Highly accurate protein structure prediction with AlphaFold

Let me start with the problem everyone in structural biology has felt in their bones. We can read life's code faster than ever, but turning that code—the amino acid sequence—into a reliable three-dimensional structure has lagged decades behind. Genomes exploded; structures trickled. As Demis Jumper and colleagues put it, experimental structure determination stayed the bottleneck, so our maps from sequence to function grew spotty just where we needed them dense. Then CASP fourteen happened, and the floor moved. AlphaFold didn't just edge out the field; it lapped it. On the blind targets in CASP fourteen, the median backbone error was under an angstrom—0.96 angstroms root mean square deviation at 95 percent residue coverage—while the next best method sat at 2.8 angstroms. All atoms told the same story: 1.5 angstroms for AlphaFold versus 3.5 angstroms for the runner-up. Think about what that means. Sub-angstrom backbone accuracy is the difference between a model that hints at a fold and a model you can build chemistry on. They even showed a single chain of 2,180 residues with the right domain packing, no human hand-tuning. And crucially, you don't just get a structure; you get self-awareness. A per-residue confidence, p L D D T, flags what to trust locally, and a global confidence, p T M, estimates how close the whole fold is to the truth. So you can tell, before you ever compare to an experimental map, whether the model is probably right. How did they do it? They reframed structure prediction as graph reasoning in three dimensions, and made the evolution signal talk to geometry at every step. The trunk of the network, which they call the Evoformer, holds two intertwined views of the protein. One is the multiple sequence alignment, or MSA—stacked related sequences that expose which positions co-vary across evolution. The other is a pairwise map—one slot for every residue pair—where the network reasons about which residues should end up close in space. Information flows continuously between these views. The MSA updates the pair map by an operation that's like an outer product running over all the aligned sequences, so "this column talks to that column" becomes "these two residues probably touch." And the pair map pushes back into the MSA by biasing attention, so spatial guesses refine which evolutionary couplings the model pays attention to. Inside the pair map, two geometric moves enforce consistency. In a triangle of residues—i, j, k—if you know two edges, you can often infer the third. AlphaFold bakes that idea in with a triangle multiplicative update, a cheaper, symmetric cousin of attention that uses two edges to update the third, and a triangle-aware self-attention that adds a bias for the "missing edge." Together, they nudge the pair map toward a geometry you could actually realize in three dimensions. This is the inductive bias part of the story: not hand-coded energy terms, but wiring that makes geometric sense. Then, and only then, the network commits to coordinates. The structure module gives each residue its own local coordinate frame—think of a tiny rigid scaffold riding on each carbon alpha—and treats the whole chain as a residue gas. Within those frames, it prioritizes backbone geometry so side chains have a stable anchor. The key move here is invariant point attention, or IPA. Each residue produces a handful of three-dimensional query and key points in its local frame; the model projects them into a common global frame, lets them attend to each other, then pulls the result back into each local frame. That sounds abstract, but it buys something concrete: if you rotate or translate the whole protein in space, the computation—and the outcome—doesn't change. After IPA, an equivariant update actually moves those local frames in three dimensions in a way that stays faithful to rigid body physics. And, interestingly, the module is allowed to "break the chain" locally so it can fix bad kinks without being shackled to exact peptide geometry; stereochemical clean-up comes later with a short Amber relaxation. That clean-up removes local violations but doesn't change the global accuracy, which tells you the network itself is placing atoms well. What keeps all this honest is the loss function. Instead of just comparing one set of coordinates to another, AlphaFold measures error in a frame-aware way. For each residue, it aligns the predicted local frame to the true local frame and asks, across many such alignments, how far the predicted atom positions are from the true ones. Those distances feed a clamped L one penalty so outliers don't dominate. That's the frame-aligned point error, or FAPE. Around it, the team added a few extra training signals. From the pair map, they predict a distogram, essentially a histogram over inter-residue distances, with a cross-entropy loss. From the final activations, they predict p L D D T, the per-residue accuracy. And on the sequence side, they mask out chunks of the MSA and train a BERT-like objective to reconstruct them, which sharpens the network's sense of evolutionary covariation. Then they made the model teach itself. After training a strong model, they used it to predict structures for roughly 350,000 diverse sequences from Uniclust thirty, filtered down to the high-confidence ones, and mixed that "pseudo-labeled" dataset back into training from scratch. It's a kind of self-distillation. The effect was large: more confident, more accurate predictions, especially when true structures were scarce. And the last trick is recycling. They run the Evoformer and the structure module, take the output, and feed it right back into the same stack for another pass—up to four iterations in their analyses. Each pass refines the coordinates and the pair map in lockstep. You can actually watch the structure tighten up over the 48 Evoformer blocks per iteration, with 192 intermediate structures across four recycles. What about the data fuel behind this? The model thrives on deep MSAs, and the team built those with huge databases. Two names matter because you can see their fingerprints in the results: BFD, a massive metagenomic sequence collection, and MGnify, another metagenomics resource. When they removed BFD from the MSA search, mean global distance test score—GDT, a fold-level metric—dropped by about 0.4 points; removing MGnify cost 0.7. Remove both, and the hit was big: a 6.1 point drop on average, with a minority of targets falling by more than 20. That maps to a simple intuition. You need diversity in the alignment to read coevolution. But there's a threshold. AlphaFold's accuracy climbs quickly as the effective number of sequences rises toward a few dozen, and then the gains taper beyond roughly a hundred. In plain English: the MSA helps lock in the coarse fold early, and later refinement leans more on learned geometry than on ever-deeper alignments. Speed and scale matter if you want to move a whole proteome. On a single NVIDIA V one hundred, a no ensemble run takes about 0.6 minutes for a 256 residue protein, 1.1 minutes for 384 residues, and around 2.1 hours for a 2,500 residue behemoth. Ditching ensembling—a technique they used in CASP fourteen for model selection—costs little in accuracy and buys roughly an eightfold speedup. Memory scales roughly with the square of the sequence length, so a 16 gigabyte card fits proteins up to around 1,300 residues comfortably; larger ones need multiple GPUs. That's not desktop friendly for the biggest monsters, but it's perfectly workable in the cloud, and it puts proteome-scale predictions within reach. Did all this translate beyond the carefully curated CASP targets? They checked. Jumper and colleagues assembled a validation set of 10,795 protein chains deposited in the Protein Data Bank after their training cutoff. The same accuracy patterns held up on these unfiltered, real-world structures. More importantly, the model's confidence estimates tracked reality. Per-residue p L D D T maps to actual per-residue l D D T almost linearly—nearly a one-to-one slope with a small negative offset—and across chains, predicted T M score, p T M, maps to actual T M score with a slope just under one and an offset of about 0.07. Correlations were strong, with Pearson's r around 0.85 for the global metric. In practice, that means you can treat the model's own confidence as a trustworthy guide: high p L D D T, you can place side chains; lower p L D D T, be cautious or look for help from experiments. There are limits, and they're instructive. Proteins that live mostly by touching other proteins—bridging domains in large complexes, or chains whose key contacts are heterotypic rather than within-chain—are harder. AlphaFold did well on homomers, even when the monomers are interwoven; they highlight a homotrimer from PDB entry six S K zero that came out right without feeding in the stoichiometry. But interfaces between different chains, or assemblies where the single-chain context is weak, are still challenging. That squares with the MSA story: if the signal for within-chain couplings is thin, the model has less to grab onto. It doesn't mean complexes are off-limits; it means the single-chain engine isn't the whole job. One of my favorite parts of the paper is how they "opened the hood" to watch the gears turn. They trained a lightweight structure head for each of the 48 Evoformer blocks, froze the rest of the network, and produced a movie of intermediate structures across recycling rounds. Some targets—the transporter LmrP in CASP, for instance—snap into the right fold within the first few layers. Others, like the SARS-CoV-2 Orf8 protein, need many layers and nearly the full stack to untangle. The trajectories are smooth after the initial blocks, which tells you the model doesn't lurch from guess to guess; it tightens and rearranges in a controlled way. Take away recycling, and accuracy drops markedly; put it back, and hard cases resolve with minimal extra training time. Layer by layer, you can see why the exchange between the MSA and the pair map matters: the outer product updates push pairwise contacts into focus, the triangle operations enforce consistency, and IPA turns that into physical coordinates without ever losing track of global symmetry. Let's zoom back out. The headline numbers from CASP fourteen—0.96 angstroms median backbone error for AlphaFold versus 2.8 angstroms for the next best—weren't a parlor trick. They flowed from design choices that let evolutionary context and three-dimensional geometry inform each other all the way through the network. Confidence estimates weren't tacked on; they were learned from the same final activations that built the coordinates, and they calibrated cleanly to reality. The data engine behind the model—deep alignments from resources like BFD and MGnify—mattered most when the evolutionary signal was thin and mattered less once you crossed that "dozens of sequences" threshold. And the whole thing ran in minutes to hours on accessible hardware, with per-residue maps that told you where to trust, where to check, and where to hand off to cryo-electron microscopy or molecular replacement in crystallography. In fact, the authors note those workflows are already using the models in practice. If you want a cautious look forward, it's this. Heteromeric complexes and assemblies remain the next frontier; the single-chain machinery excels at homomers but strains when the key contacts live across chains. And sequences that simply don't have deep alignments—orphans from the far reaches of the tree of life—still pose a challenge. But the template for progress is clear. Keep the geometric priors. Keep the bidirectional flow between coevolution and three-dimensional reasoning. Keep training signals that tie local frames to atomic positions. And, as Jumper and colleagues showed, let the model learn from its own confident predictions when data are scarce. That combination cracked a decades-old problem. It's hard to think of a better reminder that in biology, as in life, structure emerges when the right kinds of information finally get to talk to each other.