P3CMQA: Single-Model Quality Assessment Using 3DCNN with Profile-Based Features

Abstract

Model quality assessment (MQA), which selects near-native structures from structure models, is an important process in protein tertiary structure prediction. The three-dimensional convolution neural network (3DCNN) was applied to the task, but the performance was comparable to existing methods because it used only atom-type features as the input. Thus, we added sequence profile-based features, which are also used in other methods, to improve the performance. We developed a single-model MQA method for protein structures based on 3DCNN using sequence profile-based features, namely, P3CMQA. Performance evaluation using a CASP13 dataset showed that profile-based features improved the assessment performance, and the proposed method was better than currently available single-model MQA methods, including the previous 3DCNN-based method. We also implemented a web-interface of the method to make it more user-friendly.

Protein sequence‐to‐structure learning: Is this the end(‐to‐end revolution)?

“…network with additional input features including MSA profile, predicted secondary structure, and solvent accessibility. 105 Finally, DeepAccNet showed remarkable performance in the CASP14 refinement category. This architecture extends Ornate by adding 1D and 2D input features coming from sequence and Rosetta energy terms 63 (Table 3).…”

“…Sato-3DCNN by Sato and Ishida [104] used an idea similar to that of Ornate with oriented local frames but did not automatically learn the atom type embeddings. 3DCNN_prof (or P3CMQA) extended this network with additional input features including MSA pro-file, predicted secondary structure, and solvent accessibility [105]. Finally, DeepAccNet showed remarkable performance in the CASP14 refinement category.…”

“…A non-invariant 3D CNN Ornate [103] A local frame-based 3D CNN model with learned atom embeddings Sato-3DCNN [104] A local frame-based 3D CNN 3DCNN_prof (P3CMQA) [105] Extends Sato-3DCNN with predicted features and PSSMs SE(3)-3DCNN [101] An invariant 3D CNN based on [102] trained for protein complexes iPhord & DeepMUSICS 3D CNNs TopQA [125] 3D CNN with explicit rotations and automatic scaling to fit into a unit cube DeepAccNet [61] Extends Ornate with 1D and 2D geometrical features to predict per-residue model accuracy and also inter-residue distance signed error Graph representations Graph-QA [106] GCN with representation learning, explicit modeling of both sequential and 3D structure, geometric invariance, and computational efficiency S-GCN [107] Molecular-graph-based method where angular information is accounted for using spherical harmonics GQArank GCN with many features, including PSSM and predicted geometrical properties DeepML A classical GCN LAW 5-layer GCN followed by a 3-layer 1D CNN Tessellations, 2D manifolds, and point clouds VoroCNN [117] A CNN built on a hierarchical 3D Voronoi tessellation of a protein molecule VoroMQA-dark [118] A CNN-based extension of VoroMQA [119] EDN [111] A point-cloud representation of the atomic structure combined with rotationequivariant, hierarchical convolutions…”

Protein sequence-to-structure learning: Is this the end(-to-end revolution)?

“…Sato-3DCNN by Sato and Ishida [88] used an idea similar to that of Ornate with oriented local frames but did not automatically learn the atom type embeddings. 3DCNN_prof (or P3CMQA) extended this network with additional input features including MSA profile, predicted secondary structure, and solvent accessibility [89]. Finally, DeepAccNet showed remarkable performance in the CASP14 refinement category.…”

“…A non-invariant 3D CNN Ornate [87] A local frame-based 3D CNN model with learned atom embeddings Sato-3DCNN [88] A local frame-based 3D CNN 3DCNN_prof (P3CMQA) [89] Extends Sato-3DCNN with predicted features and PSSMs SE(3)-3DCNN [85] An invariant 3D CNN based on [86] trained for protein complexes iPhord & DeepMUSICS 3D CNNs TopQA [109] 3D CNN with explicit rotations and automatic scaling to fit into a unit cube DeepAccNet [44] Extends Ornate with 1D and 2D geometrical features to predict per-residue model accuracy and also inter-residue distance signed error Graph representations Graph-QA [90] GCN with representation learning, explicit modeling of both sequential and 3D structure, geometric invariance, and computational efficiency S-GCN [91] Molecular-graph-based method where angular information is accounted for using spherical harmonics GQArank GCN with many features, including PSSM and predicted geometrical properties DeepML A classical GCN LAW 5-layer GCN followed by a 3-layer 1D CNN Tessellations, 2D manifolds, and point clouds VoroCNN [101] A CNN built on a hierarchical 3D Voronoi tessellation of a protein molecule VoroMQA-dark [102] A CNN-based extension of VoroMQA [103] EDN [95] A point-cloud representation of the atomic structure combined with rotationequivariant, hierarchical convolutions tween distant amino acids in the sequence. When treating a raw MSA as a 2D image, the order of the sequences will also have an influence while this order may somewhat be arbitrary.…”