S4Pred

Publication     Copy citation     Github repository   

S4PRED predicts the secondary structure of a single protein sequence in the absence of homology information, achieving a Q3 score of 75.3% on the standard CB513 test set, taking only single sequences as input. Although they don't perform as their homology-based counterparts, single-sequence methods are not constrained by the requirement for evolutionary information. More accurate single-sequence approaches have the potential to improve structural modelling across the vast majority of sequence space, especially in areas of great scientific interest like viral proteins, the “dark proteome”, and de novo protein design. Academic users can download the model here.

DMPfold2

Publication     Copy citation     Github repository       

DMPfold2 is an ultrafast end-to-end deep learning method that predicts tertiary structure using only a multiple sequence alignment (MSA) as input. The network makes use of just three recurrent networks and a stack of residual convolutional layers, making the predictor very fast to run, and easy to install and use. Our approach constructs a directly learned representation of the sequences in an MSA, starting from a one-hot encoding of the sequences. When supplemented with an approximate precision matrix, the learned representation can be used to produce structural models of comparable or greater accuracy as compared to our original DMPfold method, while requiring less than a second to produce a typical model.

DMPfold

Publication     Copy citation     Github repository   Understanding the webserver results    

DMPfold uses deep learning to predict inter-atomic distance bounds, the main chain hydrogen bond network, and torsion angles, which it uses to build models in an iterative fashion. DMPfold produces more accurate models than two popular methods for a test set of CASP12 domains, and works just as well for transmembrane proteins. It produced confident models for 25% of all Pfam domains without known structures and models for 16% of human proteome UniProt entries without structures and generates accurate models with fewer than 100 sequences in some cases. Using the DMPfold method we have modelled all but ten of the proteins without templates in the JCVI-syn3.0 minimal genome. The paper is available here . A broader discussion of the use of deep learning in structural prediction is available here .

PASS: Profile Augmentation of Single Sequences

Publication     Github repository   

Profile Augmentation of Single Sequences (PASS) is a simple but powerful framework for accurately modelling single orphan protein sequences in the absence of homology information. S4PRED uses PASS to achieve an unprecedented Q3 score of 75.3% on the standard CB513 test. PASS provides a blueprint for the development of a new generation of predictive methods, advancing our ability to model individual protein sequences.

PSIPRED: Predict Secondary Structure

Publication     Copy citation     Github repository   Understanding the webserver results    

PSIPRED is a  simple and accurate secondary structure prediction method, incorporating two feed-forward neural networks which perform an analysis on output obtained from PSI-BLAST (Position Specific Iterated - BLAST). Using a very stringent cross validation method to evaluate the method's performance, PSIPRED 3.2 achieves an average Q₃ score of 81.6%. Predictions produced by PSIPRED were also submitted to the CASP4 evaluation and assessed during the CASP4 meeting, which took place in December 2000 at Asilomar. PSIPRED 2.0 achieved an average Q₃ score of 80.6% across all 40 submitted target domains with no obvious sequence similarity to structures present in PDB, which ranked PSIPRED top out of 20 evaluated methods (an earlier version of PSIPRED was also ranked top in CASP3 held in 1998). It is important to realise, however, that due to the small sample sizes, the results from CASP are not statistically significant, although they do give a rough guide as to the current "state of the art". For a more reliable evaluation, the EVA web site at Columbia University provides a continuous evaluation. NOTE that at the time of writing, the EVA site is no longer being updated. Downloads: The PSIPRED V3.2 software can be downloaded from HERE. Please note that you should read the license terms given in the README file if you wish to incorporate PSIPRED in another program or Web server. Older releases of PSIPRED can be downloaded here HERE.

DISOPRED3: Protein intrinsic disorder prediction

Publication     Copy citation     Github repository   Understanding the webserver results     Tutorial    

DISOPRED3 represents the latest release of our successful machine-learning based approach to the detection of intrinsically disordered regions. The method was originally trained on evolutionarily conserved sequence features of disordered regions from missing residues in high-resolution X-ray structures. DISOPRED2 mainly addressed the marked class imbalance between ordered and disordered amino acids as well as the different sequence patterns associated with terminal and internal disordered regions using SVMs. DISOPRED3 extends the previous architecture with two independent predictors of intrinsic disorder - a neural network and a nearest neighbor classifier - which were trained to identify long intrinsically disordered regions using data from the PDB and DisProt databases. The intermediate results are integrated by an additional neural network. DISOPRED3 was blindly tested and compared during the ninth and tenth rounds of the world-wide CASP experiments, where it was found to achieve high levels of specificity (about 99%) and therefore precision (about 75%). Indeed, the official assessment teams ranked DISOPRED3 at the top or near the top across a number of tests and evaluation measures. To provide insights into the biological roles of proteins, DISOPRED3 also predicts protein binding sites within disordered regions using a SVM that examines patterns of evolutionary sequence conservation, positional information and amino acid composition of putative disordered regions. Using a stringent test set, DISOPRED3 predictions were found to improve over existing methods, achieving approximately 20% precision and 30% recall. These results highlight the need for additional efforts in the area.

Merizo

Preprint     Copy citation     Github repository   

Merizo is a deep learning-based domain segmentation method that is trained on CATH annotations and can operate directly on PDB and AlphaFold2 structures. Notably, it makes use of AlphaFold2's Invariant Point Attention module in an input-reversed manner, using it to read a folded structure into a latent representation, where each residue embedding is then clustered into domains in a bottom-up manner. The network learns to cluster residues into domains based on an affinity learning principle, whereby the latent representation of pairs of residues are encouraged or discouraged to be similar based on whether or not the residues belong to the same domain instance.

pGenTHREADER and pDomTHREADER: Fold Recognition

Publication     Copy citation     Github repository     Understanding the webserver results    

pGenTHREADER and pDomTHREADER are fast, sensitive methods for predicting protein fold structures by comparison with protein templates of known structure. Their speed makes them suitable for full proteome annotations. pDomTHREADER is an accurate and sensitive superfamily discrimination, combining information from both sequence and structure to produce highly accurate domain alignments. The method employs the same underlying threading algorithm as pGenTHREADER, however it aligns sequences to a domain-based template library rather than a chain-based template library. The use of smaller regions of structure for templates means that different features of the alignments are required for optimal scoring. The final prediction score results from an SVM trained on a combination of 5 different feature inputs; template coverage, alignment score, template length, solvation and pairwise potentials. Compared with other superfamily discrimination methods using Hidden Markov Models and PSI-BLAST profile alignments, we found that pDomTHREADER provided higher coverage on the CATH S35 superfamilies. Additionally, pDomTHREADER produced more accurate alignments that can be used to better predict domain boundaries. Please note that the pDomTHREADER method is tuned for performance in fine superfamily discrimination, for fold recognition problems or structural annotation of very distant sequences, pGenTHREADER should be used.

DomPred & DOMSSEA : Domain Boundary Prediction

Publication     Copy citation     Github repository     Tutorial    

DomPred is a protein structural domain boundary predictor. The DomPred process runs 2 independent protein domain predictors; DomPred and DOMSSEA. The DomPred process begins by using PSI-BLAST to match a database of Pfam-A domains to the query sequence, where not clear domains can be match it then proceeds to search the nrdb90 sequence database with PSI-BLAST. The final prediction is procduced by analysing the locations of all the N and C boundaries for each hit. For the DOMSSEA process predicted secondary structure patterns in the query sequence are matched to a library of SCOP domain secondary structure patterns.

FFPRED: Protein Function Prediction

Publication     Copy citation     Github repository    Understanding the webserver results    

FFPred predicts Gene Ontology terms for human protein chains, when homology with characterized proteins can provide little aid. Predictions are made by scanning the input sequences against an array of Support Vector Machines (SVMs). Each of these SVMs examines the relationship between protein function and biophysical attributes describing secondary structure, transmembrane helices, intrinsically disordered regions, signal peptides and other motifs. The latest version FFPred3 features a larger SVM library that extends its coverage to the cellular component sub-ontology for the first time.

FFPRED-GAN: Generative Adversarial Networks

Publication     Github repository       

FFPRED-GAN provides a data augmentation method, enabling the generation of high-quality synthetic samples to improve the accuracy of function prediction methods. It uses a novel generative adversarial network to accurately learn the high-dimensional distributions of protein sequence-based biophysical features and generates high-quality synthetic protein feature samples, which can be used to augment the original training protein feature samples.

String2go

Publication     Github repository       

String2go predicts protein function using Protein-protein interaction network data. It works by assuming that interacting proteins tend to have similar functions, and uses deep maxout neural networks to learn functional representations simultaneously encoding both protein-protein interactions and functional predictive information.

DeepCov

Publication     Copy citation     Github repository   

DeepCov is a deep neural network-based method for predicting residue-residue contacts in protein sequences. It uses fully convolutional neural networks to analyze amino-acid pair frequency or covariance data derived directly from sequence alignments, without using global statistical methods. The method achieves state-of-the-art precision in contact prediction, matching or exceeding the performance of other methods such as CCMpred and MetaPSICOV2. DeepCov's performance is particularly notable in shallow sequence alignments (fewer than 160 effective sequences), indicating that it could be useful for predicting contacts in smaller sequence families.

HSPRED: Protein-Protein Interaction Characterization

Publication     Copy citation     Github repository        Tutorial

Protein-protein interactions are critically dependent on just a few ‘hot spot’ residues at the interface. These hot spots make a dominant contribution to the free energy of binding and they can disrupt the interaction if mutated to alanine. HSPred is a support vector machine(SVM)-based method which predicts hot spot residues, given the PDB structure of a complex.

DMP: DeepMetaPSICOV for Inter-Residue Contact Prediction

Publication     Github repository       

DeepMetaPSICOV (abbreviated DMP) is an interresidue contact predictor based on a deep, fully convolutional residual network and a large input feature set. Data was augmented by including partial loop deletions, simulated by probabilistically removing rows and columns in the input tensors and contact maps; linearly interpolating input tensors generated using different alignments; and by flipping input feature tensors and contact maps by 180°, corresponding to a reversal of the chain direction.

Metsite: Metal Binding Site Prediction

Publication     Copy citation     Github repository   

Metsite predicts residues forming the metal-binding site in super-families by combining sequence profile and structural information. It uses neural networks to predict six commonly occurring metal ion sites: Ca2+, Cu2+, Fe3+, Mg2+, Mn2+ and Zn2.

Protein-VAE: Design of Metalloproteins & Novel Protein Folds

Publication     Github repository       

Protein-VAE is a deep generative model that produces unsupervised design of metal binding sites in pre-existing proteins, and proteins that fold into a novel topology. The approach uses conditional variational auto-encoders to learn a compressed and compact latent representation of the training protein sequences. It aims to reduce the initial search space by learning a distribution that narrows down the space of all possible proteins to a smaller space of proteins that are more likely to be natively ordered.

DARK: Learning Deep Generative Models for De Novo Protein Design

Publication     Github repository       

DARK is a tool for designing novel protein sequences that are not found in nature, using a deep generative sequence model. It rapidly samples diverse and ordered structures for new protein sequences and is built using a framework that trains the model on synthetic sequences. The resulting model outperforms other sampling-based approaches and can generate high-quality candidates for de novo protein design problems. Overall, DARK represents a step towards automating and improving the process of protein design, which has applications in the development of novel therapeutics and materials.

MEMSAT & MEMSATSVM

Publication     Copy citation     Github repository        Downloads       

MEMSAT V3 is a widely used all-helical membrane protein prediction method based on the MEMSAT method. The MEMSAT algorithms calculate the most probable length, location and topological orientation for each transmembrane segment. They make use of scores compiled from membrane protein data about the propensity of each amino acid to be in one of five states (inside loop, outside loop, inside helix end, helix middle and outside helix end) and a dynamic programming algorithm to search through all possible topological models by a process of expectation maximization. The method was benchmarked on a test set of transmembrane proteins of known topology. From sequence data MEMSAT was estimated to have an accuracy of over 78% at predicting the structure of all-helical transmembrane proteins and the location of their constituent helical elements within a membrane.

MEMSATSVM is highly accurate predictor of transmembrane helix topology. It is capable to discriminating signal peptides and identifying the cytosolic and extra-cellular loops. The original paper is available here.

MEMPACK : Transmembrane helix contact prediction

Publication     Copy citation     Github repository       

MEMPACK is a membrane helix packing predictor. The process leverages MEMSATSVM predictions to predict possible inter-helix interactions. The final step a helix packing is produced that orients the helices such that the greatest number of predicted interactions face one another.

Memembed

Publication     Copy citation     Github repository        Downloads       

Memembed developed in 2013 provides the code to predict knowledge-based membrane potentials, calculated from a statistical analysis of transmembrane protein structures, coupled with a genetic and direct search algorithms, and demonstrated its use in positioning proteins in membranes, estimating membrane thickness, refinement of transmembrane protein models and in decoy discrimination.

The future of the PSIPRED workbench

Website     Key collaborator: Dr Daniel Buchan     Publication     Copy citation    

Development of the PSIPRED workbench is an ongoing collaboration with Dr Daniel Buchan of Goldsmiths, Univerity of London.

Genome3D

Website     Publication     Copy citation    

Genome3D is a freely available resource that provides consensus structural annotations for representative protein sequences taken from a selection of model organisms. The group provides 3D Models based on DomSerf and domain predictions based on pDomTHREADER.

CATH

Website     Lead researcher: Professor Christine Orengo     Publication     Copy citation    

CATH is a classification of protein structures downloaded from the Protein Data Bank. Protein domains into superfamilies when there is sufficient evidence they have diverged from a common ancestor.

ELIXIR

Website     Publication     Copy citation    

ELIXIR coordinates and develops life science resources across Europe so that researchers can more easily find, analyse and share data, exchange expertise, and implement best practices. This makes it possible for them to gain greater insights into how living organisms work. PSIPRED is an ELIXIR-UK service.

ISMB: Institute of Structural and Molecular Biology

Website     Key collaborators Dr Renós Savva and Dr Cen Wan    

ISMB was founded in 2003 to promote multi-disciplinary research at the interface of structural-, computational- and chemical biology at UCL and Birkbeck. The Jones lab is an active member of ISMB, and collaborates with Dr Renós Savva on protein design, and Dr Cen Wan on functional prediction.