Accurate prediction of protein structures and interactions using a three-track neural network

Let’s start at the moment the ground shifted. The Critical Assessment of protein Structure Prediction, or CASP14, made it clear that protein structure prediction wasn’t a straight line from sequence to shape anymore. DeepMind’s AlphaFold2 redrew the map, and almost immediately Baek, DiMaio, Anishchenko and their colleagues asked a simple question with big consequences: what if you stop forcing the problem into one direction? What if a model reasons about sequence, inter-residue geometry, and actual three-dimensional coordinates at the same time? Their answer is RoseTTAFold, a three-track network that thinks in one-dimensional, two-dimensional, and three-dimensional all at once. One track reads the amino-acid sequence and its evolutionary hints, one lives in pairwise distances and orientations between residues, and one carries the coordinates that define the fold. Crucially, these tracks talk to each other, constantly, in both directions. When the three-dimensional track proposes a backbone twist, the two-dimensional track can nudge it by tightening distances. When the two-dimensional map suggests a contact, the sequence track can ask whether evolution supports it. That constant cross-checking is the point. In head-to-head tests, the three-track design beat the team’s own two-track baselines and outperformed the leading automated servers on the CASP14 targets. It didn’t top AlphaFold2, but it lived in the same neighborhood, which, at that moment, was astonishing. Now, getting that to run in practice takes some choreography. Full-length proteins blow up memory quickly, so the team trained and ran the model on many discontinuous crops: two segments that together span about two hundred sixty residues, sampled across the sequence. Think of it like panning across a massive mural, learning the local scenes, then stitching them into a whole. And there are two ways out of the network. In one route, the model’s predicted distances and orientations are handed to pyRosetta, which builds an all-atom model using classical energy functions and sampling. In the other, a final layer that is SE three–equivariant—meaning it respects rotations and translations in three-dimensional space—generates backbone coordinates directly, end to end. The second route is fast. For proteins under roughly four hundred residues, you can get backbones in about ten minutes on a single RTX 2080 graphics processing unit. If you want full-atom detail through pyRosetta, budget around an hour on a mid-sized central processing unit box after about five minutes of network time. Compared with some alternatives, the end-to-end path is memory-thrifty; they report running models for larger proteins on roughly eight gigabytes of graphics processing unit memory where others needed triple that. Here’s where the three-track idea really changes the game: complexes. Traditional modeling builds each subunit separately and docks them like puzzle pieces. RoseTTAFold skips that. Feed in multiple sequences, let the network see inter-chain context, and it predicts the complex directly. That’s flexible-backbone docking by construction, not as a bolt-on. In two- and three-chain tests, many models hit Template Modeling scores, or TM-scores, above zero point eight, which is a strong indicator that the overall fold and arrangement are right. And the method scales. They modeled the four-chain IL-12 receptor bound to IL-12 and fit it into cryo-electron microscopy maps with the correct topology, even pointing to a shared contact between IL-12p35 Y189 and IL-12R beta 2 G115, an echo of packing seen in the related IL-23 system. If you’ve ever stared at multiple sequence alignments, you know why this works. Co-evolving residues whisper about contacts. The deeper the alignment, the clearer the whisper. Baek and colleagues showed that giving the network more sequences improves complex accuracy; inter-chain co-variation helps place the parts. They also saw a benefit from combining predictions across many discontinuous crops rather than relying on a single full-length pass. It’s a bit like taking multiple exposures to capture a high-dynamic-range photo—you average them, and details pop. To ease memory limits while still tapping larger alignments, they experimented with a Perceiver-style cross-attention module: start from a compact seed alignment and attend over thousands more sequences, then distill the signal back into a manageable form. In their released models, they used the top one thousand sequences; the early Perceiver results looked promising but needed more training to benchmark cleanly. Speed and accuracy are only part of the story. What really matters is whether the models help you do science. On that front, the wins are concrete. For X-ray crystallography, molecular replacement lives or dies on whether your search model is close enough to the truth. In four cases where templates and older predictions failed, RoseTTAFold cracked it. The SLP example stands out: their model superimposed ninety-five C alpha atoms within three angstroms, with a backbone root-mean-square deviation of zero point ninety-eight angstroms—while the closest template only covered fifty-four positions and was off by about one point seven angstroms. That’s the difference between a solved structure and a stalled project. In cryo-electron microscopy, predicted models slid into density maps and guided refinement. The p101 protein G beta gamma-binding domain inside a phosphoinositide three-kinase gamma complex, for instance, lined up with the beta-sheet density to about three angstroms over the core, which is accurate enough to interpret the interface. On benchmarks, the pattern repeats: this integrated one-dimensional, two-dimensional, and three-dimensional reasoning buys you reliability. In the blind Continuous Automated Model Evaluation, or CAMEO assessment—sixty-nine medium and hard targets over a set period—RoseTTAFold outperformed all the other servers running at the time. While Baek and colleagues are clear that they didn’t match AlphaFold2 across CASP14, they consistently beat the other automated groups and their own earlier two-track variants. It’s a reminder that smart architecture plus careful training can move the needle even when the headline crown stays with someone else. Where the method really starts to serve biology is in turning structures into hypotheses. Take G protein–coupled receptors, those shape-shifting membrane sensors that toggle between inactive and active states. For human G protein-coupled receptors without solved structures, the team generated models for both states. Even without close templates, the predicted shapes compared favorably to curated G protein-coupled receptor model databases. They also packaged each model with an accuracy estimate using DeepAccNet, a learned score that correlated well with actual quality on tests. That’s valuable if you’re trying to decide whether a pocket is trustworthy enough to model a ligand or whether a cytoplasmic loop might plausibly touch a G protein. Then there’s the broad sweep across disease-related human proteins. They targeted six hundred ninety-three domains that lacked nearby structural homologs but carried clinical weight—mutations, phenotypes, you name it. More than a third came back with high predicted confidence, a local Distance Difference Test score, or lDDT, above zero point eight. To calibrate that, on CASP14 targets with similar scores, the average backbone error was about two point six angstroms. That’s good enough to reason about folds, active sites, and whether a substitution might jam the works. Three case studies make it real. The TANGO2 protein, tied to metabolic crises in children, came out as an N-terminal nucleophile hydrolase, with the right catalytic geometry. Disease mutations clustered where you’d expect trouble. Variants like R26K, R32Q, and L50P sat near the catalytic center and could disrupt chemistry; G154R wedged a bulky, charged side chain into a hydrophobic core, a recipe for misfolding. The ADAM33 protein, implicated in asthma, revealed a lipocalin-like barrel in its prodomain, and conserved residues lined a patch consistent with a cysteine-switch that locks the metalloprotease quiet until activated. Ceramide synthase one, a key enzyme in sphingolipid metabolism, showed six transmembrane helices forming a central cavity with conserved residues—His182, Asp213, Trp298—that make chemical sense. A pathogenic H183Q right next door? That tracks with a direct hit to the active site environment. Let’s circle back to complexes for a second, because the implications ripple outward. Modeling multi-chain assemblies directly from sequence isn’t just faster—it’s different. Because the chains fold in the context of each other, you capture induced fit and subtle packing that rigid docking tends to miss. The IL-12 receptor story illustrates that. Fitting a four-chain model into cryo-electron microscopy density and seeing a conserved contact like the IL-12p35 Y189 to IL-12R beta 2 G115 link, plus a plausible role for the IL-12R beta 2 N-terminal peptide in binding, gives you immediate, testable ideas. And you got there in about thirty minutes on a single twenty-four gigabyte graphics processing unit, not after days of docking and manual tweaking. Availability matters here. This isn’t a closed box. The group put a public server online through the Robetta portal, and they released code and trained weights on GitHub and Zenodo. They also pushed out model sets for communities that need them—G protein-coupled receptors in different activation states with confidence scores, and structures that assisted experimental deposits like the phosphoinositide three-kinase gamma complex. That openness is part of why the method spread so quickly: you could try it today on your favorite protein, and you still can. So, what’s the shape of the lesson? Integrating sequence, geometry, and coordinates in one loop lets the model reconcile constraints as it goes, instead of patching them together afterward. That design shows up in every result we’ve talked about: speed without giving up accuracy, flexible complexes instead of rigid guesses, and enough reliability to solve crystallographic phases or interpret a fuzzy electron microscopy map. It isn’t AlphaFold2, and Baek and colleagues never claim it is. But it’s close enough, and different enough, to expand what’s possible—especially for complexes and for groups that need a fast, open tool. Looking ahead, the path to better will sound familiar: train directly on complexes so the model learns inter-chain grammars, lean harder on templates where they exist, and scale up the sequence signal with smarter attention over massive alignments. The Perceiver-style experiments hint at one route to that. Hardware will keep easing the constraints. But the core idea is already here, already useful: let the protein tell its story in one-dimensional, two-dimensional, and three-dimensional at once, and listen with a model built to hear all three.