Loading…
Conditional diffusion model that generates P-type ATPase backbone conformations in a specified E1, E1P, E2P, or E2 functional state.
No providers recorded yet. Browse all providers
A P-type ATPase pumps its substrate by cycling through four conformations — E1, E1P, E2P and E2 — so the cycle, not any single snapshot, is the function. A predictor trained to return one answer per sequence collapses it: given human secretory-pathway Ca2+-ATPase (hSPCA1) and custom templates for all four states, ColabFold returned models whose TM-scores to the four experimental structures were nearly identical, unable to distinguish E1 from E2P. What is missing is not a better single structure but a sampler that can be told which state to produce.
PtypeATPaseGenerator, from Jingtian Xu and Yong Wang at the College of Life Sciences, Zhejiang University, answers that with classifier guidance over a frozen diffusion backbone. Two of the four terms in its sampling score — the unconditional backbone score and the sequence-given-structure term — are taken unchanged from Chroma; one new component, a four-way state classifier, is trained; and an analytic membrane term is added. Because guidance updates the coordinates being sampled rather than any weights, a new target needs no retraining: the paper generates 60 conformations of hSPCA1, a protein deliberately excluded from training, from a single input structure.
The sampling schedule is the second idea. Rather than denoising from pure noise, the method injects noise into a known reference structure only as far as time T' = 0.65 and then denoises back under the conditioners — the authors liken it to heating and annealing. A lower T' keeps samples close to the reference; a higher one widens the spread. Conformational diversity therefore becomes an explicit dial, which unconditional generators do not offer.
The state classifier is a graph neural network encoder initialized from Chroma's GNNEncoder weights — four layers, node features of dimension 512 and edge features of dimension 192, with edges drawn from the 20 nearest backbone atoms plus 70 distance-weighted far-field neighbours — followed by eight-head attention pooling at hidden dimension 512 and an MLP with one 64-unit hidden layer emitting a four-way softmax. It was trained with cross-entropy and Adam for 200 epochs at batch size 8, the learning rate decayed linearly from 1e-5 to 1e-6, on an 80/20 split of 1,247 structures: 301 experimental structures curated from the InterPro IPR023298 superfamily in the PDB (released before May 2023) plus 946 frames from 5–10 ns CHARMM36m molecular dynamics simulations of four template systems in a DOPC bilayer. Validation accuracy was 0.85 at a loss of 0.3, weakest on the structurally heterogeneous E1 class.
Generation used classifier weight 1 with gradients clipped at max-norm 10, sequence weight 0.75, membrane weight 0.1 at an RMSD threshold of 4 Å, 500 denoising steps and a probability-flow ODE reverse process. Starting from the Ca2+-bound E1 cryo-EM structure of hSPCA1 (PDB 8IWP) and sampling 15 times per state, the 60 generated backbones matched the corresponding experimental hSPCA1 structures (8IWP, 8IWR, 8IWW, 8IWS, 8IWT and 8IWU) at TM-scores above 0.70, with per-domain RMSD near 2.0 Å. The ablation raises TM-score from 0.47 with no conditioners to 0.70 with all three, at roughly 480 seconds per structure on one NVIDIA GeForce RTX 3090. Outputs are backbone-only and need a separate side-chain modelling step.
The model suits structural biologists who hold one or a few experimental structures of a P-type ATPase and want physically plausible models of the states they lack. Generation runs from a single input structure plus a state label in about ten minutes per conformation on one consumer GPU, far cheaper than the microsecond-scale molecular dynamics that transitions in this family would demand, and more state-specific than template-based homology modelling. Generated ensembles can seed further simulation, supply state-specific starting points for docking, or interpolate gaps in a cryo-EM series; side chains must be rebuilt with a tool such as Modeller before atomistic use.
The work is a compact demonstration that conditional guidance over an existing protein diffusion model can add a capability the base model never had — functional-state control and membrane awareness — for the price of training one small classifier. That composability is the transferable lesson, and it is what distinguishes the approach from fine-tuned AlphaFold 2 variants and from unconditional ensemble generators such as AlphaFlow. The limits are real and the authors state them: the classifier covers four classes within a single transporter family, it is biased toward well-represented metastable states — accuracy fell sharply on the Ca2+-bound E1P-ADP conformation, which is 2.1% of the E1P training data — and the method depends on curated experimental structures and simulation trajectories, so extending it to other membrane protein families will take further work. The preprint has not been peer reviewed, the released code is documented as incomplete pending a fuller deposit, and running it requires an access token for Chroma's weights.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.