Structure-Based Protein Function Prediction using Graph Convolutional Networks

Abstract

Recent massive increases in the number of sequences available in public databases challenges current experimental approaches to determining protein function. These methods are limited by both the large scale of these sequences databases and the diversity of protein functions. We present a deep learning Graph Convolutional Network (GCN) trained on sequence and structural data and evaluate it on~40k proteins with known structures and functions from the Protein Data Bank (PDB). Our GCN predicts functions more accurately than Convolutional Neural Networks trained on sequence data alone and competing methods. Feature extraction via a language model removes the need for constructing multiple sequence alignments or feature engineering. Our model learns general structure-function relationships by robustly predicting functions of proteins with ≤ 30% sequence identity to the training set. Using class activation mapping, we can automatically identify structural regions at the residue-level that lead to each function prediction for every protein confidently predicted, advancing site-specific function prediction. De-noising inherent in the trained model allows an only minor drop in performance when structure predictions are used, including multiple de novo protocols. We use our method to annotate all proteins in the PDB, making several new confident function predictions spanning both fold and function trees.

Unsupervised protein embeddings outperform hand-crafted sequence and structure features at predicting molecular function

“…Our work continues upon two recent studies involving protein representation learning (Heinzinger et al, 2019) and its combination with contact maps applied in AFP (Gligorijevic et al, 2019). We confirm the power of the unsupervised ELMo embeddings in capturing relevant biological information about proteins (Heinzinger et al, 2019).…”

“…Equation (1) describes the diffusion of information about each amino acid to the neighboring residues, where the neighborhood is defined by the graph. We tested the model proposed by (Gligorijevic et al, 2019) that has three convolutional layers (GCN3_{E,1h,SA}_CM, Fig. S2).…”

“…S9-S11 (SM8) show all pairwise comparisons between the tested methods. The 3-layer GCN proposed in (Gligorijevic et al, 2019) trained with the ELMo embeddings (GCN3_E_CM) was worse than the 1DCNN_E based on all three metrics (S min =0.52, ROCAU C=0.75, Fmax=0.48, compared to S min =0.51, ROCAU C=0.77, Fmax=0.50).…”

“…These embeddings demonstrated competitive performance in both amino acid and protein classification tasks, such as inferring the protein secondary structure, structural class, disordered regions, and cellular localization (Heinzinger et al, 2019;Kane et al, 2019). Other works also trained LSTMs to predict the next amino acid in a protein sequence using the LSTM hidden state at each amino acid as a feature vector (Gligorijevic et al, 2019;Alley et al, 2019). Finally, a transformer neural network was trained on 250 million protein sequences, yielding embeddings that reflected both protein structure and function (Rives et al, 2019).…”

“…Only (Gligorijevic et al, 2019) have explored the effectiveness of a pre-trained sequence model in AFP, but it was done in combination with protein structure information using a GCN. We suspect that a deep pretrained embedding can be powerful enough to predict protein function, in which case the structural information would not offer any significant performance improvement.…”