling robot programmed with Claude Code ran the in vitro screening, while LLMs , especially ChatGPT Pro, were integrated into “virtually every step” of the project.

Applied Scientific Intelligence (ASI) has built VCBench, a new benchmark for testing whether single-cell foundation models are actually moving toward virtual cells.

A new preprint evaluates five models - Geneformer, scGPT, UCE, TranscriptFormer and Arc State - across five testable capabilities: perturbation prediction, cross-species transfer, gene-regulatory-network inference, cross-modal prediction and temporal ordering.

The results ain’t pretty. As ASI’s Lukas Weidener put it on X, the models were tested against “simple baselines you could write in an afternoon.” On four of five tasks, “the afternoon code” won. Only TranscriptFormer clearly beat the strongest baseline, on RNA-to-protein prediction. Even there, the paper flags a tradeoff: the same representation that helped with protein prediction damaged its ability to order cells across time.

Why it matters: This extends a warning BAIO covered in Issue 10, when a Nature Biotechnology review argued that biological AI is still full of proof-of-concept benchmarks, weak standardization and models that do not consistently beat simple baselines. VCBench turns that warning into a harsher virtual cell audit. The promise is still huge. But for investors betting that bigger single-cell foundation models are already becoming virtual cells, this is another rude awakening.

Did you know? BAIO first covered ASI in Issue 21, when its Alexandria literature agent tried to read scientific figures as evidence. It returned in Issue 28 with RefusalBench, a benchmark for how frontier models handle biological research prompts across risk levels.

Insilico Medicine has signed another enormous-looking AI drug discovery deal. This one is with SK Biopharmaceuticals, the South Korean CNS (central nervous system) drug company behind the epilepsy medicine Xcopri.

The headline number is up to $2.5 billion. The real near-term number is smaller: Insilico is eligible for up to $18 million in upfront and near-term milestone payments. The rest depends on development, regulatory and commercial milestones, plus single-digit royalties if any drugs reach the market.

The collaboration will focus on neuroimmune disorders: conditions where the nervous system and immune biology intersect, including neuroinflammatory, neurodegenerative and rare neurological diseases. Insilico will use its Pharma.AI platform for target validation, generative chemistry and molecule optimization. SK brings CNS development and commercialization.

Why it matters: BAIO covered Insilico’s $2.75 billion Lilly deal in Issue 12. And the company has never been shy about its ambitions. In a new Wall Street Journal profile, Alex Zhavoronkov talks about making people live “much, much longer” and “creating the Olympus of drugs,” while the company’s latest deal adds another $2.5 billion headline number to its AI drug discovery story. But the SK agreement is also a reminder to read the small print. Only up to $18 million is upfront or near-term. The rest depends on drugs surviving development, regulators and the market.

Did you know? Insilico is hiring.

AI Bio IQ turns the benchmark sprawl around life-science models into one navigable leaderboard. It pulls public, source-backed results across biology, biosecurity, pharma R&D, therapeutic discovery and bioinformatics, then shows model scores, tradeoffs and underlying benchmarks.

☑️ Aidoc gets FDA Breakthrough status for AI-drafted radiology reports.

☑️ Boehringer buys into Immunai’s single-cell AI platform.

☑️ Post-training can make biology models better - and worse.

☑️ Tessel releases Sesame for pocket-conditioned molecule generation.

☑️ AI drug discovery needs trustworthy biological data.

☑️ Reya.ai joins NVIDIA Inception.