A Protein Backbone in a Minute: The Real Progress, Bottlenecks, and Risks of AI Biology

In its 2023 AI executive order, the White House drew a line that most language-model watchers barely noticed. Reporting kicks in at 10^26 FLOPs for language models, but at 10^23 FLOPs for biological ones—a thousand-fold difference. Host Nathan Labenz flags that gap at the top of this Cognitive Revolution episode, along with a newer rule that DNA synthesis companies screen orders against known dangerous sequences. After walking through how small today's protein models already are, he says both moves look "extremely smart right now."

The guest is Amelie Schreiber, a computational biochemist and AI researcher whose graduate training started in pure mathematics, moved into applied math and data analysis, and then into deep learning. She spends her days on proteins, small-molecule drugs, and nucleic acids, with some side work on environmental and materials problems. Other than AGI, she says biochemistry may be "the most important problem we can be working on" because the health impact is so direct. The invitation itself came from a tweet: she listed her top ten AI tools for biology, and Labenz realized he did not recognize a single name on the list.

Labenz calls this the most future shock of any episode he has recorded, and he rewound the tape more than once while preparing. Because the picture sat so far outside his prior map, he checked his high-level reading with people deep in the field; one of his smartest friends said, simply, that is exactly what is going to happen. One number from later in the conversation makes the shock concrete: on a laptop, RFdiffusion can generate a three-dimensional protein backbone that nature never made in about a minute. Traditional structure determination, by contrast, still meant making enough protein, coaxing it into a crystal that does not occur in nature, and interpreting X-ray scatter—an approximation of the shape the chain actually takes in a cell.

Labenz sketches biology as a nested systems problem. DNA is code; RNA is both messenger and machine; proteins fold into odd three-dimensional machines; small molecules carry signals. Those pieces bind, activate, and shut one another off in a vast interaction network, and his working estimate is that humanity understands perhaps five to ten percent of it. Schreiber says he has "hit the nail on the head": even deciding whether two proteins interact is hard, let alone modeling how they fit, how tightly, and how transiently.

Treating disease, in this picture, is two steps. First, find which pathway in that poorly mapped network has gone wrong. Second, design something that intervenes—without setting off knock-on effects no one can predict. Schreiber's teaching example is cancer's PD-1 / PD-L1 pair: a protein on the tumor binds a protein on the immune cell and turns the immune response off. The design move is to build a binder that occupies one side of that interface, blocking the stand-down signal so the immune system can attack again. The same logic, she notes, extends to small-molecule drugs aimed at a particular pocket and to DNA and RNA designs with specified chemical properties.

Before these models, the work ran through wet labs and molecular dynamics. Directed evolution mutates a protein in animals and waits; once you go beyond single-point changes, the combinatorial explosion becomes unwieldy. The computational tradition was molecular dynamics: integrate Newtonian physics, maybe with quantum extras, for hours or days on GPUs, and still miss states if the trajectory is too short. The ideal ground truth, Schreiber says, is the Boltzmann distribution—an energy landscape whose valleys are metastable conformations and whose peaks are fleeting transitions. Labenz's Slinky on the floor is the low-energy coil; stretch it and let go, and the rebound is the transition, while water molecules batter the chain until it occasionally flops from one valley to another.

AlphaFold 2, Schreiber reminds listeners, takes a twenty-letter amino-acid sequence and returns one static low-energy structure plus a confidence map. That output matches the older experimental diet: X-ray crystallography of protein crystals that do not exist in nature, so the "true" cellular shape was always an approximation. Proteins jiggle; temperature, partners, and ligands deform them. A single snapshot cannot tell you how often two residues touch, or which other shapes the chain occupies.

Distributional Graphormer, released around mid-2023, generalizes that picture. It is a diffusion model—kin to DALL-E, working on proteins instead of pixels—and it emits an ensemble of conformations plus transition paths between metastable states. MIT's AlphaFlow retrains an AlphaFold 2 architecture with flow matching, a generalization of diffusion; run it for 10,000 conformations, cluster them, and you can see which residue-residue contacts are transient. AlphaFold Multimer, meanwhile, scores the interface between two chains, from puzzle pieces that click to pieces that barely adhere. Schreiber's research-project pitch is to do the AlphaFlow trick on Multimer, so the field can estimate how fleeting a protein-protein contact really is—work she says has not been done yet.

The models still fail in instructive ways. KaiB, the literature's favorite fold-switching protein, spends about 10 percent of its time in one conformation and 90 percent in another, gated by circadian rhythm, yet AlphaFold 2 and AlphaFlow both cling to the rarer, slightly higher-energy fold and miss the ground state. The popular hack is MSA subsampling: different clusters of evolutionary relatives emphasize different shapes, so swapping clusters sometimes yields the other fold, but Schreiber says it is not robust. Multiple sequence alignments themselves are evolutionary ledgers—regions that barely change across deep time are usually load-bearing, and mutating them often wrecks function or stability. Labenz reaches for the wartime airplane diagram: the bullet holes on returning planes mark survivable damage, while the blank patches are where hits never came home, and conserved sequence is the biological version of those blanks.

ESMFold makes the physics-optional point even sharper: it is a BERT-style protein language model trained to fill in masked amino acids, then topped with AlphaFold 2's Evoformer folding head, and it never sees forces or potentials. It underperforms AlphaFold 2, which also uses templates and MSAs, yet it still folds most proteins in minutes. Attention maps in these models recapitulate residue contact maps, a finding dissected in the roughly three-year-old paper *Bertology Meets Biology*, which mines attention for active sites, binding sites, and motifs. Labenz's hypothesis, which Schreiber largely accepts, is compression: networks learn higher-order concepts such as recurring motifs the way language models trained only on next-token prediction still grow internal features for justice and fairness. Some teams even train with physics constraints and no experimental data, and still sample ensembles orders of magnitude faster than molecular dynamics.

Schreiber's sharpest warning is about evaluation, not architecture. Most protein-interaction models, she says, do not generalize, because people split train and test without regard to sequence or structural similarity; near-duplicate chains leak across the cut, scores look excellent, and the overfitting is invisible. Labenz maps it onto language-model practice: if MMLU sat in the training data, you can no longer trust MMLU. DiffDock-L, a larger docking model, split data by the domains involved in each interaction and trained with confidence bootstrapping, a reinforcement-learning-like scheme; the result, she says, generalizes much better out of distribution. Anyone crossing over from NLP should treat a random 80/20 split on biological sequences as a data leak until proven otherwise.

The generative stack is shifting from Transformers toward diffusion and flow matching, and it is no longer protein-only. RoseTTAFold All-Atom predicts complexes of proteins with small molecules, nucleic acids, and metals, and returns a score for whether an interaction is likely and how strong. Models such as ProteinDT and MoleculeSTM go further: they use CLIP-style contrastive learning on captions paired with molecules, so a natural-language prompt describing chemical properties can generate a protein or a small molecule that matches. Schreiber's reaction is unvarnished. Having a model that generates molecules from plain descriptions is, in her words, "crazy" and "amazing."

Her favorite workhorse is RFdiffusion, which takes a RoseTTAFold backbone and trains it as a denoiser of 3D protein structures. It does not see sequence, so LigandMPNN—the all-atom successor to ProteinMPNN—has to invent an amino-acid string that will actually fold into the proposed shape. Unconditional generation needs only a length; the model also designs high-affinity, high-specificity binders, and it does motif scaffolding, extracting a useful fragment and building a new protein that holds that fragment in place. String two binding motifs into one scaffold and you get an adaptor that grabs protein A and protein B at once, which is how she talks about rewriting interaction networks rather than merely blocking them. On the PD-1 / PD-L1 example, the loop is generate, redesign with partial diffusion and biased sequences, then check AlphaFold Multimer's ipTM score; David Baker's lab has used the same family's symmetric oligomers as biosensors and as drug-delivery cages.

Speed is the headline: Schreiber can generate a fairly large backbone in about a minute on a laptop, and sequence design per backbone is under a couple of minutes, so an afternoon can yield hundreds or thousands of candidates. When those designs are synthesized, thermal stability, specificity, and binding affinity are very often high, and partial diffusion plus sequence redesign usually lifts weak starting points dramatically. RFdiffusion binders tend to be specific enough that off-target effects stay limited, which is why she argues that injecting animals and waiting out directed evolution is starting to look unnecessary. The slowest step in the loop is not generation but AlphaFold or AlphaFold Multimer validation, and the models themselves are small—on the order of a few hundred million parameters. One H100 is plenty, and a Colab notebook runs RFdiffusion on a free T4.

If binders can be that specific, Labenz asks, why are we not already curing a lot of diseases? Schreiber's answer is that diseases are messy: multiple targets, multiple interaction types, whole networks to modulate, and the hard judgment is which node to touch. Compute is not the limiter—one good GPU, even a single H100, runs the design stack. She compares adoption to YouTube views on RFdiffusion tutorials: not many, because the tools are new, they need a conda environment and some code, and there are not many experts yet. Labenz's analogy is military. The missiles already fly straight; picking the target is now the dangerous decision.

The other throttle is the clinic. Current approval was not built for high-throughput design, so a candidate still crawls through trials on a timescale of years. Labenz wonders whether trials will go from a roll of the dice to something you are disappointed to see fail; Schreiber thinks speed-up is coming, and that it will jump again once dynamics models are widely used. Platforms such as 310 AI are trying to erase the coding barrier with a chat Copilot that function-calls out to ProteinMPNN, LigandMPNN, DiffDock, and, soon, likely RFdiffusion. A user can say "increase this protein's affinity for that target" or "dock this small molecule" and get a result without setting up environments. Schreiber thinks that kind of interface will matter as much as the models, because GPT-level tools did not change productivity until people actually used them.

In parallel, sequence foundation models are climbing. Evo is a DNA language model trained on about 300 billion nucleotides of bacterial and phage genomes—a narrow slice Schreiber thinks was chosen partly for safety, and one she found disappointing because it leaves so much biology out. Even so, it can score gene essentiality from downstream perplexity, and she says predicting whether a mutation helps or hurts is already "not a hard thing": ESM had the skill years ago, Evo reported state of the art, and DeepMind's AlphaMissense scored every single-point change in human proteins. Three hundred billion nucleotides is modest next to the 15 trillion tokens Meta used for Llama 3, so the ceiling is still high. Labenz expects multimodal models that predict how a cell's state evolves, then support counterfactual perturbations—change this, watch the trajectory.

His two-year sketch is two overlapping curves. First, a backlog of targets found the hard way can now be hit with the new design stack—the low-hanging fruit. Then, as those are worked through, later generations of models like Evo start identifying new targets by perturbing a learned picture of the whole causal graph. None of this assumes AGI; these are still narrow tools. Agents are arriving anyway: FutureHouse is building a system, on GPT-4 or an in-house model, that searches literature, proposes hypotheses and targets, and is pairing that with a self-driving lab. OpenAI's partnership with Moderna is another sign of the same marriage, and Schreiber expects the research agent and the molecule-surgery tools to converge "in the next year."

Schreiber is excited and nervous in the same breath. She expects health problems to recede and lifespan to stretch, and she also expects someone to try to target a genetic group. Labenz's update is that the familiar debate over whether GPT-4 is more helpful than Google for a bioweapon already feels quaint: these systems generate new matter, not just retrieve recipes. He keeps a long zoom in mind. Earth's first mass extinction on the Wikipedia list is atmospheric oxygenation—a metabolism whose waste product killed almost everything else—so a brute-force search through biological space has no obvious conceptual ceiling.

Schreiber tries to put the panic in the right place. Computationally designing a molecule is one step; synthesis, wet-lab work, and a delivery mechanism are non-trivial and tightly watched. A random person with Llama 3 is "very unlikely" to run that full gauntlet, even with a capable agent. The actors to watch are research companies that already have wet labs and are building agents to run them. That is a now problem, not a future one: such agents already exist, most people cannot use them, and those firms need safety teams, red-teaming, protection against espionage, and hiring that screens for more than skill.

On offense versus defense she analogizes to cybersecurity, where attackers usually hold the edge. Strong enough agents could flip that, which she treats as an argument for keeping language-model and agent progress fast: the best agent may be the best defense. A University of California team has published work on universal vaccines that cover a wide range of mutants, with a method they describe as highly transferable—helpful if a designed pathogen ever needs an answer as fast as the COVID record of a few days from sequence to vaccine design. Labenz adds the property that makes biology unlike nuclear missiles: warheads do not spontaneously leave their silos and self-replicate, and the list of lab leaks is long. A bioweapons arms race, he says, is a trap that has to be avoided in the first place.

She ends on a claim about time, not technology. "The time that we have with the people we care about is in my opinion the most valuable thing that we have." If people can live healthy lives with those they love for much longer, she hopes the usual finite-resource frame starts to loosen, and that some cultural threats fade because they were never only technical. Labenz calls it a beautiful place to stop. For readers who want to do something this week, RFdiffusion's Colab notebook plus LigandMPNN and an ipTM check is a full beginner pipeline; 310 AI and FutureHouse are the interfaces to watch; and AlphaFold 3, which landed shortly after taping, folds proteins, nucleic acids, small molecules, and ions in one model.