From Static Structures to Moving Molecules: Where AI for Protein Design Stands After AlphaFold 3

Schreiber opens with David Baker, who shared a Nobel Prize, saying she is a big fan of his lab's work and hears he is also a very good advisor. She finds it annoying that popular videos gave AlphaFold most of the fanfare while skipping the history of Baker's protein engineering, which predates AlphaFold and was initially physics-based through Rosetta.

In her telling, things changed when the lab adopted AI, around the time AlphaFold arrived. The group built RoseTTAFold 2, a structure predictor much like AlphaFold 2, then brought in generative modeling people and fine-tuned it into the diffusion model RFdiffusion. It is more than two years old but still very useful, and the lab also built a sequence design model so designed structures could be turned into sequences that fold into them.

She notes that five laureates across physics and chemistry won in large part because of AI techniques, which says a lot about AI's impact. There is plenty of hype, she says, but also real signal, and in biochemistry in particular she thinks we are still at the very beginning of the curve.

Schreiber highlights AlphaFold 3 and ESM3 as the two big models released since the previous episode. AlphaFold 3 predicts complexes involving not only proteins but also RNA, DNA, small molecules and even ions. She hasn't tested how accurate the ion predictions are, but if they are good, that is useful for several things she cares about.

The model isn't open source, and she says open reproductions so far haven't satisfied her, though a couple of groups are working on it and she hopes for results by the end of the year or early next. An API exists, but its licensing restrictions mean it can't really be used in a drug design process or where you want IP, such as patenting or selling a molecule. Academic, fully public research would be fine.

Why does it matter? Complex prediction lets researchers build out interaction networks, which are key to understanding disease mechanisms and deciding whether you need a binder, an inhibitor or an allosteric modulator. Labenz floats the idea of brute-forcing which molecules in a cell interact, noting that perhaps only 5 to 10 percent of interactions in a cell are mapped.

Schreiber describes the standard binder workflow from the Baker lab. RFdiffusion designs a binder backbone, ProteinMPNN or LigandMPNN designs sequences for it, and a structure predictor validates that the sequences fold as intended. At scale people may generate tens or hundreds of thousands of candidates, so they screen them using the interface PAE score from a single-sequence AlphaFold 2 run.

Because high-throughput screening needs speed, the lab trained a lighter, smaller model by stripping layers from a structure predictor. It turned out that it did better at predicting protein-protein interactions even though it was slightly worse at structure. Schreiber still thinks structure awareness is essential for any good interaction predictor and that geometry should transfer to other targets, such as protein-DNA binders.

The same approach was applied to the human proteome, predicting all pairwise interactions and finding many new, high-quality ones with a good area under the precision-recall curve. Schreiber explains that this metric captures how much you recover versus how many false positives you bring in. She suggests a flow-matching version of AlphaFold could make this screening even faster.

The slow step in interaction screening isn't the model but generating multiple sequence alignments (MSAs). Schreiber is testing ideas to avoid or improve them, especially for "twilight zone" proteins that have very low sequence identity to others, making deep alignments hard to build.

One method uses a protein language model such as ESM2 to find homologues by comparing embeddings with cosine similarity, instead of comparing sequences. Since structure is more conserved than sequence, this can pick up relationships that sequence similarity misses. She is also trying to perform the alignment itself in embedding space, which sometimes yields odd-looking alignments.

Labenz summarizes MSAs as a way to identify what is conserved and therefore likely critical. Schreiber says the ideal setup would be an indexed vector database, but she isn't yet sure it will be much faster, since finding homologues depends on how well you query the database. The speed question remains to be determined.

Asked what AlphaFold 3 still can't do, Schreiber says a lot. Static prediction is fine for tasks like binder design or motif scaffolding, but a new enzyme catalyzing a new reaction is trickier. Catalysis is inherently dynamic, so you must consider the catalytic residues, the binding pocket's side chains, and dynamics up to about 20 angstroms away from the catalytic sites.

She lays out a pipeline. EnzymeFlow, a flow-matching model, generates a catalytic pocket; RFdiffusion scaffolds it into a full protein; LigandMPNN designs the sequence since it understands the substrate; and another Baker lab model docks the substrate and gives conformational ensembles to rank designs. Structure validation follows, and the loop can be iterated to boost turnover.

She says what is needed beyond AlphaFold 3 is the modeling of conformational ensembles. That sets up her enthusiasm for MDGen, which she says could serve this role, possibly after some fine-tuning.

MDGen is a generative flow-matching model that Schreiber likens to a Sora for molecules. Trained on the smallish Atlas dynamics dataset, it outperformed AlphaFlow. Beyond generating trajectories from a static structure, it can upsample coarse simulations, interpolate between two metastable states (useful for binding free energy), and in-paint parts of a trajectory.

The in-painting excites her most: masking part of the molecule and having it fill in a new piece whose dynamics match the rest. It was trained only on single proteins, but trained on complexes it could enable new binder designs with favorable dynamics. She says the capability is brand new and the community hasn't had time to digest it.

If such a model generates trajectories a thousand times faster than molecular dynamics, a month-long simulation becomes a few hours, and tasks that weren't high-throughput suddenly are. She thinks the small loss in accuracy would be outweighed by testing a thousand times more hypotheses. She doesn't claim it will replace molecular dynamics, but expects it to take a chunk out of it.

Labenz recaps that the frontier is dynamics versus statics, picturing enzymes as assembly lines where the machinery must have the right shape and also move in the right order. Reactions that won't happen at room temperature, or would kill a cell if heated, need an enzyme that physically guides the molecules. He likens the process to his "do the hard part first" rule for AI app development: find the core active-site geometry, then scaffold a larger protein around it.

He connects this to an earlier guest's work on machine-learned simulations of salt water, where a model stood in for slow numerical methods, allowing longer runs that revealed crystallization. Asked what in-painting is good for, Schreiber offers two uses: binder design, once trained on complexes, and enzyme design by in-painting around a catalytic pocket.

In both cases dynamics matter. A designer might want a binder whose interaction lasts for a certain time or has a certain probability of dissociating, and for enzymes the dynamics away from the active site must be compatible with those around it.

Asked about the ecosystem, Schreiber divides applications into industrial and medical. Industrial work moves faster because it doesn't face the safety process needed for human use, although safety can still matter. She says the heavy-duty software and workflows are mostly closed source, inside big drug companies collaborating with smaller AI labs.

One exception is Nvidia, which is working on open endpoints (NIMs) designed to run whole workflows of models rather than a single one. Otherwise open source still means picking models, setting up environments, downloading weights and sometimes training by hand. Labenz observes that nobody outside a few edge companies has tied it all together; there is no Zapier for this.

On whether the next AlphaFold would absorb all this, she says for now dynamics models still need a static structure as input, so it is a two-step process. The real next generation, she says, would be a single model doing structure prediction and dynamics together.

Schreiber hesitates to give a single success rate. For motif scaffolding, such as joining two binding domains into a bispecific protein, success is relatively high, perhaps one hit in 50 designs sent to the wet lab after ranking. A hit means the design shows good binding to both targets.

Enzymes are harder because of the active-site arrangement, side-chain conformations and dynamics that must be modeled. She guesses something like one in a thousand, though it depends on the target and models, and is still much higher than with prior methods. For comparison, she cites AlphaProteo, where some targets had success rates in the 80-something percent range while others were around 9 percent.

On how enzyme tasks are defined, Schreiber says the catalytic geometry is actually very precise and fragile, so the task is really a geometry. EnzymeFlow, for example, is conditioned on a reaction and an enzyme commission number, then generates the catalytic pocket with the sites arranged correctly.

Asked how ambitious design can get, for example creating a CRISPR analog, Schreiber says pretty ambitious. At Rosetta conference this year she saw stereoselective enzymes, mechanical degraders that pulled apart the needle complex of bacteria so that they couldn't infect cells, and multi-step enzymes with four steps of catalysis. Her advice is not to expect a miracle but to expect surprising things.

Many of these problems translate into mechanical engineering problems for people who are creative, persistent and mechanically minded. Practitioners build intuition by looking at structures in viewers like Mol*. Schreiber says that often you don't need full molecular dynamics: with an open and a closed conformation, designing a binder to lock a protein open is a purely structural problem.

Schreiber covers PepFlow and a companion antibody model, both RFdiffusion-like. Small disordered peptides behave much like antibody CDR loops, and she wonders why nobody trained on both. PepFlow uses several kinds of flow matching, including SO(3), toric and Euclidean ones, and co-designs structure and sequence in one model, for peptides under about 30 residues. She cautions that without wet lab validation she doesn't fully trust these methods yet.

She explains flow matching as a generalization of which diffusion is a special case, with benefits in training stability, data efficiency and notably faster inference. Peptides, she adds, can sometimes achieve better binding affinity than antibodies, and small ones may be manufactured by processes similar to those for small molecules.

On ESM3, she is lukewarm about its finite function vocabulary of a few hundred tokens and would prefer an open-vocabulary, text-prompted design like ProteinDT. She also finds the phrase "simulating 500 million years of evolution" a bit misleading, since it isn't like in silico directed evolution. She does note a new open fine-tune of ESM3 that generates conformational ensembles.

For the next six months Schreiber expects improvements to dynamics models, such as MDGen trained on bigger and better data, and models that sample Boltzmann distributions. Beyond that, she expects multiple models chained into cohesive workflows run by agents that tune hyperparameters, assess outputs and propose hypotheses. If such a process scales, you can churn out molecules, and design becomes less guess-and-test and more deliberate.

She has seen language models used for literature search, output assessment and iteration, sometimes with small fine-tuned 8-billion-parameter models and a bigger one orchestrating. She would be surprised if someone didn't have a real banger by early next year. Labenz notes that AI-scientist-style systems still write mediocre papers; she replies that her measure of success is a well-designed molecule in the wet lab, and hypothesis and target generation interest her more than writing publications.

Her closing message is that the field moves fast, with things very different every six months, and needs more people because agents will have to scale the work. She encourages anyone interested in biochemistry and AI to get in soon, calling it fast-paced but beautiful.