Structure-based Deep Fusion models were recently shown to outperform several physics- and machine learning-based protein-ligand binding affinity prediction methods. As part of a multi-institutional COVID-19 pandemic response, over 500 million small molecules were computationally screened against four protein structures from the novel coronavirus (SARS-CoV-2), which causes COVID-19. Three enhancements to Deep Fusion were made in order to evaluate more than 5 billion docked poses on SARS-CoV-2 protein targets. First, the Deep Fusion concept was refined by formulating the architecture as one, coherently backpropagated model (Coherent Fusion) to improve binding-affinity prediction accuracy. Secondly, the model was trained using a distributed, genetic hyper-parameter optimization. Finally, a scalable, high-throughput screening capability was developed to maximize the number of ligands evaluated and expedite the path to experimental evaluation. In this work, we present both the methods developed for machine learning-based high-throughput screening and results from using our computational pipeline to find SARS-CoV-2 inhibitors.

BACKGROUND: Most of the currently used methods for protein function prediction rely on sequence-based comparisons between a query protein and those for which a functional annotation is provided. A serious limitation of sequence similarity-based approaches for identifying residue conservation among proteins is the low confidence in assigning residue-residue correspondences among proteins when the level of sequence identity between the compared proteins is poor. Multiple sequence alignment methods are more satisfactory--still, they cannot provide reliable results at low levels of sequence identity. Our goal in the current work was to develop an algorithm that could help overcome these difficulties by facilitating the identification of structurally (and possibly functionally) relevant residue-residue correspondences between compared protein structures. RESULTS: Here we present StralSV (structure-alignment sequence variability), a new algorithm for detecting closely related structure fragments and quantifying residue frequency from tight local structure alignments. We apply StralSV in a study of the RNA-dependent RNA polymerase of poliovirus, and we demonstrate that the algorithm can be used to determine regions of the protein that are relatively unique, or that share structural similarity with proteins that would be considered distantly related. By quantifying residue frequencies among many residue-residue pairs extracted from local structural alignments, one can infer potential structural or functional importance of specific residues that are determined to be highly conserved or that deviate from a consensus. We further demonstrate that considerable detailed structural and phylogenetic information can be derived from StralSV analyses. CONCLUSIONS: StralSV is a new structure-based algorithm for identifying and aligning structure fragments that have similarity to a reference protein. StralSV analysis can be used to quantify residue-residue correspondences and identify residues that may be of particular structural or functional importance, as well as unusual or unexpected residues at a given sequence position. StralSV is provided as a web service at http://proteinmodel.org/AS2TS/STRALSV/.

During microbial evolution, genome rearrangement increases with increasing sequence divergence. If the relationship between synteny and sequence divergence can be modeled, gene clusters in genomes of distantly related organisms exhibiting anomalous synteny can be identified and used to infer functional conservation. We applied the phylogenetic pairwise comparison method to establish and model a strong correlation between synteny and sequence divergence in all 634 available Archaeal and Bacterial genomes from the NCBI database and four newly assembled genomes of uncultivated Archaea from an acid mine drainage (AMD) community. In parallel, we established and modeled the trend between synteny and functional relatedness in the 118 genomes available in the STRING database. By combining these models, we developed a gene functional annotation method that weights evolutionary distance to estimate the probability of functional associations of syntenous proteins between genome pairs. The method was applied to the hypothetical proteins and poorly annotated genes in newly assembled acid mine drainage Archaeal genomes to add or improve gene annotations. This is the first method to assign possible functions to poorly annotated genes through quantification of the probability of gene functional relationships based on synteny at a significant evolutionary distance, and has the potential for broad application.

Here we introduce a quantitative structure-driven computational domain-fusion method, which we used to predict the structures of proteins believed to be involved in regulation of the subtilin pathway in Bacillus subtilis, and used to predict a protein-protein complex formed by interaction between the proteins. Homology modeling of SpaK and SpaR yielded preliminary structural models based on a best template for SpaK comprising a dimer of a histidine kinase, and for SpaR a response regulator protein. Our LGA code was used to identify multi-domain proteins with structure homology to both modeled structures, yielding a set of domain-fusion templates then used to model a hypothetical SpaK/SpaR complex. The models were used to identify putative functional residues and residues at the protein-protein interface, and bioinformatics was used to compare functionally and structurally relevant residues in corresponding positions among proteins with structural homology to the templates. Models of the complex were evaluated in light of known properties of the functional residues within two-component systems involving His-Asp phosphorelays. Based on this analysis, a phosphotransferase complexed with a beryllofluoride was selected as the optimal template for modeling a SpaK/SpaR complex conformation. In vitro phosphorylation studies performed using wild type and site-directed SpaK mutant proteins validated the predictions derived from application of the structure-driven domain-fusion method: SpaK was phosphorylated in the presence of (32)P-ATP and the phosphate moiety was subsequently transferred to SpaR, supporting the hypothesis that SpaK and SpaR function as sensor and response regulator, respectively, in a two-component signal transduction system, and furthermore suggesting that the structure-driven domain-fusion approach correctly predicted a physical interaction between SpaK and SpaR. Our domain-fusion algorithm leverages quantitative structure information and provides a tool for generation of hypotheses regarding protein function, which can then be tested using empirical methods.

BACKGROUND: MannDB was created to meet a need for rapid, comprehensive automated protein sequence analyses to support selection of proteins suitable as targets for driving the development of reagents for pathogen or protein toxin detection. Because a large number of open-source tools were needed, it was necessary to produce a software system to scale the computations for whole-proteome analysis. Thus, we built a fully automated system for executing software tools and for storage, integration, and display of automated protein sequence analysis and annotation data. DESCRIPTION: MannDB is a relational database that organizes data resulting from fully automated, high-throughput protein-sequence analyses using open-source tools. Types of analyses provided include predictions of cleavage, chemical properties, classification, features, functional assignment, post-translational modifications, motifs, antigenicity, and secondary structure. Proteomes (lists of hypothetical and known proteins) are downloaded and parsed from Genbank and then inserted into MannDB, and annotations from SwissProt are downloaded when identifiers are found in the Genbank entry or when identical sequences are identified. Currently 36 open-source tools are run against MannDB protein sequences either on local systems or by means of batch submission to external servers. In addition, BLAST against protein entries in MvirDB, our database of microbial virulence factors, is performed. A web client browser enables viewing of computational results and downloaded annotations, and a query tool enables structured and free-text search capabilities. When available, links to external databases, including MvirDB, are provided. MannDB contains whole-proteome analyses for at least one representative organism from each category of biological threat organism listed by APHIS, CDC, HHS, NIAID, USDA, USFDA, and WHO. CONCLUSION: MannDB comprises a large number of genomes and comprehensive protein sequence analyses representing organisms listed as high-priority agents on the websites of several governmental organizations concerned with bio-terrorism. MannDB provides the user with a BLAST interface for comparison of native and non-native sequences and a query tool for conveniently selecting proteins of interest. In addition, the user has access to a web-based browser that compiles comprehensive and extensive reports. Access to MannDB is freely available at http://manndb.llnl.gov/.

MOTIVATION: Specific and sensitive ligand-based protein detection assays that employ antibodies or small molecules such as peptides, aptamers or other small molecules require that the corresponding surface region of the protein be accessible and that there be minimal cross-reactivity with non-target proteins. To reduce the time and cost of laboratory screening efforts for diagnostic reagents, we developed new methods for evaluating and selecting protein surface regions for ligand targeting. RESULTS: We devised combined structure- and sequence-based methods for identifying 3D epitopes and binding pockets on the surface of the A chain of ricin that are conserved with respect to a set of ricin A chains and unique with respect to other proteins. We (1) used structure alignment software to detect structural deviations and extracted from this analysis the residue-residue correspondence, (2) devised a method to compare corresponding residues across sets of ricin structures and structures of closely related proteins, (3) devised a sequence-based approach to determine residue infrequency in local sequence context and (4) modified a pocket-finding algorithm to identify surface crevices in close proximity to residues determined to be conserved/unique based on our structure- and sequence-based methods. In applying this combined informatics approach to ricin A, we identified a conserved/unique pocket in close proximity (but not overlapping) the active site that is suitable for bi-dentate ligand development. These methods are generally applicable to identification of surface epitopes and binding pockets for development of diagnostic reagents, therapeutics and vaccines.

Rapid advances in the genomic sequencing of bacteria and viruses over the past few years have made it possible to consider sequencing the genomes of all pathogens that affect humans and the crops and livestock upon which our lives depend. Recent events make it imperative that full genome sequencing be accomplished as soon as possible for pathogens that could be used as weapons of mass destruction or disruption. This sequence information must be exploited to provide rapid and accurate diagnostics to identify pathogens and distinguish them from harmless near-neighbours and hoaxes. The Chem-Bio Non-Proliferation (CBNP) programme of the US Department of Energy (DOE) began a large-scale effort of pathogen detection in early 2000 when it was announced that the DOE would be providing bio-security at the 2002 Winter Olympic Games in Salt Lake City, Utah. Our team at the Lawrence Livermore National Lab (LLNL) was given the task of developing reliable and validated assays for a number of the most likely bioterrorist agents. The short timeline led us to devise a novel system that utilised whole-genome comparison methods to rapidly focus on parts of the pathogen genomes that had a high probability of being unique. Assays developed with this approach have been validated by the Centers for Disease Control (CDC). They were used at the 2002 Winter Olympics, have entered the public health system, and have been in continual use for non-publicised aspects of homeland defence since autumn 2001. Assays have been developed for all major threat list agents for which adequate genomic sequence is available, as well as for other pathogens requested by various government agencies. Collaborations with comparative genomics algorithm developers have enabled our LLNL team to make major advances in pathogen detection, since many of the existing tools simply did not scale well enough to be of practical use for this application. It is hoped that a discussion of a real-life practical application of comparative genomics algorithms may help spur algorithm developers to tackle some of the many remaining problems that need to be addressed. Solutions to these problems will advance a wide range of biological disciplines, only one of which is pathogen detection. For example, exploration in evolution and phylogenetics, annotating gene coding regions, predicting and understanding gene function and regulation, and untangling gene networks all rely on tools for aligning multiple sequences, detecting gene rearrangements and duplications, and visualising genomic data. Two key problems currently needing improved solutions are: (1) aligning incomplete, fragmentary sequence (eg draft genome contigs or arbitrary genome regions) with both complete genomes and other fragmentary sequences; and (2) ordering, aligning and visualising non-colinear gene rearrangements and inversions in addition to the colinear alignments handled by current tools.

There has been a significant increase, fueled by technologies front the human genome project, in the availability of nucleic acid sequence information for viruses and bacteria. This paper presents a computer-assisted process that begins with nucleic acid sequence information and produces highly specific pathogen signatures. When combined with instrumentation using the polymerase chain reaction, the resulting diagnostics are both specific and sensitive. The computational and engineering aspects of converting raw sequence data into pathogen-specific and instrument-ready assays are presented. Examples and data are presented for specific pathogens, including foot-and-mouth disease virus and the human immunodeficiency virus.

BACKGROUND: Prediction of protein structures is one of the fundamental challenges in biology today. To fully understand how well different prediction methods perform, it is necessary to use measures that evaluate their performance. Every two years, starting in 1994, the CASP (Critical Assessment of protein Structure Prediction) process has been organized to evaluate the ability of different predictors to blindly predict the structure of proteins. To capture different features of the models, several measures have been developed during the CASP processes. However, these measures have not been examined in detail before. In an attempt to develop fully automatic measures that can be used in CASP, as well as in other type of benchmarking experiments, we have compared twenty-one measures. These measures include the measures used in CASP3 and CASP2 as well as have measures introduced later. We have studied their ability to distinguish between the better and worse models submitted to CASP3 and the correlation between them. RESULTS: Using a small set of 1340 models for 23 different targets we show that most methods correlate with each other. Most pairs of measures show a correlation coefficient of about 0.5. The correlation is slightly higher for measures of similar types. We found that a significant problem when developing automatic measures is how to deal with proteins of different length. Also the comparisons between different measures is complicated as many measures are dependent on the size of the target. We show that the manual assessment can be reproduced to about 70% using automatic measures. Alignment independent measures, detects slightly more of the models with the correct fold, while alignment dependent measures agree better when selecting the best models for each target. Finally we show that using automatic measures would, to a large extent, reproduce the assessors ranking of the predictors at CASP3. CONCLUSIONS: We show that given a sufficient number of targets the manual and automatic measures would have given almost identical results at CASP3. If the intent is to reproduce the type of scoring done by the manual assessor in in CASP3, the best approach might be to use a combination of alignment independent and alignment dependent measures, as used in several recent studies.