GraphQA: protein model quality assessment using graph convolutional networks

Abstract

Motivation: Proteins are ubiquitous molecules whose function in biological processes is determined by their 3D structure. Experimental identification of a protein's structure can be time-consuming, prohibitively expensive and not always possible. Alternatively, protein folding can be modeled using computational methods, which however are not guaranteed to always produce optimal results. GraphQA is a graph-based method to estimate the quality of protein models, that possesses favorable properties such as representation learning, explicit modeling of both sequential and 3D structure, geometric invariance and computational efficiency.

Results: GraphQA performs similarly to state-of-the-art methods despite using a relatively low number of input features. In addition, the graph network structure provides an improvement over the architecture used in ProQ4 operating on the same input features. Finally, the individual contributions of GraphQA components are carefully evaluated.

Availability and implementation: PyTorch implementation, datasets, experiments and link to an evaluation server are available through this GitHub repository: github.com/baldassarreFe/graphqa.

Supplementary information: Supplementary data are available at Bioinformatics online.

Fig. 1.

Protein QA. GraphQA predicts local and global scores from a protein’s graph using message passing between chemically bonded or spatially close residues. CASP QA algorithms score protein models by comparison with experimentally determined conformations

Fig. 2.

Protein representations for learning. Sequential representations for LSTM or 1D-CNN fail to represent spatial proximity of non-consecutive residues. Volumetric representations for 3D-CNN fail instead to capture sequence information and are not rotation invariant. Protein graphs explicitly represent both sequential and spatial structure, and are geometrically invariant by design

Fig. 3.

Joint plots of LDDT and GDT_TS scores on CASP13. The marginal plots show the distribution of true versus predicted scores

Fig. 4.

Trade-off between the number of message-passing layers and the connectivity of the protein graph (CASP11)

Fig. 5.

Ablation study of node (top) and edge (bottom) features (validation results on CASP 11). All node features improve both local and global scoring. DSSP features are marginally more relevant for LDDT. Richer edge features benefit LDDT predictions the most, while bringing little improvement to GDT_TS

Fig. 6.

Gradient magnitude of predicted LDDT score w.r.t. the edges of the input graph (T0773). In the edge matrix, a darker red indicates a higher magnitude. The attributions for residue 20 (left) and 60 (right) reveal the long-range dependencies between residues captured by GraphQA