There is an old joke in biology and medicine: when you read an exciting headline, mentally add "in mice" to the end of it. Nathan Labenz, host of The Cognitive Revolution, brought it up with Alex Rives, head of science at Biohub, in an episode published October 20, 2026. Rives used to lead ESM, a protein language model project, and his answer to the joke drew on what that work taught him.
A gap that should be smaller
Labenz said the gap confused him. Humans and mice share a great deal of DNA, and much of the core machinery inside their cells is conserved, meaning it has stayed largely unchanged through evolution. Given that, he said, treatments developed for "mouse cancer or mouse Alzheimer's" ought to work in humans more often than they seem to.
He offered two ways to think about closing that gap. One is to build linked datasets that connect results in model organisms, the lab species researchers study in place of people, to results in humans. The other is less hopeful about animals. Perhaps "there's like no amount of zebrafish data" that will add up to the insight researchers want into humans, and only a great deal of human data will get them there.
What lab animals got right, and two ways they mislead
Rives began by defending model organisms. Most of the fundamental advances in the molecular understanding of disease, and of the biological pathways behind it, came from experiments in model systems, he said. Even yeast, which he called "the simplest eukaryote" (eukaryotes are the broad group of organisms whose cells have a nucleus, humans included), has revealed mechanisms that matter for human disease.
He agreed with Labenz, though, that they are "not perfect proxies by any means," and said leaning on them too heavily can go wrong in two opposite directions. A researcher can think they have "a cure to cancer, but it only works in mice." Or they can miss a mechanism that matters in humans because the mouse has nothing that stands in for it. In the first case the animal points to something that fails in people. In the second, a real human mechanism never shows up in the animal work at all.
There is "no simple solution to this problem," Rives said. Biology keeps pushing for better experimental stand-ins for human tissue. The latest, he said, are organoids: small tissue-like structures grown in lab culture that reproduce more and more of the complexity of real tissues.
The lesson from protein models
Rives said AI could help in another way, and he pointed to his own field. Protein models learn from sequences of amino acids, the chemical units that make up proteins. Two of the best known are ESM, a language model for protein sequences, and AlphaFold, Google DeepMind's system for predicting a protein's three-dimensional shape. Neither "came from training a model just on human proteins," Rives said. "They really came from training a model across, you know, all of evolution."
In his telling, that breadth was required, and calling it helpful undersells it. "It's not even that it's relevant. It's essential," he said. A model trained only on human proteins, he argued, would not be an accurate predictive model.
The 2021 papers describing the two systems back up the general point that breadth helps. In the ESM paper, Rives and colleagues trained models on collections of up to 250 million protein sequences. Each model learned by guessing amino acids that had been hidden from it, using the rest of the sequence as context. When the team compared training sets sampled in different ways, broader and more diverse sampling produced models that generalized better. They also tested the models on entire protein families and structural classes left out of training, to see whether what the models learned carried over to distant relatives.
The AlphaFold paper shows the same dependence from another angle. AlphaFold uses alignments of evolutionarily related proteins, drawn from large sequence collections that include metagenomic data (genetic material read in bulk from environmental samples). These alignments show which parts of a protein stay the same across species and which change together. AlphaFold's accuracy generally drops when a protein has few known relatives to align against.
Sal Candido of Biohub made a related point on a separate panel released by the Latent Space podcast on October 10, 2026. He said that adding messy metagenomic sequences to training still improves protein models, even though much of that data does not represent a real, whole protein.
From proteins to cells
Rives's argument carries the protein lesson up to cells and whole animals. Much of biology is conserved "across all the different domains of life," he said, and the closer you get to humans, the more is shared. Among vertebrates, he pointed to transcription factors, proteins that switch genes on and off, and to enhancers and promoters, stretches of DNA that control when and how strongly a gene is used. Rives said these parts, and the way they interact, are preserved across vertebrates: "that language is preserved."
What changes from species to species, he said, is how those parts are wired together. "The programs are completely rewired, but they're kind of built with the same, same building blocks." On that reading, a mouse runs its own program, built largely from parts a human also uses. That, Rives said, is why there is so much to learn from many different systems at once.
He separated two strategies. One is the search for a perfect experimental proxy for human biology, which he said "really just isn't going to be possible." The other is a predictive model that captures, integrates and learns from everything those many systems reveal. Rives said there is "a power" in trying to build such a model, one that can "use all of that information to make far better predictions for human biology."