A compact neural network learns to generate shape-shifting protein structures

The work

Technical · from the record

Phanto-IDP is a variational autoencoder for protein backbones: a graph convolutional encoder reads backbone N, CA and C atoms as a graph, a variational layer compresses them, and a three-block transformer decoder outputs Cartesian coordinates. It was trained separately on each target system using 1 μs all-atom MD trajectories from which 50 000 conformations were extracted, covering 10 intrinsically disordered proteins and four ordered proteins under 200 residues. On a held-out set of 12 500 conformations, average backbone reconstruction RMSD was 0.511 Å for RS1, 0.885 Å for PaaA2 and 2.714 Å for α-synuclein, and the model generated 50 000 α-synuclein conformations in about 50 s on a Tesla V100 after 30.91 h of training. When trained on a short MD trajectory, 5.62% of the generated structures were in a folded state, compared with 18.3% sampled by replica exchange molecular dynamics, and latent-space interpolation between two selected conformations produced continuous transition paths.

How AI was used

Converged all-atom MD trajectories were simulated with the ESFF1 force field and OPC3 solvent model for 1 μs per protein, and 50 000 conformations were extracted at 0.02 ns intervals, shuffled and split into training, validation and test sets (50%/25%/25% for the main systems; 80% training for the short AAQAA3 trajectory). Each conformation was reduced to its N, CA and C backbone atoms and converted into an atom-level crystal graph with one-hot atom-type node features, a 30-nearest-neighbour adjacency and SE(3) equivariant edge features built from local frames computed with the mylddt toolset. A graph convolutional encoder with edge gating produced atom embeddings, a fully connected layer reduced dimensionality to residue-level features, and a variational layer produced mean and log-variance matrices from which latent codes were drawn by reparameterisation at sampling temperature T = 0.02; a decoder of three transformer blocks, each with a self-attention layer and an update module, emitted backbone Cartesian coordinates. Training used a frame-aligned point error reconstruction loss computed in Gram–Schmidt local frames plus a KL term whose weight was increased dynamically to avoid posterior collapse, for 400 epochs per system on one Tesla V100. The trained decoder was then run on seeds from a normal distribution to generate new conformation ensembles and on linear interpolations between the latent codes of two selected conformations; generated backbones were passed through a side-chain refinement step, and ensembles were compared with MD and REMD trajectories and with published experimental observables.