(Re)Share #60 - Protocol Me Maybe
Generative genomics | Reanimated tissue | Agent payments | Antibody prediction
Greetings from the West coast where I’m spending the week getting caught up on desk work before I’m back on the road. Next week I’m scheduled to be in NYC for the AI + Bio summit, assuming there is still a city to visit. Between the mayoral race and the insane flooding it’s anyone’s guess. Loads of fascinating stuff to cover from the four corners of the deep tech world so let’s get to it.
Shameless Plug
Over the last couple of months I have been researching the norms and benchmarks of EU / UK University TTOs required for startup spin-outs. I think our standards are broken but I want to test that theory by polling the investor community. If you’re an early stage VC and can spare 3 minutes I would greatly appreciate your participation in a short survey to gauge your experience with TTOs and your perspective on how we might improve. In my conversations with founders it seems like there is a lot of misinformation on what investors actually expect in these deals so I will be open sourcing the results to hopefully strengthen our shared ecosystem.
Stuff Worth Sharing
Underpants genomes - DeepMind’s newest release, AlphaGenome, is designed to decode human genomic mysteries at a previously impossible scale. It can process up to 1 million DNA base pairs at single-nucleotide resolution and simultaneously predict thousands of molecular properties (gene start sites, RNA splicing patterns, protein binding, etc.) across hundreds of cell types. It does this by comparing predictions between reference and mutated sequences—an A/B test of sorts. The model was able to efficiently score the impact of genetic variants in seconds and outperformed prior models across two dozen benchmarks at half the compute cost. This marks the first unified model that can handle “long-range” genomic context with high precision, meaning we can now unlock how non-coding mutations reshape gene regulation.
Brain freeze - Spoiler alert: this is the weirdest article of the issue. Biotech startup Bexorg is restoring cellular life to human brains after death—not to revive them, but to test preclinical drugs. They do this by perfusing artificial blood through donor brains that have been pronounced dead (no electrical activity), which can continue cellular life for up to 24 hours. The company currently has four BrainEx perfusion machines and has tested around 100 human brains to date, but has lofty plans to process 500 this year and 1,000 in 2026. Important to note, as it was called out several times in the article, the approach deliberately prevents consciousness, which I guess makes sense ethically / morally but that’s where my interest obviously goes. What this approach does enable is the study of drug distribution, neurological response, and biomarker release monitoring within authentic brain tissue. This obviously raises a lot of bioethical questions, but I find this to be wildly exciting and sorely needed to tackle the growing epidemic of neurodegenerative decline.
Protocol me maybe - I’m refreshing my thinking on Agentic Payments, which have enormous promise but will never achieve it as long as payment rails remain a bottleneck. A new open protocol from Coinbase, x402, proposes a fix by turning HTTP 402 (“Payment Required”) into a functional standard for machine-native payments. Instead of API keys, subscriptions, or prepaid credits, services return a 402 response with pricing and wallet details. Agents could then respond with a signed onchain payment and gain instant access to the paywalled service or dataset all natively built for stablecoins. There is a lot of activity going on with single use / agent-owned virtual cards but I’m broadly uninterested in such an incremental approach. Novel rails are a primitive for micropayments and pay-per-use monetization ($0.01 / article, $0.005 / API call). Given the immense distribution advantage of HTTP I’m super, super excited about this.
Discover band - “AI meets science” is a topic I bring up a lot in (Re)Share as I’m actively exploring it for investment and have penned a few working hypotheses on my blog. This SciAgents paper approaches the topic in a more holistic manner. Rather than pushing the frontier of a specific domain capability (gene therapy, protein modification, etc.), SciAgents presents a multi-agent AI framework designed to automate and enhance scientific discovery more generally. It powers LLMs with ontological knowledge graphs and assigns role-based expertise for MoE collaboration. At the heart of the approach is a team of specialized AI agents, each assigned a specific role in the discovery process. Agents operate on a shared knowledge graph constructed from ~1,000 papers. This multi-agent coordination mimics human scientific workflows, including iterative critique, validation, and refinement, resulting in a scalable, context-aware scientific reasoning engine that can be applied across disciplines. The research optimizes for novelty in hypothesis generation, calibrated through external validation via literature review. As we’ve seen from other leading labs, the MoE architecture is extremely promising.
Anti-poison pill - A new tool called LightShed can remove AI-protective measures that artists embed in digital artwork to prevent AI training models from scraping and learning their styles. We’ve talked about some of the data poisoning solutions in past issues (Nightshade, Metacloak), and I continue to believe there is a meaningful company to be built in generative AI prevention / authentication. LightShed effectively strips away embedded protective layers, allowing images to be ingested and learned by AI models as if no safeguarding existed. It highlights the limitations of adversarial perturbation as a long-term defense and reflects the broader reality of cybersecurity. Measures that once offered security can now be undermined with little effort. As adversarial techniques fall short, the debate will naturally shift toward legal protections and licensing regimes, but I remain convinced that this can only be solved through technological means—at least at scale.
Signal and noise - Researchers at UC Davis have developed a BCI that translates neural signals into speech in near real time, edging closer to restoring natural conversation for those with paralysis. The implant targets the speech motor cortex, decoding phonemes and prosody (speech building blocks—I had to look that up) directly from brain activity. AI then processes signals in 80‑millisecond streams and generates speech within three seconds. Like the well-known startup Eleven Labs, this model can be trained on pre‑injury voice recordings and then synthesized into a personal vocal output. The research currently achieves speech rates of 47–90 words per minute (~160 words/min for natural speech), but it’s clearly a transformative step toward better quality of life for those who have lost speech.
Your antibody is a wonderland - Chai Discovery released Chai 2, a model that delivers zero shot antibody and miniprotein discovery with double digit success rates. The big claim to fame is that in a single round of testing (20 candidates per target) Chai 2 achieved a 16% hit rate for antibodies and identified effective binders for over half of 52 novel antigens. This is notable because it was done without relying on existing templates or extensive lab screening. For miniproteins, Chai 2 showed a 68% hit rate, often reaching picomolar binding affinities. The model was built on a multimodal, all atom architecture that allows it to craft new complementarity determining regions (CDRs) and compresses discovery timelines from months to weeks.
I’m physick and tired of.. - Physical engineering, and specifically design improvement, remains relatively untapped by automation systems. This paper proposes eAGI, a collection of engineering-focused agents that can reason across domains, interface with modeling tools, and generate end-to-end system designs. Think of it as an extension of the MoE solutions, but aimed at mechanistic properties like CFD, FEM, and other structural integrity measures. The piece also introduces a benchmarking framework to evaluate these agents using an adapted version of Bloom’s taxonomy. The hardest tasks—those involving tradeoffs, creativity, and domain nuance—are still out of reach, but the paper is more about laying the groundwork for what’s to come.
Birds of a tether - The Ukraine theater of war has a new player: unjammable fiber optic drones. On the ground reporting suggests that drone strikes now account for the majority of attack missions and 70 to 80 percent of combat casualties. Despite their prevalence, a growing challenge has been the ease of signal jamming through electronic warfare. Fiber optic drones solve this by launching with long reels of optical cable that connect directly to soldier controllers. These systems are nearly immune to jamming, offer high resolution video control, and can reach ranges of up to 20 kilometers. Of course they have those pesky scissors to deal with.