AlphaFold3 and the State of Protein Structure Prediction

For a decade the hard problem in structural biology had a clean shape: given an amino-acid sequence, predict the folded structure. AlphaFold2 effectively closed that problem in 2021 for single proteins, and the field moved on. The problem that actually matters to anyone designing a drug is harder and messier. Proteins do not act alone. They bind small molecules, sit on DNA and RNA, coordinate metal ions, and assemble into complexes. A folding model that stops at the isolated chain answers the wrong question.

AlphaFold3, announced by Google DeepMind and Isomorphic Labs on 8 May 2024, is an attempt to answer the right one. It predicts the joint structure of a complex — proteins together with nucleic acids, ligands, ions, and modified residues — in a single model. That shift, from folding one chain to predicting how molecules sit against each other, is the part worth understanding, because it is where structure prediction stops being an academic curiosity and starts touching the bench.

What actually changed under the hood#

AlphaFold2 was built around two ideas: an Evoformer that reasoned over a multiple-sequence alignment to extract evolutionary signal, and a structure module that placed residues using protein-specific geometry — frames, torsion angles, the whole vocabulary of a backbone. That geometry is exactly what made it a protein model. You cannot describe a zinc ion or a drug-like small molecule in the language of backbone torsions.

AlphaFold3 throws out the protein-specific output stage. The architecture replaces the Evoformer with a simpler Pairformer and hands the final structure generation to a diffusion module that operates directly on raw atom coordinates. It begins with a cloud of atoms at random positions and iteratively denoises them into a plausible structure, the same generative recipe that powers image diffusion, applied to 3D coordinates instead of pixels.

That choice is what unlocks the generality. A diffusion model over atoms does not care whether an atom belongs to a protein, a base pair, or a ligand — it predicts a position either way. The model no longer needs a bespoke representation for every chemical entity. It needs one representation, atoms in space, and a denoiser that has learned what real molecular arrangements look like. The cost is that diffusion can produce confident-looking structures in regions with no real constraints, a failure mode the authors handle with cross-distillation and ranking rather than pretending it away.

Where it helps, measured honestly#

The headline result is protein-ligand prediction, because that is the interaction drug discovery lives or dies on. On the PoseBusters benchmark — a set of recent protein-ligand structures chosen to avoid training-set leakage — AlphaFold3 placed the ligand within 2 angstrom RMSD of the true pose in roughly 76% of cases, without being given any pocket information, and did so more accurately than conventional docking tools that were handed the binding site. Predicting where a small molecule sits on a protein, from sequence and chemistry alone, is a different capability from anything the docking field shipped before.

The model reports similar gains on protein-nucleic-acid interactions, beating specialized RNA and DNA structure tools, and on antibody-antigen complexes, which are notoriously hard. None of this means the predictions are correct often enough to skip the experiment. It means the starting hypothesis is good enough to prioritize which experiments to run, which is the entire game when wet-lab cycles cost weeks and money.

The honest caveats#

Two limits deserve to be stated plainly, because the marketing tends to skip them. First, AlphaFold3 predicts static structures. A drug target is a moving object, and the conformation that matters for binding may not be the one the model returns. Second, confidence is uneven. The model emits per-atom confidence scores, and for ligands and disordered regions those scores are doing real work — a low-confidence pose is a coin flip, not a result. Teams that treat every prediction as ground truth get burned; teams that treat the confidence metrics as a triage filter get value. The discipline is the same one we apply to any AI implementation: the model proposes, the validation disposes.

What it means for drug discovery, concretely#

The practical effect is on the front of the pipeline, not the end. A medicinal-chemistry program spends an enormous amount of early effort guessing which targets are tractable, which binding modes are plausible, and which of a thousand candidate molecules deserve synthesis. A model that predicts complex structures fast and cheaply changes the cost of asking those questions. You can screen binding hypotheses computationally before committing a single assay.

This is the same architectural logic we see in every domain where a Data Platform sits in front of an expensive physical process. In a Hospital Management System you do not run a confirmatory test on every patient; you triage with cheap signals and reserve the expensive workup for the cases that warrant it. Structure prediction is triage for the wet lab. It does not remove the lab. It decides where the lab points.

What it does not do is hand you a drug. A predicted complex is a hypothesis about geometry, not a statement about binding affinity, selectivity, toxicity, or whether the molecule survives a liver. The gap between “this ligand probably binds here” and “this is a medicine” is most of the work of drug discovery, and no structure model closes it. The win is that the earliest, cheapest, most uncertain stage got faster and more informed.

Where it sits next to docking and simulation#

It helps to place AlphaFold3 against the tools it partly displaces. Classical docking searches for the lowest-energy pose of a known ligand in a known pocket, using a scoring function hand-built over decades. It is fast, interpretable, and only as good as that scoring function — which is to say, often wrong on novel chemistry. Molecular dynamics simulates the physics of atoms over time and can capture motion that AlphaFold3 cannot, but it is expensive enough that you run it on a handful of systems, not a library of thousands. AlphaFold3 occupies a different niche: it predicts a plausible bound pose directly, with no pocket specified and no force field tuned, across a far wider range of molecular types than any one of those tools handles. It does not replace docking or dynamics. It reorders when you reach for each. The cheap, broad, learned predictor goes first to narrow the field; the expensive, physics-grounded methods go second on the survivors. That layering — cheap model to triage, expensive method to confirm — is the same pattern that shows up whenever a learned approximation meets a rigorous-but-slow ground truth, and getting the order right is most of the engineering.

The access fight, which is not a footnote#

The release of AlphaFold3 turned into one of the more revealing episodes in computational biology, and it is worth understanding because it tells you something about how this technology will actually be governed.

The Nature paper landed without code. For a field built on reproducibility, publishing a method you cannot run was a problem, and the journal drew open criticism for it. DeepMind’s initial answer was a hosted AlphaFold Server with hard daily limits and — pointedly — no ability to model arbitrary protein-ligand interactions, the single most commercially valuable capability. More than a thousand scientists signed an open letter objecting that a landmark result had been gated behind a restricted web form.

The reason for the gating was not subtle. Isomorphic Labs, the Alphabet drug-discovery company spun out of DeepMind, uses AlphaFold3 internally as a commercial asset. Open weights for predicting drug binding are precisely the thing a drug-discovery company does not want to hand to every competitor for free.

In November 2024, under pressure, DeepMind released the code and made model weights available to academics on request, under a non-commercial license. That is a real improvement and a real boundary. Academic and non-profit researchers can now run the model. Commercial use is off the table without a separate arrangement. The episode is a preview of how frontier scientific models will be released for years: capability and openness traded against the commercial value of the thing, with the most valuable capabilities the last to open up.

The strategic read for builders#

If you are building anything downstream of structure prediction, the licensing terms are part of the architecture, not a legal afterthought. A pipeline anchored on weights you may only use non-commercially is a pipeline with a ceiling. The open-source ecosystem around the field — independent reimplementations, the broader family of folding and design models — exists partly because that ceiling is real and teams need to route around it.

The deeper lesson is the one we give every client weighing a frontier model into their stack: separate the capability from the provider. The capability here, all-atom complex prediction by diffusion, is now public knowledge and will be re-implemented and improved by many groups. The specific weights and their license are a vendor decision that can change under you. Build your Operational Automation so that the structure-prediction step is a swappable component with a defined interface, benchmark the options on your own targets, and never let a single licensed artifact become load-bearing. The science is moving too fast, and the access rules too unpredictably, to bet a program on one model’s terms of service.

---

Structure prediction is a triage layer, not a magic answer. We build the validation loops, benchmarks, and swappable model interfaces that turn it into real AI implementation. Talk to our engineers.