Loading…
Full-atom protein model accuracy estimation, regressing per-atom lDDT with an SE(3)-transformer over a heavy-atom graph of the modeled structure.
No providers recorded yet. Browse all providers
In a photograph, neighboring pixels are more strongly correlated than distant ones, and a 2D convolution is built to exploit that. Protein distance maps break the assumption: entries far apart in the map often move together, because the residues they index sit side by side in three dimensions. DeepAccNet, the CASP14 estimation-of-model-accuracy (EMA) method from the same lab, learned ResNet kernels over such maps, then added 3D convolutions scanning a 24 Å box around each Cα to recover what the 2D stack missed.
Pluto removes the distance map entirely. It represents a modeled structure as a graph over its heavy atoms and processes that graph with an SE(3)-transformer, a self-attention network equivariant to rotation and translation of the input coordinates by construction — the right symmetry for a task whose answer must not change when a structure is rotated. Every heavy atom is a node, and the network regresses that atom's full-atom lDDT: how far it has drifted from where the experimental structure would place it. The hard part is reach: a dense atom graph exhausts GPU memory, a sparse one leaves atoms many hops apart. Pluto keeps dense local webs among side-chain heavy atoms and adds sparse Cα-to-Cα "highway" edges, so any atom can reach any other in a few message-passing steps.
Naozumi Hiranuma developed Pluto at the Institute for Protein Design and presented it in his 2022 University of Washington doctoral dissertation, supervised by David Baker, as the graph-based successor to DeepAccNet. It sits alongside single-model accuracy estimators such as DeepUMQA and GraphGPSM, and is unrelated to PLUTO, the histopathology foundation model that shares its name.
Pluto's graph places a node on every heavy atom, keeping the 32 nearest Cα–Cα neighbors, the 12 nearest heavy–heavy, and the 8 nearest Cα–heavy. The final network uses 8 layers, 2 degrees, and 48 channels, trained with Adam at a learning rate of 1e-5 across 8 A100 GPUs against a mean-squared-error loss on true full-atom lDDT. Training data is the DeepAccNet corpus — roughly two million conformations spanning 21,000 protein sequences, with decoys from RosettaCM comparative modeling, native-structure perturbation, trRosetta-guided folding, and RoseTTAFold — extended with PDB entries deposited by February 2020 at a 4 Å resolution cutoff, adding 14,210 sequence clusters. On held-out test structures Pluto beats DeepAccNet-Standard at p < 0.05, though the absolute margin is small at roughly 0.0005 MSE. It assigns crystal structures a median predicted global lDDT of 0.976, and is less accurate on side-chain heavy atoms than on backbone atoms, bulky ones hardest. An ablation building the atom graph only within a 16 Å ball around one residue scores 0.0292 MSE against Pluto's 0.011 — the case for the highway edges.
Accuracy estimation turns a pool of candidate models into a decision: which prediction to submit, which region to refine, which design to carry to the bench. Pluto serves the last directly: applied to de novo miniprotein binders designed against EGFR, IL-7Rα, IL-6R, and SARS-CoV-2 receptor-binding domain targets, it scores whole complexes as a filter for designs that will not bind, benchmarked by ROC analysis against Rosetta's ddG interface energy; the accuracy estimators pick out successful binders about as well as the energy-based score. Featurization inherits DeepAccNet's Rosetta dependency, so the model is a component of a Rosetta pipeline rather than a standalone package.
Pluto's contribution is architectural rather than a leap in accuracy. It shows that an equivariant graph transformer over atoms can edge out a strong convolutional EMA baseline, and that sparse connectivity makes whole-protein atom graphs tractable — a pattern that recurs in later all-atom work from the same lineage, including the Vilya-1 macrocycle model. The dissertation is candid about the limits: the margin over DeepAccNet is significant but small, and because EMA methods learn the error modes of whatever generated their decoys, staying current requires periodic retraining as structure predictors improve. Chapter 7 of the same thesis extends the graph approach to peptides and protein–DNA complexes as GAAP. Neither code nor weights were released; the dissertation states that both are available upon request.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.