To Improve Protein Sequence Profile Prediction through Image Captioning on Pairwise Residue Distance Map

Abstract

Protein sequence profile prediction aims to generate multiple sequences from structural information to advance the protein design. Protein sequence profile can be computationally predicted by energy-based method or fragment-based methods. By integrating these methods with neural networks, our previous method, SPIN2 has achieved a sequence recovery rate of 34%. However, SPIN2 employed only one dimensional (1D) structural properties that are not sufficient to represent 3D structures. In this study, we represented 3D structures by 2D maps of pairwise residue distances. and developed a new method (SPROF) to predict protein sequence profile based on an image captioning learning frame. To our best knowledge, this is the first method to employ 2D distance map for predicting protein properties. SPROF achieved 39.8% in sequence recovery of residues on the independent test set, representing a 5.2% improvement over SPIN2. We also found the sequence recovery increased with the number of their neighbored residues in 3D structural space, indicating that our method can effectively learn long range information from the 2D distance map. Thus, such network architecture using 2D distance map is expected to be useful for other 3D structure-based applications, such as binding site prediction, protein function prediction, and protein interaction prediction.

Protein Design with Deep Learning

“…This last approach reaches state-of-the art prediction accuracy for a single residue given its environment, approximately 57% [56]. This metric should not be confused with the native sequence recovery rate, which compare the whole designed sequence with the native sequence and which is usually between 30 and 40% depending on the test set [49,54,63,65].…”

“…Neural networks designed to predict sequences often chose to output a tensor which can be directly interpreted as a Position-Specific Scoring Matrix [84]. In most cases, these scores are transformed to a probability distribution over amino acids using a "softmax" function at the final layer [50,[54][55][56][57][58]. Softmax combines exponentiation with normalization to transform arbitrary real vectors into vectors that can be interpreted as a multinoulli probability distribution.…”

“…Distance map can also be used directly to represent an input structure. SPROF extended SPIN2 by incorporating 3D information in the form of a contact map [54]. They processed the map as an image, and the corresponding sequence as a caption.…”

Structure-aware Protein Solubility Prediction From Sequence Through Graph Convolutional Network And Predicted Contact Map

“…However, these methods are mostly based on LSTM and CNN and didn't utilize spatial information of protein molecules. Though our recent studies indicated that protein structure could be well represented by the residue-pairwise distance matrix through CNN (Chen, et al, 2020;Zheng, et al, 2020), the contacted structural information is only implicitly included that can't fully utilize the relations between contacted residues.…”

DenseCPD: Improving the Accuracy of Neural-Network-Based Computational Protein Sequence Design with DenseNet

“…Chen and coworkers developed the SPROF method, which uses the two-dimensional map of the pairwise residue distance as the input, and reached an accuracy of 39.8%, representing a 5.2% improvement over that of SPIN2. 41 Yu et al used an interesting approach that translates amino acid sequences into musical compositions and trained a recurrent neural network to generate protein sequences. 42 Greener et al used a variational autoencoder to generate protein sequences conditioned on protein structures.…”