Accurate structure prediction of biomolecular interactions with AlphaFold 3

If you've followed the AlphaFold story, you know the punchline for proteins: we can predict single-chain structures remarkably well. But the living cell doesn't do single chains. It's a tangle of proteins clasping other proteins, wrapping around RNA and DNA, grabbing small molecules, and bristling with post-translational decorations. Abramson and colleagues set out to tackle that full, messy theater with one model. They call it AlphaFold 3, or AF3, and the idea is simple to say but audacious to pull off: one system that can place every atom in proteins, nucleic acids, ligands, ions, and modified residues, all at once, and do it better than the specialized tools built for each niche. To make that leap, they rethought the architecture. Instead of AlphaFold 2's Evoformer, which leaned hard on multiple sequence alignments, AF3 runs a slimmer "Pairformer" trunk. Think of two streams of information: a per-residue track and a pairwise interaction track that describes how any two positions relate. AF3 keeps those and largely lets go of the heavy multiple sequence alignment machinery; the alignment input is downshifted and most of the learning flows through the pair representation. Then comes the big shift: rather than predicting a protein backbone frame and torsion angles, AF3's Structure Module is a diffusion model that operates directly on raw atomic coordinates. It uses no rotational frames or equivariant tensor gymnastics. It takes coordinates that have been deliberately noised and learns to denoise them in a multiscale way that first nails local stereochemistry when the noise is low. As the noise rises, it learns to assemble the big picture of how all the parts sit together. That move pays off in generality. Because the head works on atoms, not torsions, it doesn't care whether that atom belongs to a peptide, a base in RNA, a zinc ion, or a drug-like ligand. There's no special casing or bond library to maintain for exotic components. You hand it the inputs—polymer sequences, any residue modifications, and, for small molecules, a Simplified Molecular Input Line Entry System string—and the trunk distills those into the pair and single representations. The diffusion head then generates coordinates in one coherent shot. At inference time, they lean into the stochasticity: run multiple random seeds, draw several diffusion samples per seed, and you get a small gallery of candidates. In their default setup, that's five seeds with five samples each, so that adds up to 25 structures, and a ranking function—built around the inter-chain template match score, or ipTM—picks the most plausible one, with a harsh penalty if a pose has clashes or chirality errors. That ranking only works if the model knows how confident it is, and AF3 has a head for that too. It predicts per-atom predicted local distance difference test, or pLDDT; a predicted aligned error—the predicted aligned error matrix that shows where domains might shift relative to each other; and a distance error matrix called predicted distance error, or PDE, that summarizes how far the predicted inter-atomic distances are likely to be off. They also inoculated the model against a classic diffusion-era problem: hallucinating tidy helices where none should be. Training incorporated cross-distillation from AlphaFold-Multimer version 2.3, which tends to render disordered regions as long, floppy loops. Learning from that pattern taught AF3 to leave the spaghetti as spaghetti when the biology calls for it. The training and sampling schedule is carefully staged. They start with smaller crops and end larger—think windows growing from roughly 384 residues up to 768—so the trunk gets used to bigger assemblies. Early in training, each optimizer step generates around twelve thousand diffusion samples; in fine-tuning, that drops to about eight thousand as the model shifts from exploration to polishing. That's a lot of denoising, and it shows up in the learning curves. Local structure, the intra-chain details, comes together fast: those metrics sail past 97 percent of their eventual peak within the first twenty thousand steps. Interfaces are harder. Protein–protein contact quality doesn't reach that same 97-percent mark until roughly sixty thousand steps in, a nice quantitative picture of how global organization lags behind local stereochemistry during learning. So how does the model actually perform when you take it out of the training loop and into benchmarks? Let's start with small molecules, because docking is a famously tough space. AF3 was put through PoseBusters, a set of 428 recent protein–ligand complexes, and scored on a strict criterion: success means the ligand ends up within 2 angstroms of the native pose when you align around the pocket. Against that yardstick, AF3 beat the classical docking workhorse AutoDock Vina by a mile—Abramson and colleagues report a Fisher exact p-value of two point twenty-seven times ten to the negative thirteen—and it also outstripped RoseTTAFold All-Atom with an even stronger signal at four point forty-five times ten to the negative twenty-five. The takeaway is straightforward. Even when you don't feed AF3 the crystallographic context—just the protein sequence and the ligand's Simplified Molecular Input Line Entry System—it places ligands where traditional scoring functions often don't. It isn't just bare proteins and ligands either. AF3 handles covalent chemistries and glycosylation, and they quantify that the same way: what fraction of cases put the sugar in the right pocket pose? On high-quality data, multi-residue glycans land within 2 angstroms in 42.1 percent of cases; single-residue glycans do slightly better at 46.1 percent. That's not perfect—it reflects how complex carbohydrate chemistry can be—but it's meaningful traction across a class many docking tools side-step or treat with bespoke add-ons. Move from sugars to nucleic acids and the pattern holds. On protein–RNA and protein–DNA assemblies—particularly those under about a thousand residues—AF3 predicts interfaces more accurately than RoseTTAFold version 2 for nucleic acids. On low-homology sets where templates can't save you, the gains are statistically clean: for RNA targets, the Wilcoxon signed-rank test comes in at a p-value of one point six times ten to the negative seven, and for DNA, it's a p-value of five point two times ten to the negative twelve. In the CASP fifteen RNA exercises, AF3 edges past both RF2NA and AIchemy_RNA2 on the overlapping targets they all attempted, though there are still individual CASP fifteen cases where AIchemy_RNA2 takes the lead. That nuance matters. It shows that AF3's generalist approach buys you consistency across categories, but there are still corners where a specialist can win. Protein–protein interfaces show a similar story, and the antibody–antigen case is the stress test everyone cares about. Across a broad set of complexes, AF3's interface quality—measured by DockQ, which blends contact overlap, interface root-mean-square deviation, or RMSD, and fraction of native contacts—improves over AlphaFold-Multimer version 2.3. If you set a modest success bar at DockQ above 0.23, the paired comparison across targets yields a p-value of one point eight times ten to the negative eighteen in favor of AF3. And within those complexes, monomer quality ticks up too: single-chain local distance difference test improvements are highly significant, with a p-value of one point seven times ten to the negative thirty-four. That's the general lift. The targeted antibody analysis adds a practical twist: sampling depth matters. When the team ranked models drawn from roughly a thousand random seeds, both the percentage of correct interfaces and the percentage of very high-accuracy models rose. The trend isn't just eyeballed; the gains carry p-values of two point zero times ten to the negative five for overall correctness and zero point zero zero nine for the very best category. If you simply compare the top-ranked model chosen from the usual five seeds against the top-ranked from that deeper thousand-seed pool, the advantage for the deeper search is again statistically solid at a p-value of six point five times ten to the negative five. Translation for practitioners: for difficult paratopes and epitopes, it pays to cast a wider net. Underneath these scores is a confidence system you can actually use. AF3's atom-level predicted local distance difference test tracks how much to trust the local geometry, the predicted aligned error matrix tells you which domains might be sliding around, and the predicted distance error distance errors fill in the interface uncertainty. In practice, those measures correlate well with interface quality metrics like DockQ or interface local distance difference test, or iLDDT, so you can translate familiar cutoffs and set thresholds without resorting to bespoke heuristics for each modality. None of this is magic, and the paper is candid about the rough edges. AF3 gives you a static snapshot. Multiple random seeds don't conjure a Boltzmann-weighted ensemble in solution, and large-scale motions remain beyond its reach. Stereochemistry isn't flawless: chirality violations show up in about 4.4 percent of cases in the PoseBusters tests, even with a ranking penalty that tries to demote those poses. Atom clashes are still more common in the complex assemblies—think protein–nucleic acid complexes with more than 100 nucleotides and a few thousand residues—where getting the last steric centimeter right is genuinely hard. And because AF3 still uses evolutionary information, just in a different way, the depth of your multiple sequence alignment matters: shallower alignments tend to correlate with lower single-chain local distance difference test. There's also the meta-message that comes from how the model was built. Diffusion, done this way, is well tuned to multi-component biology. By decoupling representation learning in the trunk from coordinate generation in the head, AF3 can learn the grammar of interfaces in the pairwise track and then render atoms for proteins, bases, and ligands without switching parameterizations. It also sidesteps some of the contortions earlier structure modules needed to remain consistent with torsion angle libraries. In effect, it uses the same denoising playbook whether it's shrinking a wobbly methyl group into place or pulling an entire ribonucleoprotein complex into register. And it's not just success by benchmark. The sampling and ranking protocol tells you how to use the model. Generate a small panel of candidates—25 by default—and let the confidence head and inter-chain template match-based score sort them while automatically down-ranking models that collide or flip chirality. For antibodies or other complex interfaces, spend the compute and expand the seed pool; the statistics say you'll be rewarded. For ligands, don't hesitate to put in modified residues or grafted glycans; the model has learned enough of that chemistry to place them plausibly. If you're wondering how to get your hands on it, the authors say AF3 will be available as a non-commercial server, with some restrictions around ligands and covalent modifications. That's a pragmatic compromise while the field figures out where to draw lines on proprietary chemistry and safety. The big picture is still clear. As Abramson and colleagues argue, a single diffusion-based framework can now cover most of the biomolecular landscape with state-of-the-art accuracy, often beating the specialists on their home turf and providing usable confidence maps alongside the coordinates. I'll end with the right kind of speculation—brief and cautious. Bringing proteins, nucleic acids, and small molecules under one generative roof opens a path to modeling pathways end-to-end, not just parts in isolation. You could imagine quickly triaging designs for an RNA-binding protein, a glycoengineered antibody, and a covalent inhibitor in the same workflow, with the same confidence readouts. But AF3 is still a snapshot tool, not a simulator. The next frontier is marrying these static but accurate pictures with dynamics and energetics. For now, though, the center of gravity has shifted. We don't have to pick a different tool for every kind of biomolecule. One model can speak most of the languages of the cell, and it's fluent enough to be useful.