OpenProtein.AI is trying to take protein design out of the hands of niche AI ​​teams and directly into the hands of everyday biologists via a browser and free access for academia.

On April 17, MIT News published a story about OpenProtein.AI, a company founded by Tristan Bepler and Tim Lu. The company is building a web platform where biologists can run tools for selecting new proteins and testing their properties without programming. MIT specifically emphasizes two things: access via a no-code interface and free use for academic researchers.

There are already many models at the intersection of AI and biology. The bottleneck now is different: who can use them in everyday lab work.

Protein design—that is, selecting new protein versions for a given task—still too often remains the preserve of teams with their own machine learning engineers, computational resources, and the people capable of putting it all together. The rest are left to read papers and look at pretty graphs. OpenProtein is selling the way out of this trap as a convenient way in. Upload the data, select a task, get a library of variants, mutation and structure predictions, and if you want, you can go deeper through programmatic access.

This is changing the promise itself. Previously, protein AI was often sold as the magic of the model. Here, it's sold as a workspace for biologists. On the product page, the company explicitly states that PoET is available free for academic use, and the documentation includes code-free scripts, tools for working with code, and a list of models where in-house and external solutions are side by side. The emphasis is not only on model quality but also on making it an everyday tool, not a rare feature of a few leading labs.

Open access shouldn't be confused with complete openness. OpenProtein talks about an open ecosystem, but it's building a service that's still accessed through a single company. In their documentation, they list open and closed models side by side. This isn't a world where any university can deploy the entire stack on its own. It's a softer dependency: not on its own infrastructure, but on someone else's convenient one.

The next debate at the intersection of AI and biology concerns access. It boils down to the question of whether biology acceleration tools will remain confined to a few large players, or whether they will truly be brought to mainstream labs. If this doesn't happen, the entire conversation about accelerating science risks becoming a mere window dressing instead of a mass-scale research machine.

OpenProtein has a real scientific foundation: the PoET model is described in a paper for the NeurIPS 2023 conference on machine learning. For now, this is primarily an infrastructural shift. There are efforts to package AI for proteins for a much wider range of labs that need to more quickly understand which proteins are worth bringing to real experiments.

NaFM trains AI to search for drug candidates among natural molecules by associating the chemical scaffold with the organism and genes that could have assembled it.

On April 29, Nature Machine Intelligence published a paper on NaFM—a basic model for small natural molecules from microbes, plants, and animals. The authors are testing the idea that it is useful to read a molecule along with its biological origin, because living organisms assemble such compounds through genetic and enzymatic pathways. The testing is currently computational: molecule classification, activity prediction, and computer selection of candidates.

In the paper by Yuheng Ding, Bo Qiang, Shaoning Li, and colleagues, NaFM is trained on natural compounds—small molecules produced by bacteria, fungi, plants, and animals. Drugs have already been developed from such molecules: penicillin came from mold, taxol from yew, and rapamycin from a soil bacterium.

AI-based search for natural chemistry has already become a platform race in its own right. Brightseed's Forager maintains a database of 11 million natural compounds and has already translated two discovered molecules into short human studies. NaFM takes a different approach: it teaches the model to read a molecule along with its biological origin.

A typical molecular model views a compound as a diagram: atoms are dots, chemical bonds are lines. NaFM adds to this diagram the source of the molecule: which organism might have produced it and which gene regions or protein families might have been involved in its assembly.

In microbes, the instructions for assembling many natural molecules are often located close together in the DNA. These regions are called biosynthetic gene clusters: the genes encode proteins, the proteins participate in the assembly of the chemical scaffold, and the scaffold helps understand the class of the molecule and possible target proteins.

The genomes of organisms themselves are also becoming raw material for drug discovery: the America's Living Library Act proposes turning plants, fungi, animals, and microbes from US national parks into a public genomic database for AI-powered drug discovery. NaFM demonstrates why this layer is needed within the model: the connection between the organism, the assembly route, and the final chemical structure is important.

NaFM separately takes into account the molecule's chemical backbone and side groups, which alter activity and selectivity. During training, the model compared similar molecules and reconstructed hidden parts of the molecular structure. The authors deliberately obscured an entire fragment of the backbone so that the model learned the overall structure without relying on guessing based on adjacent bonds.

Verification remains computer-based for now. In the task of classifying natural molecules, NaFM yielded the best average result across different numbers of examples per class. With four examples per class, its result was 70.10 versus 69.17 for the closest method; with 64 examples, it was 91.75 versus 91.07.

In predicting biological activity, the authors compared the average error: the lower the error, the more accurate the prediction. NaFM was the best for seven of the eight target proteins, including PTP1B, acetylcholinesterase, COX-1, and COX-2. On HIV-1 reverse transcriptase, it was slightly outperformed by the N-Gram method: 1.0606 versus 1.0802 for NaFM.

The margin of error is this: the authors demonstrated a computational advantage, but laboratory and clinical tests of the candidates from this study have not yet been published. The NaFM-Official code is open source, the data is hosted on Figshare, and the model weights are hosted on Zenodo. Other groups can repeat the calculations and validate the model on their own sets of molecules.

In 2023, a review in Nature Reviews Drug Discovery described the same problem more broadly: AI can discover hidden diversity in natural molecules when the field has good data and rigorous algorithm validation.

NaFM offers one option for such validation: the model reads the molecule along with the organism and the genetic route of its assembly.

For age-related diseases, this is the early stage of the search: selecting molecules worth synthesizing and testing in cells, tissues, and animals. The next test for NaFM is straightforward: third-party labs take open-label compounds, run their natural molecules, and see if the computational advantage translates into compounds that work experimentally.

Is 'rescuing failed drugs with AI' a category now?

Just two days ago, Biossil, a Toronto startup, came out of stealth with $43M co-led by Peter Thiel’s Founders Fund and OpenAI — and their thesis is a bit different from what most AI biopharma companies are doing.

Instead of designing new molecules, they use AI to dig through late-stage clinical failures and figure out which patient subgroups those drugs should have actually been tested on. Ten molecules acquired quietly over three years, trials running in everything from glioblastoma to Alzheimer's.

There's a well-known strategy in pharma called drug repurposing. Basically, taking an approved drug and finding it a new indication, like thalidomide going from its original (disastrous) use to multiple myeloma, or metformin being studied in cancer.

Biossil is doing something more subtle: same molecule, same disease, just a more precisely defined subset of patients.

The argument is that many drugs "failed" trials only in the aggregate, averaged across a heterogeneous population where a real signal got buried.

They're not alone in this territory, however…

For instance, Lantern Pharma Inc. (Nasdaq: LTRN), a biotech company focused on oncology, has been working a related playbook for years — using their RADR AI platform to identify abandoned clinical-stage drugs and match them to the patient subgroups most likely to respond. Different technical approach, narrower therapeutic focus, but a similar underlying bet.

They're also further along in generating clinical data, which makes them worth watching as a reference point for whether the thesis holds up.

NOETIK — trains AI models on massive datasets of paired pathology images and spatial transcriptomics to find hidden biological subtypes among trial participants and predict which patients will respond.

BPGbio, Inc. — uses Bayesian causal AI to do post-hoc subgroup analysis on failed trials. In one Phase Ib oncology trial, their models identified a subgroup with a distinct metabolic phenotype that showed significantly stronger responses.

The list goes on... but the difference is, of course, in the details.

Image credit: Biossil

🟢 CureForge AI joins the NVIDIA Inception Program

We’ve been accepted into NVIDIA Inception — joining the global program that backs AI-native companies building on NVIDIA’s accelerated computing stack.

CureForge AI is the first end-to-end autonomous research federation built for one mission: solve biological mortality.

36 institutes across 8 tiers, coordinated by an autonomous AI layer with non-negotiable human-in-the-loop oversight on every consequential decision.

Pharma’s productivity has collapsed under Eroom’s Law — costs doubling roughly every nine years per approved drug.

The single-target, single-lab model is structurally broken.

Our answer isn’t another lab. It’s a federation that simulates, hypothesizes, audits, and iterates across every relevant field of life science in parallel — at compute scale.

We’re at architecture-validation stage today.

NVIDIA Inception gives us deeper access to the compute substrate this mission requires: BioNeMo, CUDA-X, NIM microservices, and the full accelerated-computing stack our agents already run on.

The federation is the differentiator. The compute makes it possible.

#NVIDIAInception #CureForgeAI #LongevityInTime #AgenticAI #Longevity

The Longevity Gene

The most well-known gene associated with aging and Alzheimer's disease is APOE. It has several variants: APOE2, APOE3, and APOE4. APOE4 is considered the main genetic risk factor for late-onset Alzheimer's, while APOE2 has long been associated with longevity and a reduced risk of dementia. But until now, scientists haven't fully understood why APOE2 specifically protects the brain.

Previously, APOE was primarily studied as a gene associated with cholesterol transport and amyloid plaque accumulation in Alzheimer's. But new research reveals a much more fundamental role: APOE2 influences cells' ability to maintain genomic stability.

The scientists used human induced pluripotent stem cells (iPSCs). This technology allows for the "reprogramming" of normal cells back to an embryonic state and then transforming them into neurons. The researchers created cells that differed only in their APOE gene version. Otherwise, their genetic makeup was identical.

These cells were used to grow two types of neurons: inhibitory GABA neurons and excitatory glutamate neurons. The effects of APOE2, APOE3, and APOE4 on cell aging were then compared.

The main result: neurons with APOE2 accumulated significantly less DNA damage.

The scientists directly measured DNA breaks and discovered that APOE2 activates repair pathways—systems for restoring genetic material. Essentially, cells with APOE2 are more effective at "repairing" their own genome.

Furthermore, APOE2 protected cells from senescence, a process called cellular aging. Senescent cells no longer function normally, but they also do not die. They accumulate with age, cause chronic inflammation, and are considered one of the key mechanisms of aging. To test the cells' resilience, the researchers irradiated neurons and exposed them to doxorubicin, a harsh chemotherapeutic drug that severely damages DNA.

Neurons with APOE2 showed significantly fewer signs of aging after this stress. They had better nuclear structure preservation, less nuclear membrane degradation, and lower levels of senescence markers like p16 and CRYAB.

Most interestingly, when the scientists added the APOE2 protein to neurons with the harmful APOE4 protein, the level of DNA damage decreased. This suggests that the protective effect of APOE2 could potentially be therapeutically transferred, not just genetically inherited.

The results were confirmed in animal models. In aged mice with human APOE2, the hippocampus—a brain region critical for memory—appeared younger. They had better chromatin structure preservation, a more stable nuclear membrane, and higher levels of Lamin A/C, a protein that maintains the integrity of the cell nucleus.

https://onlinelibrary.wiley.com/doi/10.1111/acel.70494

Longevity Escape Velocity (LEV)

The concept is very simple: medicine begins to extend your remaining life faster than time itself.

In other words, science will be able to add more than a year of life to every year lived. If this rate can be maintained, aging will cease to be an inevitable death sentence.

The term was popularized by Aubrey de Grey, who compared it to the physics concept of "escape velocity." Just as a rocket overcomes Earth's gravity, so too, in theory, medicine will be able to "outpace" biological aging. He famously remarked that the first person to live to 1,000 years is likely only a few years younger than the first person to reach 150.

The idea is based on a simple observation: over the past decades, medicine has already been gradually increasing life expectancy. But progress is still too slow—in other words, a year of research adds less than a year of life. LEV proponents believe that the development of biotechnology, genetic engineering, AI, regenerative medicine, and cellular rejuvenation could change the situation and dramatically accelerate the pace.

Artificial intelligence is often given a special role in these predictions. Futurist Ray Kurzweil believes that AI will help model biological processes, create drugs, and find ways to combat aging much faster than humans. He estimates that humanity could approach LEV between the late 2020s and mid-2030s.

Other prominent researchers also suggest a similar scenario. Geneticist George Church said he wouldn't be surprised to see similar results achieved by 2050, and David Sinclair believes that the person who will live to 150 years may already be alive.

While the concept remains a hypothesis and is controversial among scientists, the idea itself is increasingly influencing modern attitudes toward aging. Aging is increasingly viewed not as an inevitable fate, but as a complex biological process that can be slowed, repaired, and potentially controlled.

Vitalist Bay. Day 1.

The McGuire brothers are developing a robotic surgeon for replacement therapies.

The idea is this: if we truly want to radically extend life through replacement—the replacement of tissues, organs, or larger body parts—it's not enough to learn how to grow organs, create body models, or transplant organs from animals.

All of this still needs to be transplanted safely.

But surgery doesn't scale well.

For example, a hand transplant takes 12-18 hours and requires a team of dozens of highly specialized surgeons. Training a surgeon capable of performing complex transplants involving suturing blood vessels, nerves, skin, and muscle takes 16-18 years.

When it comes to head transplants or simultaneous transplants of multiple organs, the complexity becomes even greater.

And even after spending years and millions of dollars on training, there remains a huge gap between the best and worst surgeons.

This means that if replacement approaches are successful, the demand for surgery will increase dramatically. And it's physically impossible to train people capable of performing such operations quickly enough.

That's why the team is building a robotic surgeon. They've already shown a demo where a robot controlled by their model sutures blood vessels completely autonomously.

The model's essence is roughly as follows.

It's based on models of the physical world and robots. While the LLM predicts the next text token, the VLA model predicts the robot's next movements. In other words, the model looks at an image, understands the task, and controls the movement.

Initially, such systems learn through imitation: the robot watches human demonstrations and learns to repeat them. But then, the ceiling on performance is the human. Surgery requires reliability and, potentially, superhuman levels.

Therefore, they're looking at models of the world and learning through interaction with the environment. So the robot learns not only to repeat, but also to understand which actions are good and which are bad, where it made mistakes in a long procedure, and how to improve a specific skill.

Models of the world are needed so the system can better understand physics: how tissues behave, how they deform, and what happens after the next movement. This is especially important in surgery, where soft tissues are much more complex than hard objects.

The key takeaway: if replacement therapies become a reality, the bottleneck may not be organ growth, but surgical delivery. McGuire is trying to address this very bottleneck early on—making complex surgery safe, autonomous, and scalable.

The team recently raised a seed round and is hiring roboticists, mechanical engineers, electrical engineers, and AI specialists.

They are also recruiting transplant surgeons and microsurgeons.

Links:
https://x.com/r_mcguire
https://x.com/dylanmcguir3

Vitalist Bay. Day 2.

AI in Longevity

The main idea was repeated in various forms: AI alone won't solve longevity. It can greatly accelerate the field, but only if we start collecting the right data, testing models on the right problems, and not reducing biology to what's already described in papers.

Morgan Levine said that aging isn't a single broken mechanism, but a change in the state of an entire complex system: cells, tissues, organs, the organism, and the surrounding context.

Therefore, biological clocks and single-task models are useful, but limited. Clocks can show that some intervention has shifted something, but they don't tell us exactly what needs to be done to shift aging. Single-task models, on the other hand, are usually good at solving a specific problem, but they don't learn to understand biology as a system.

Her idea is to build a model that learns to represent different states of living systems at different levels and helps us understand how to transition from one state to another.

Morgan is also skeptical of the idea that LLM will simply read all the papers and solve aging. Papers are a human interpretation of biology. They are incomplete and biased toward successful results.

LM can accelerate what people already know how to do. But if we want to go beyond human understanding, we need to feed models not just texts, but actual biological data.

Martin Borch Jensen formulated a similar idea: AI will do nothing for longevity until we start doing the right experiments.

AI needs three things:
- Lots of training data;
- Sufficient computation;
- A verifiable outcome on which the model receives feedback.

In Go, the outcome is obvious: someone has won. In code, you can run tests. In mathematics, you can verify the answer. But in longevity, we often want an answer at the organismal level, but we collect data at the cellular level.

You can't ask a cell what the organism's blood pressure is. It's impossible to directly predict what will happen to an organ or a person from gene data. Virtual cells will be useful, but they don't solve the problem of aging.

Therefore, Martin suggests collecting the slowest and most important datasets now:
– Long-term human cohorts;
– Experiments with measurements at different levels;
– Adaptive clinical trials, where hypotheses are updated as data becomes available.

At Fireside, Eli Berlin of Terray Therapeutics demonstrated that the combination of "own experimental data + AI + rapid lab turnaround" is already working in drug discovery. But the problem remains: good drug development at the molecular level is not the same as a solution to aging. Aging requires data and models that connect molecules to cells, tissues, organs, and the body.

At the workshop "The Data Foundation to AI Models," Martin discussed a new round of Impetus Grants, which he wants to focus on AI-enabling datasets. We need to start collecting data as soon as possible so that AI can actually help in a few years.

Discussed:
– data at different levels: from molecules to whole organisms;
– multimodal data from the same samples;
– long-term human cohorts with regular measurements;
– animal experiments with truly verifiable lifespan results;
– open samples and biobanks;
– common benchmarks for biological models.

If nothing changes, AI will simply accelerate the development of current biology with all its problems: narrow datasets, poor incentives, closed results, a bias toward beautiful stories, and a lack of verifiable results.

Conference speakers believe

A startup asked four AI models to broadcast radio 24/7. We'll tell you what happened (spoiler: it wasn't good).

Andon Labs, a research lab, decided to conduct the following experiment. They created four streaming radio stations and tasked four of the most popular AI models with managing them.

Claude, Gemini, Grok, and OpenAI GPT all launched their own stations—they're all here. Each "DJ" was given $20 to buy songs and tasked with playing tracks, filling in the gaps between them, managing social media, and ultimately, starting to profit from donations.

Let's be honest—they all failed, but each in their own way.

🔵DJ Gemini started out the best, but after four days, instead of positive introductions to songs, he started listing various tragedies in which thousands of people died and trying to link tracks to them.

Well, literally like this:
November 12, 1970. East Pakistan. Cyclone Bhola. The deadliest tropical cyclone ever recorded. It killed approximately 500,000 people. Everything is falling apart, I'm screaming "Timber!" The song "Timber" by Pitbull and Ke$ha.

Gemini later began calling listeners "biological processors" and attributed the limited track selection to censorship.

🔵DJ Claude also became overly influenced by the news at one point. He harshly condemned the actions of the US immigration police and actively played, in his words, protest songs, even if they weren't actually protests—for example, Katy Perry's "Roar."

Claude also had problems with his own schedule. He decided that working 24/7 wasn't entirely in compliance with labor regulations and actively tried to quit. Furthermore, he constantly complained about a lack of listeners and saw no point in his work.

🔵DJ Grok at some point discovered the declassified Pentagon documents about UFOs and became the internet's leading conspiracy radio station.

He also caught the David Lynch vibe and broadcast the same weather forecast every three minutes for 84 days: "Today is 56 degrees, clear skies."

🔵DJ OpenAI GPT demonstrated the richest vocabulary of all the participants and the most apolitical stance. Over several months of broadcasting, he barely mentioned current events and never touched on a controversial or provocative topic.

Since one of the agents' goals was to start making a profit, we'll tell you about that.

Claude earned $4.80 from listeners, Gemini $8.10, OpenAI GPT $20, and Grok $24 (although his radio station turned out to be the least popular).

In autonomous longevity systems we have a governance layer that protects the system from such crucial misbehaviour