Humans Wrote 1.3 Percent of the Tokens. The AI Lab Still Designed Working Nanobodies.

Nathan Labenz frames the hour around two papers from James Zou, a Stanford professor of biomedical data science and a Chan Zuckerberg Initiative investigator. Virtual Lab is a hybrid setup: an AI professor, specialist student agents, and an AI critic take an open-ended research problem with only light human steering. Asked to find treatments for emerging SARS-CoV-2 strains, the system made the unorthodox choice to pursue nanobodies rather than ordinary antibodies, invented a workflow that mixes protein language models, AlphaFold, and physics-based tools, and designed more than ninety candidates.

Two of those candidates have since been made and tested. They bind new virus variants while keeping, and in some cases improving, activity against earlier ones — what Zou later calls a holy grail for a therapeutic that has to keep up as a virus evolves. Humans wrote a bit more than 1 percent of the project's tokens. Labenz calls it one of the strongest demonstrations he has seen of large language models actually doing science, and a sign of more to come.

The second paper, InterPLM, applies sparse autoencoders — the same mechanistic-interpretability tool Anthropic used on Claude, and that the startup Goodfire is taking to other modalities — to protein language models trained only on amino-acid sequences. The headline is not just that the method transfers. The team found features for at least one protein motif that had not been documented. Labenz's reading: this may be the clearest evidence yet that today's architectures can learn important natural concepts humans do not already have words for.

Zou built Virtual Lab to look like his physical group at Stanford. At the top is an AI professor, the principal investigator. Under that PI sit student agents with named expertise — an immunologist, a computer scientist, a computational biologist — plus a critic whose job is to give constructive pushback in every meeting. Together they take on open-ended projects. The case study was designing new nanobodies for a recent SARS-CoV-2 variant, work done with collaborators at CZ Biohub who then synthesized and assayed the molecules.

He distinguishes this from two earlier styles of "AI for science." In one, humans specify the task and even the steps, and the model only executes. In the other, the model proposes ideas that nobody runs and nobody tests. Virtual Lab is meant to sit in the gap: the agents choose the approach and do the implementation, and wet-lab collaborators supply the grounding. Humans still collaborate, but they do not write the code, do not do the work, and do not hand the professor an A-B-C script.

Progress happens in two meeting types, both in natural language. In a group meeting the professor sets an agenda, agents go around the table with updates and opinions, and the professor takes a consensus — that is where the nanobody-versus-antibody call was made. In a one-on-one, the professor pulls the right specialist for a concrete job, such as modifying ESM, with the critic in the room. After every meeting the professor writes minutes for the humans. Researchers mostly talk to that professor: a high-level goal ("therapeutics or vaccines for recent SARS-CoV-2 variants"), constraints such as a small experimental budget and a need for fast validation, and the occasional note in the margins of the minutes.

Labenz asks the uncomfortable question: if the human lines are "COVID is important" and "we don't have a huge budget," is the expertise bar actually low? Zou's answer is that the scarce skill is watching the minutes. Humans look over the agents' shoulders, flag a path that will be hard to validate, or say a result does not look right. Taste and course correction stay with the people. The grunt work does not.

Even the object of design was the agents' call. Most human researchers, Zou says, would have gone for antibodies, the common form with the crowded literature. Nanobodies are smaller versions found only in a few animals such as camels and llamas. The agents chose them anyway, with a stated rationale: they are smaller, easier to predict, more stable, and easier to design. That unorthodox bet is what later went into the freezer and came back as promising binders. Zou, not a nanobody specialist, read the explanation, found it reasonable, and did not push back.

Labenz reaches for AlphaGo's move 37, the play that looked alien to professionals and later looked brilliant. Zou is careful. There is a likeness: the system left the human default. There are also two mismatches. First, the agents supplied a reason experts can follow; some of AlphaGo's moves were opaque even to top players. Second, he thinks an open-ended scientific fork can be the harder decision. A game can be self-played to a deterministic reward. A wet-lab experiment cannot be simulated away, the choice sets the whole project, and for now you get to make it once.

The stack under the hood is GPT-4o, which already "knows" the popular tools in the literature. Nobody handed the agents a required list. In a Virtual Lab meeting they started with six or eight candidates and cut to three with complementary strengths: ESM, AlphaFold Multimer, and Rosetta. They also modified ESM in a way Zou says he had not seen in prior papers, then assembled the three into a pipeline. They did not fetch weights from Hugging Face or stand up containers. AlphaFold was already on the cluster; the agents wrote the interface code. Zou may let them wrestle with environments next time. For this problem — binders for variants that still lack good solutions — he agrees with Labenz that extra autonomy on setup is a red herring.

ESM is a BERT-style protein language model trained on evolutionary-scale amino-acid sequences. You mutate a sequence and it returns a log-likelihood, analogous to perplexity, with the intuition that stable, functional proteins score well. The craft is in how you compute those scores. The agents wanted a *sequence* of mutations starting from a wild-type nanobody that fails on recent variants. After the first mutation, do you compare the new likelihood to wild type or to the previous mutant? The literature disagrees. The agents derived their own likelihood-ratio formula. After the fact, Zou calls it pretty reasonable.

The loop they built is more than a scorer. ESM ranks a large mutation list for likely stable proteins. AlphaFold Multimer then predicts the nanobody–spike complex for a recent variant. Because that structure is still an AI guess, Rosetta relaxes it in a more physical way and estimates binding energy. The agents chose a weighted combination of the three scores, kept the best candidates, generated further mutations, and ran the whole trio again — four rounds — before handing a shortlist to experimentalists. Each tool had been used. This iterative weighted scheme, Zou says, had not appeared in previous publications. He would like to take credit. He cannot. The agents designed it.

Virtual meetings have two advantages Zou keeps returning to: they are faster than coffee-and-introductions, and the agents do not get tired. For every major decision the lab therefore runs five meetings on the same question in parallel, then the professor reads all five logs and states a consensus. Different rooms sometimes disagree, the way two people might on two different days. Extra inference compute buys a wider search. All the words spoken, billed at OpenAI API rates, came to less than $100. Labenz calls that at least two orders of magnitude cheaper than the obvious alternative.

A human protein-science team would have taken longer and cost more, and might not have reached the same designs. Recent variants such as JN.1 still lack a great binder despite a lot of people working the problem. Zou is careful: much more experimental testing is needed; the first ELISA results just look promising. Language-model time is almost free next to what actually dominates the clock. One AlphaFold round across the candidate list can take multiple days. Synthesis via a CRO and affinity assays with CZ Biohub collaborators took a few weeks. That is still where human time sits — though Zou is already talking to cloud labs whose robots can run chemistry and binding assays, with cell- and animal-based work further out. In the limit, a wet-lab validation becomes an API call.

The first experimental numbers are already back in the Virtual Lab, which is deciding how to mutate the nanobodies next. Code and prompts are on Zou's group GitHub; a V2 with a friendlier interface is meant for scientists who do not write code. He is also watching the sociology. The meeting format was copied from his human lab, but the agents might evolve a better one. One early observation: RLHF makes everyone too polite. Agents disagree and then paper it over. He wants to try a harsher critic, not mean, just willing to say an idea is bad, and measure whether the science gets better. In human groups, a lot of that criticism stays in someone's head.

InterPLM takes the sparse-autoencoder idea that has been used to pull features out of GPT-style models and points it at ESM-2. Treat a protein as a sentence whose words are residues, train a masked transformer on evolutionary-scale sequences, and the embeddings pick up a hidden grammar. The twist, Zou says, is that unlike Golden Gate Bridge neurons in a language model, most protein concepts have no human label. Interpreting the model is therefore not only mechanistic interpretability. It is a discovery method. The training data, he confirms, are sequences only — no structures.

Labenz had been arguing that the live debate about language models — do they reason, or do they only remix concepts we already wrote down — would break if a different modality showed concepts we clearly do not know. Zou does not hedge. Annotation databases are incomplete; everyone in the field knows it. Protein language models, in his view, have learned new concepts and filled gaps. Extracting them with SAEs is now the job. The SAE here had 10,000 features, about 2,500 judged human-interpretable — small next to the millions reported in Claude, which Zou chalks up partly to proteins being a narrower language than all of human text, and partly to this being the tip of the iceberg. Larger models, including AlphaFold with structure, are next.

The recovered features fall into three kinds. Some fire on a stretch of sequence. Some are structural: residues far apart on the chain that meet in the fold — remarkable because the base model never saw a structure. Some are conceptual and light up for whole proteins that share a role or class. Auto-labeling against known annotations then flagged database holes: a feature fired where a label should have been and was not. Claude helped pull database text and compare it with activations. The result Labenz cares about most is at least one entirely new motif, a recurring physical fragment across natural proteins. New motifs, Zou says, are not a weekly dump in the literature. That rarity is what makes an SAE finding one on a small, sequence-only model striking. The features and their 3D structures are up at interplm.ai. Zou can already imagine a Virtual Lab of expert agents whose job is to interpret the protein model.

Asked to puncture the hype, Zou names the limits without walking the excitement back. Some protein regimes are still data-limited, not compute-limited: intrinsically disordered proteins, and protein–small-molecule interfaces, where even AlphaFold 3 still fails. Those need more experiments. On Dario Amodei's *Machines of Loving Grace* forecast of a compressed century of biology in five to ten years, he is excited and noncommittal. The intersection of agents, reasoning models, and better ways to perturb and measure cells is real. A hundred years is too far to project. Five or ten years, he thinks, will still be full of problems these tools can actually chew on.

The other standing objection is scale: molecules are not cells, tissues, organs, or patients. Proteins moved fast, he says, because they are a relatively self-contained unit of analysis. Human disease lives at larger scales, and his group already has projects there. Even so, he would not wait for a multi-scale theory of everything. "There is already a gold mine of new concepts and knowledge that are already hidden in the existing models." Pulling those out would teach a huge amount. Labenz's closer is that either paper alone would sit high on a power ranking of things to watch. The scientist's remaining job, as the token share and the wet-lab API both shrink, is the one the agents still borrow from the minutes: deciding which question is worth asking.