Molecular recognition features (MoRFs) are small (10-70 residues) intrinsically disordered regions in proteins that undergo a disorder-to-order transition upon binding to their partners. MoRFs are implicated in protein-protein interactions, which serve as the initial step in molecular recognition. MoRFs are disordered prior to binding to their partners, whereas they form a common 3D structure after interacting with their partners.[1][2] As MoRF regions tend to resemble disordered proteins with some characteristics of ordered proteins,[2] they can be classified as existing in an extended semi-disordered state.[3]
Categorization
MoRFs can be separated in 4 categories according to the shape they form once bound to their partners.[2]
complex-MoRFs (combination of the above categories)
MoRF predictors
Determining protein structures experimentally is a very time-consuming and expensive process. Therefore, recent years have seen a focus on computational methods for predicting protein structure and structural characteristics. Some aspects of protein structure, such as secondary structure and intrinsic disorder, have benefited greatly from applications of deep learning on an abundance of annotated data. However, computational prediction of MoRF regions remains a challenging task due to the limited availability of annotated data and the rarity of the MoRF class itself.[4] Most current methods have been trained and benchmarked on the sets released by the authors of MoRFPred[5] in 2012, as well as another set released by the authors of MoRFChibi[6][7][8] based on experimentally-annotated MoRF data. The table below details some methods available as of 2019 for MoRF prediction (related problems are also touched upon).[9]
Protein Intrinsic Disorder and Protein Binding Sites
Multistage component prediction (utilizing neural network, Support Vector Machine, and K-nearest neighbour models) for protein disorder prediction. Also uses an additional Support Vector Machine to interpolate binding regions from the disorder predictions.
Multiple logistic regression models based on predicted disorder, amino acid properties, and sequence composition. The result is aligned with transferred annotations from a functionally-annotated database.
Hierarchy of different in-house MoRF prediction models:
MoRFchibi: Utilizes Bayes rule to combine the outcomes of two support Vector Machine modules using amino acid composition (Sigmoid kernel) and sequence similarity (RBF kernel).
MoRFchibi_light: Utilizes Bayes rule to combine MoRFchibi and disorder prediction hierarchically.
MoRFchibi_web: Utilizes Bayes rule to combine MoRFchibi, disorder prediction and PSSM (MSA) hierarchically.
Neural network for predicting both long and short disordered regions. Semi-disorder can be linearly interpolated from its predicted disorder probabilities (0.4<=P(D)<=0.7).
Bidirectional Long Short-Term Memory network for predicting intrinsic disorder. Semi-disordered regions can be linearly interpolated from its predicted disorder probabilities (0.28<=P(D)<=0.69).
Transfer learning from the large disorder prediction tool SPOT-Disorder2[21] (which itself utilizes an ensemble of Bidirectional Long Short-Term Memory networks and Inception ResNets).
^Sharma R, Bayarjargal M, Tsunoda T, Patil A, Sharma A (January 2018). "MoRFPred-plus: Computational Identification of MoRFs in Protein Sequences using Physicochemical Properties and HMM profiles". Journal of Theoretical Biology. 437: 9–16. Bibcode:2018JThBi.437....9S. doi:10.1016/j.jtbi.2017.10.015. hdl:10072/376330. PMID29042212.
^Zhang T, Faraggi E, Li Z, Zhou Y (2017). "Intrinsic Disorder and Semi-disorder Prediction by SPINE-D". In Zhou Y, Kloczkowski A, FaraggiR, Yang Y (eds.). Prediction of Protein Secondary Structure. Methods in Molecular Biology (vol. 1484). Vol. 1484. New York: Springer. pp. 159–174. doi:10.1007/978-1-4939-6406-2_12. ISBN9781493964048. PMID27787826.
