CASP14 | Nov 30 – 4, 2020 | Virtual Symposium | Viewing Hall

Short Abstract: The recent success of Deep Learning (DL) methods in structural bioinformatics has improved the quality of predicted structures significantly. However, the vast majority of state-of-the-art predictors still rely on evolutionary information captured by multiple sequence alignments (MSAs), making structures of proteins with few evolutionary relatives tough to predict. Additionally, creating high-quality MSAs is not trivial: the parameters for the alignment process need to be chosen on an individual basis in order to add enough, yet diverse sequences. This is done with the goal of obtaining a rich set of sequences that model structural constraints, whilst avoiding the inclusion of sequences with diverging structure. We present EMBER (EMBedding-based inter-residue distance predictor), a novel DL method used to predict inter-residue distance maps. EMBER enriches the traditional MSA based input by sequence embeddings, represented by the hidden states of Natural Language Processing (NLP) systems such as BERT1 and ELMo2 trained on protein sequence sets. We use deep dilated residual convolutional networks with many layers, similar to Alphafold3 and ProSPr4. We also used the previously successful approach of training on crops of 64x64 residues instead of full samples. This enabled data augmentation and allowed more efficient mini-batching. Evolutionary information from the MSA is represented as plmDCA5 parameters obtained by CCMpred6. In addition to evolutionary information and embeddings, inputs include: the relative position of the crop w.r.t. the overall sequence, the normalized length of the sample and the log-normalized number of effective sequences as inputs. The latter is primarily intended to allow the model to learn how to weigh embeddings vs. DCA constraints based on alignment quality. Our training and validation sets were based on ProteinNet127, but we also added the available samples of the free-modeling category from CASP13 for further validation. Similar to other recent methods, we favored the prediction of distance probability distributions instead of binary contacts by using 42 bins representing distance intervals between 2 and 22 Angstrom. Since EMBER was developed during CASP14, we submitted predictions from multiple models, which were based on slightly different input combinations. For some of the samples with very sparse MSAs, we submitted predictions from models trained exclusively on embeddings and without evolutionary information.

To ask a question to the presenter click here

Short Abstract: Recent advances in machine learning, bioinformatics and the understanding of the folding problem have enabled efficient predictions of protein structures with moderate accuracy, even for targets when there is little information from templates. All-atom molecular dynamics simulations provide a route to refine such predicted structures, but unguided atomistic simulations, even when lengthy in time, often fail to eliminate incorrect structural features that would allow the structure to become energetically favorable owing to the necessity of making large scale motions and overcoming energy barriers for side chain repacking. In this study, we show that localizing frustration at atomic resolution by examining the statistics of the energetic changes that occur when the local environment of a site is changed allows one to identify the most likely locations of incorrect contact. The global statistics of atomic resolution frustration in structures that have been predicted using various algorithms provide strong indicators of the structural quality when feeded over a database of 20 targets from previous CASP experiments. Residues that are more correctly located turn out to be more minimally frustrated than more poorly positioned sites. These observations provide a diagnosis of both global and local quality of predicted structures, and they can be used as guidance in all-atom refinement simulations of the 20 targets. Refinement simulations guided by frustration turn out to be quite efficient and significantly improve the quality other structures.. d

To ask a question to the presenter click here

Keywords:

Poster:

Short Abstract: We present a novel deep learning approach to assess the quality of protein models. Our neural network builds on a point-based representation of the atomic structure and rotation-equivariant convolutions at different levels of structural resolution. The convolution filters are constructed based on a truncated series of spherical harmonics and the network hierarchically aggregates information at different levels --- from all atoms over alpha carbons to one global score for the entire protein model. These combined aspects allow the network to recognize structural motifs at different scales and independent of spatial orientation. Starting with the coordinates and chemical element type of each atom, the network learns to predict the global quality metric GDT_TS end-to-end. Our method does not use physics-inspired energy terms and does not rely on the availability of additional information, such as sequence alignments of multiple proteins. We report state-of-the-art results in scoring protein models submitted to recent rounds of CASP. We invite you to learn more by visiting our poster and watching our video presentation.

Video not uploaded

To ask a question to the presenter click here

Keywords:

Poster:

Short Abstract: Here we describe our participation in the CASP14 Assembly category, as part of the 4th common CASP-CAPRI Assembly Prediction challenge (CAPRI Round 50). We have participated as human predictors, human scorers, and server scorers, in all the 18 proposed targets, consisting in four hetero-dimers (A1B1), six homo-dimers (A2), two homo-trimers (A3), two homo-tetramers (A4), one hetero-nonamer (A3B3C3), one homo-20mer (A20), one hetero-27mer (A6B3C12D6), and one homo-240mer (A240). However, for of these targets (T171/T1063, T172/H1066, T173/H1069 and T181/H1103) were later cancelled due to lack of available structure. METHODS For each assembly, the models of the individual subunits were taken from the ZHANG, RaptorX, and QUARK CASP-hosted servers (only the best prediction for each server was used). In cases with no available models at the CASP-hosted servers, we modelled them with MODELLERv9.19. Using the available structural models of the individual subunits as above described, we modelled all or some binary interactions in the assembly by ab initio docking. As human group we applied our pyDock [1] docking and scoring pipeline, in which we used FTDock and ZDOCK 2.1 to generate 12,000 rigid-body docking poses, which scored by pyDock. We also participated with our pyDockWEB server [2]. In homo-oligomers, docking poses not satisfying the expected symmetry (e.g. C2 for homo-dimers, C3 for homo-timers, etc.) were removed. Additionally, we checked if there were available templates for all or part of the assembly interfaces, using BLAST as well as the released predictions from the ZHANG, QUARK, RaptorX, MULTICOM-CONSTRUCTand ROSETTA CASP-hosted servers. Finally, all the generated models (either ab initio or template-based) were scored with pyDock. The number of available templates and their reliability determined the percentage of template-based complex models included in the top 5 and 10 submitted models. In the scorers experiment, we applied pyDock scoring and used the same criteria to rank the docking models as in predictors. As human scorers we introduced more human intervention than as server scorers, i.e., removing loops with non-realistic conformations, and re-scoring some of these models afterwards. RESULTS While the evaluation of our models are not yet available at the time of writing this abstract, we will describe here the specific details of our submissions (not considering the cancelled targets). We submitted models generated only by ab initio docking in those targets for which we could not find available templates (T169/T1054, T174/T1070, T178/T1083, and T179/T1087). But for the majority of targets, we could find potentially suitable templates for all or some of the predicted interfaces, so our submissions combined template-based and combined models. Ab initio docking was favored in targets T164/T1032 and T176/T1078, while template-based modeling was favored in targets T166/H1045, T167/T1050, and T168/T1052. On the other side, in target T180/T1099, consisting in the assembly of a virus capsid with icosahedral symmetry, only template-based modeling was used. In the remaining targets, in order to build the full assembly we combined template-based docking for some interfaces and ab initio docking for the other ones, using symmetry restraints when needed. Finally, target T170/H1060 was a challenging hetero-27mer, in which we applied an ad-hoc modeling procedure. This assembly was formed by three rings with different composition and stoichiometry, which were independently modelled by combining ab initio docking and template-based modeling. The final assembly of the modelled rings was done with the help of ab initio docking, selecting only models in which the symmetry axes of the rings were aligned. The same criteria was used in the scorers experiment. [1] Cheng,T.M.-K., Blundell.T.L. & Fernandez-Recio,J. (2007) pyDock: electrostatics and desolvation for effective rigid-body protein-protein docking. Proteins. 68, 503-515. [2] Jimenez-Garcia,B., Pons,C. & Fernandez-Recio,J. (2013) pyDockWEB: a web server for rigid-body protein-protein docking using electrostatics and desolvation scoring. Bioinformatics. 29, 1698-1699.

Video not uploaded

To ask a question to the presenter click here

To ask a question to the presenter click here

Short Abstract: We present our Particle Swarm Optimisation (PSO) pipelines for prediction of ternary and quaternary structures of proteins. Both pipelines are using the same PSO engine to optimise relative position and orientation of a ligand relative to a receptor for SwarmDock and the backbone dihedral angles of linkers between protein fragments for SwarmLoop. We have recently developed a new version of SwarmLoop to generate models from primary and secondary information. We first run PSIPRED to predict the secondary structure of the protein. We then rank the linkers according to their lengths and run the SwarmLoop protocol to assemble flanking blocks of secondary structure elements iteratively until the protein is folded. We tested the new ab initio protocol using a short 70 residues protein (PDB id 1WHZ), from PSIPRED file, we have been able to generate a model with 56 points GDT-TS.

To ask a question to the presenter click here

Keywords:

Poster:

Short Abstract: Protein inter-residue contact and distance prediction are two key intermediate steps essential to accurate protein structure prediction. Distance prediction comes in two forms: real-valued distances and ‘binned’ distograms, which are a more finely grained variant of the binary contact prediction problem. The latter has been introduced as a new challenge in the 14th Critical Assessment of Techniques for Protein Structure Prediction (CASP14) 2020 experiment. Despite the recent proliferation of methods for predicting distances, few methods exist for evaluating these predictions. Currently only numerical metrics, which evaluate the entire prediction at once, are used. These give no insight into the structural details of a prediction. For this reason, new methods and tools are needed. We have developed a web server for evaluating predicted inter-residue distances. Our server, DISTEVAL, accepts predicted contacts, distances, and a true structure as optional inputs to generate informative heatmaps, chord diagrams, and 3D models. All of these outputs facilitate visual and qualitative assessment. The server also evaluates predictions using other metrics such as mean absolute error, root mean squared error, LDDT score, and contact precision. The visualizations generated by DISTEVAL complement each other and collectively serve as a powerful tool for both quantitative and qualitative assessments of predicted contacts and distances, even in the absence of a true 3D structure.

To ask a question to the presenter click here

Keywords:

Poster:

Short Abstract: This work introduces a real-valued distance prediction method REALDIST. Using a representative set of 43 thousand protein chains, a variant of deep ResNet is trained to predict real-valued distance maps. The contacts derived from the real-valued distance maps predicted by this method, on the most difficult CASP13 free-modeling protein datasets, demonstrate a long-range top-L precision of 52%, which is 17% higher than the top CASP13 predictor Raptor-X and slightly higher than the more recent trRosetta method. Similar improvements are observed on the CAMEO `hard' and `very hard' datasets. Three-dimensional (3D) structure prediction guided by real-valued distances reveals that for short proteins the mean accuracy of the 3D models is slightly higher than the top human predictor AlphaFold and server predictor Quark in the CASP13 competition.

To ask a question to the presenter click here

Keywords:

Poster:

Short Abstract: CUTSP CASP14 submissions were all generated by morphing predicted structures and ranking the results. We used the same methods for all predictions as well as protein docking for complexes. We first generate multiple sequence alignments (MSAs) using HHblits with UniRef. Next, we use both supervised and semi-supervised approaches based on distance and torsion angle representations for predicting diverse protein structures. We then morph between these structures taking into account the energy of the conformation. The morphing is non-linear and allows to bypass high energy conformation barriers. We superimpose the structures onto a base structure and select the top candidates. We rank the morphed structures using a deep neural network trained to predict quality based on previous CASPs and a graph neural network predicting quality of full-atom graph protein representations. We perform docking of proteins with multiple chains. First, we predict the conformation of each chain and then use rigid-body protein docking to generate a candidate set of complexes. Finally, we rank the complexes based on their energy score, and select the top candidates.

Video not uploaded

To ask a question to the presenter click here

Keywords:

Poster:

Short Abstract: Our MULTICOM-AI protein complex structure predictor uses an ab initio deep learning-based intrachain contact prediction tool (DNCON2) as well as a template-based prediction (TBP) method to predict interchain residue-residue contacts for protein complex targets in CASP14 and CAPRI50. We name this predictor DNCON2_Inter. Interchain residue pairs are considered interchain contacts if the Euclidean distance between the heavy atoms of the residues is less than or equal to 6.0 Å. These contacts are then used to reconstruct the final structure of the multimeric complex using our Con_Complex tool which uses the distance-geometry protocol of Crystallography & NMR System (CNS). Upon comparing the dimeric contacts of T1032 with respect to the dimeric contacts of the PDB 6N64, both DNCON2_Inter and TBP method performed exceptionally well.

To ask a question to the presenter click here

Keywords:

Poster:

Short Abstract: Many algorithms and pipelines have been developed that go from sequence to structure, however, a full end-to-end model remains a challenge. Here we present our initial efforts towards a fully differentiable set of modules that go from sequences to distance/dihedral matrices to 3D coordinates. We modified the last layers of the TrRosetta NN (neural network) model1 to return a full alpha-carbon distance matrix and backbone dihedrals. A shortest path searching layer was then applied to refine distance matrices. Finally, to recover the 3D coordinates, we iteratively placed alpha-carbon atoms one at a time conditioned on both distances and dihedrals. The NN models were trained and validated on the TrRosetta benchmark set. This approach returns structures of comparable quality to those generated by the TrRosetta protocol that requires a very expensive minimization step. Going forward, we think these modules can be easily incorporated into any deep learning protocol for a full end-to-end training. 1. Yang, J. et al. Improved protein structure prediction using predicted interresidue orientations. Proceedings of the National Academy of Sciences 117, 1496-1503 (2020).

Video not uploaded

To ask a question to the presenter click here

Keywords:

Poster:

Short Abstract: Our team has participated in two sections of the competition; quality assessment of protein models (QA) and prediction of protein complexes or tertiary assemblies (TS). Overall, we participated in 78 QA and 28 TS prediction rounds. For QA we have relied on scores generated by our pre-trained 3DCNN model, adopting the protocols used in CASP13. Individual monomers were either directly modelled or obtained by ranking the predictions from CASP server stage 2 rounds. Protein assemblies were built in two main ways: (1) using a template-free approach based on the 3DCNN docking algorithm recently developed in our lab and (2) using template-based docking whenever templates were available. The 3DCNN docking algorithm was trained to predict complex three-dimensional representations from the atomic densities of a protein. Each protein is represented by 48 features, which are combined with those of the partner protein to produce a score that can be optimized by adjusting the relative position and orientation of the two proteins. Complexes of more than two proteins are assembled in a stepwise manner, one protein at a time. The CNN docking algorithm is trained on decoy conformations generated from the dataset of Huang and Zou and is tested on structures from the Protein-Protein Docking Benchmark Version 4.0.5. When good templates were found, docking poses were generated by aligning the proteins with TM-align and were scored by global score from 3DCNN-LQA algorithm. During the TS rounds, most of the protein assemblies were built using a combination of the 3DCNN docking algorithm and template-based methods, and the complexes built were ranked using the 3DCNN-LQA algorithm. Our neural network-based approach enabled us to deal with assemblies of any size and stoichiometry. By interpreting the CNN docking score as an energy, we could also estimate the binding affinity of some of the complexes.

Video not uploaded

To ask a question to the presenter click here

Keywords:

Poster:

Short Abstract: A novel version of the Recurrent Geometrical Network (RGN1) algorithm, which geometrically reasons over protein conformations, is used to predict protein structures. Two options are considered for inputs: (i) the raw amino acid sequence and position-specific scoring matrix (PSSM) of each protein and (ii) a context-based encoding of amino acid residues - AminoBert - derived strictly from raw amino acid sequences without making explicit use of any evolutionary information. Raw RGN structure predictions are subsequently refined using an energy-minimization protocol subject to dihedral constraints computed from family sequence alignments. Methods One-dimensional curves, in differential geometry, are described by the Frenet-Serret geometries (FSG). We use an improved version of the previously reported RGN1, which parameterized protein backbones (Ca atoms) using dihedral angles, that leverages the fact that protein backbones are intrinsically discrete one-dimensional curves. The improved version implements a transfer matrix formalism which enables reasoning over protein backbones using a discrete version of Frenet-Serret geometries (dFSG2). Inputs: dFSG-based RGNs are used with two different possible inputs to predict protein backbones: (i) Raw amino acid sequences and PSSMs, as previously described1 (ii) AminoBert: a reformulated version of the BERT language model3 is used to train a transformer4 over protein sequences to predict missing amino acids conditioned on the flanking sequence. Amino acid residues are thus mapped onto a higher-dimensional representation. Refinement: Raw structure predictions from dFSG-based RGNs, trained with sequence+PSSM (HMSCasper-PSSM) or AminoBERT representations (HMSCasper-Seq) are refined using a Rosetta-based protocol that first builds the remaining atoms and then alleviates steric clashes and fine-tunes folded domains. As an additional possibility, constraints coming from an orientogram populated with pairwise angular dependencies between residues, derived from a family sequence alignment and trRosetta, are imposed during energy minimization of the structure (HMSCasper-MSA). Training: For training we used (a) the dFSG-based RGN model trained on ProteinNet12 dataset (comprising UniParc + JGI metagenomes and PDB) with sequence+PSSM inputs for making predictions under the HMSCasper-PSSM group and (b) AminoBERT models trained on SCOPe datasets for making predictions under the HMSCasper-Seq group. Availability Source code for training dFSG-based RGN models as well as trained models, including PSSM and AminoBert based versions as used for the CASP14 experiment, will be available on GitHub. 1. AlQuraishi, M. End-to-End Differentiable Learning of Protein Structure. Cell Syst. (2019) doi:10.1016/j.cels.2019.03.006. 2. Hu, S., Lundgren, M. & Niemi, A. J. Discrete Frenet frame, inflection point solitons, and curve visualization with applications to folded proteins. Phys. Rev. E - Stat. Nonlinear, Soft Matter Phys. (2011) doi:10.1103/PhysRevE.83.061908. 3. Devlin, J., Chang, M. W., Lee, K. & Toutanova, K. BERT: Pre-training of deep bidirectional transformers for language understanding. in NAACL HLT 2019 - 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies - Proceedings of the Conference (2019). 4. Vaswani, A. et al. Attention is all you need. in Advances in Neural Information Processing Systems (2017).

To ask a question to the presenter click here

Keywords:

Poster:

Short Abstract: We participated in the assembly category. We predicted both homo- and hetero-oligomeric protein structures according to the oligomeric state in the CASP14 target list. Our modeling procedure was based on template-based docking method. We basically used CASP14 server models (Stage 2) as monomer models. We selected high quality monomer models using combined score of ProQ3, ProQ3D and VoroMQA. To find reliable oligomeric templates, we carried out two-step template search. When we could not obtain oligomeric templates, we used DOCKGROUND database as templates. The quality of oligomeric models were assessed by combined score of VoroMQA and SOAP-PP. Human intervention were made when necessary.

Video not uploaded

To ask a question to the presenter click here

Keywords:

Poster:

Short Abstract: Five automatic servers from Zhang group have participated in CASP14 for protein structure prediction, including Zhang-Server, QUARK, Zhang-CEthreader, Zhang-TBM and Zhang_Ab_Initio. Those pipelines were built upon the D-I-TASSER, D-QUARK, DEthreader and LOMETS3 algorithm, which in turn are based on I-TASSER, QUARK, CEthreader and LOMETS pipeline but with the following new modules: (a) FUpred, a novel domain boundary detection program, which predicts domain boundary by maximizing the number of intra-domain contacts, while minimizing the number of inter-domain contacts. (b) DeepMSA2, a new multiple sequence alignment generation tool, which is extended from DeepMSA pipeline by adding more searching stage with more metagenome sequence databases. The predicted contact map score is used for selecting the final MSA from all MSAs produced by DeepMSA2. (c) DeepPotential, a deep learning based method for predicting contact-maps, distance-maps, torsion angles and hydrogen-bond networks. It is trained by deep convolutional neural networks with a set of co-evolutionary features. (d) DEthreader, a template detection method which interplays distance Eigenvectors into CEthreader pipeline. It is utilized in Zhang-CEthreader server. (e) LOMETS3, an updated version of LOMETS threading pipeline for collecting templates for Zhang-Server, QUARK and Zhang-TBM. In LOMETS3, more contact-based threading methods have been added and the predicted contacts from DeepPotentail have been used to re-rank templates that picked up by profile-based threading methods. (f) D-I-TASSER and D-QUARK folding simulations, which are based on I-TASSER and QUARK folding simulations. The new folding simulations integrated restraints from templates and potentials from DeepPotental to guide the Replica Exchange Monte Carlo simulation.

Video not uploaded

To ask a question to the presenter click here

Keywords:

Poster:

Short Abstract: The goal of AP_1 is single chain protein structure scoring and combines our refinement protocol. AP_1 employs several characteristics, such as database search and structure retrieval without calculating pair-wise potentials and without building a fixed form potential. Methods The goal of AP_1 is to accurately score not only the topology of a protein structure, but also the side-chain positions of the high-accuracy template-based models. Our structure prediction pipeline consists of the following steps: 1. Five of the best models were picked using AP_1 from all submitted server models of CASP14. 2. Five of the best models were picked and used as the seed model for our refinement protocol. 3. Subsequently, five generated models were added to the seed models. 4. We applied AP_1 again to the above candidate models and selected the five best models to submit. In CASP14, we submitted 390 models for 78 TS regular targets.

Video not uploaded

To ask a question to the presenter click here