Automated model building and protein identification in cryo-EM maps.

Interpreting electron cryo-microscopy (cryo-EM) maps with atomic models requires high levels of expertise and labour-intensive manual intervention in three-dimensional computer graphics programs1,2. Here we present ModelAngelo, a machine-learning approach for automated atomic model building in cryo-EM maps. By combining information from the cryo-EM map with information from protein sequence and structure in a single graph neural network, ModelAngelo builds atomic models for proteins that are of similar quality to those generated by human experts. For nucleotides, ModelAngelo builds backbones with similar accuracy to those built by humans. By using its predicted amino acid probabilities for each residue in hidden Markov model sequence searches, ModelAngelo outperforms human experts in the identification of proteins with unknown sequences. ModelAngelo will therefore remove bottlenecks and increase objectivity in cryo-EM structure determination.

Spatial landmark detection and tissue registration with deep learning.

Spatial landmarks are crucial in describing histological features between samples or sites, tracking regions of interest in microscopy, and registering tissue samples within a common coordinate framework. Although other studies have explored unsupervised landmark detection, existing methods are not well-suited for histological image data as they often require a large number of images to converge, are unable to handle nonlinear deformations between tissue sections and are ineffective for z-stack alignment, other modalities beyond image data or multimodal data. We address these challenges by introducing effortless landmark detection, a new unsupervised landmark detection and registration method using neural-network-guided thin-plate splines. Our proposed method is evaluated on a diverse range of datasets including histology and spatially resolved transcriptomics, demonstrating superior performance in both accuracy and stability compared to existing approaches.

quantms: a cloud-based pipeline for quantitative proteomics enables the reanalysis of public proteomics data.

The volume of public proteomics data is rapidly increasing, causing a computational challenge for large-scale reanalysis. Here, we introduce quantms ( https://quant,ms.org/ ), an open-source cloud-based pipeline for massively parallel proteomics data analysis. We used quantms to reanalyze 83 public ProteomeXchange datasets, comprising 29,354 instrument files from 13,132 human samples, to quantify 16,599 proteins based on 1.03 million unique peptides. quantms is based on standard file formats improving the reproducibility, submission and dissemination of the data to ProteomeXchange.

Pathway analysis through mutual information.

MOTIVATION: In pathway analysis, we aim to establish a connection between the activity of a particular biological pathway and a difference in phenotype. There are many available methods to perform pathway analysis, many of them rely on an upstream differential expression analysis, and many model the relations between the abundances of the analytes in a pathway as linear relationships. RESULTS: Here, we propose a new method for pathway analysis, MIPath, that relies on information theoretical principles and, therefore, does not model the association between pathway activity and phenotype, resulting in relatively few assumptions. For this, we construct a graph of the data points for each pathway using a nearest-neighbor approach and score the association between the structure of this graph and the phenotype of these same samples using Mutual Information while adjusting for the effects of random chance in each score. The initial nearest neighbor approach evades individual gene-level comparisons, hence making the method scalable and less vulnerable to missing values. These properties make our method particularly useful for single-cell data. We benchmarked our method on several single-cell datasets, comparing it to established and new methods, and found that it produces robust, reproducible, and meaningful scores. AVAILABILITY AND IMPLEMENTATION: Source code is available at https://github.com/statisticalbiotechnology/mipath, or through Python Package Index as "mipathway."

Triqler for Protein Summarization of Data from Data-Independent Acquisition Mass Spectrometry.

A frequent goal, or subgoal, when processing data from a quantitative shotgun proteomics experiment is a list of proteins that are differentially abundant under the examined experimental conditions. Unfortunately, obtaining such a list is a challenging process, as the mass spectrometer analyzes the proteolytic peptides of a protein rather than the proteins themselves. We have previously designed a Bayesian hierarchical probabilistic model, Triqler, for combining peptide identification and quantification errors into probabilities of proteins being differentially abundant. However, the model was developed for data from data-dependent acquisition. Here, we show that Triqler is also compatible with data-independent acquisition data after applying minor alterations for the missing value distribution. Furthermore, we find that it has better performance than a set of compared state-of-the-art protein summarization tools when evaluated on data-independent acquisition data.

Retention Time and Fragmentation Predictors Increase Confidence in Identification of Common Variant Peptides.

Precision medicine focuses on adapting care to the individual profile of patients, for example, accounting for their unique genetic makeup. Being able to account for the effect of genetic variation on the proteome holds great promise toward this goal. However, identifying the protein products of genetic variation using mass spectrometry has proven very challenging. Here we show that the identification of variant peptides can be improved by the integration of retention time and fragmentation predictors into a unified proteogenomic pipeline. By combining these intrinsic peptide characteristics using the search-engine post-processor Percolator, we demonstrate improved discrimination power between correct and incorrect peptide-spectrum matches. Our results demonstrate that the drop in performance that is induced when expanding a protein sequence database can be compensated, hence enabling efficient identification of genetic variation products in proteomics data. We anticipate that this enhancement of proteogenomic pipelines can provide a more refined picture of the unique proteome of patients and thereby contribute to improving patient care.

Toward an Integrated Machine Learning Model of a Proteomics Experiment.

In recent years machine learning has made extensive progress in modeling many aspects of mass spectrometry data. We brought together proteomics data generators, repository managers, and machine learning experts in a workshop with the goals to evaluate and explore machine learning applications for realistic modeling of data from multidimensional mass spectrometry-based proteomics analysis of any sample or organism. Following this sample-to-data roadmap helped identify knowledge gaps and define needs. Being able to generate bespoke and realistic synthetic data has legitimate and important uses in system suitability, method development, and algorithm benchmarking, while also posing critical ethical questions. The interdisciplinary nature of the workshop informed discussions of what is currently possible and future opportunities and challenges. In the following perspective we summarize these discussions in the hope of conveying our excitement about the potential of machine learning in proteomics and to inspire future research.

Survival analysis of pathway activity as a prognostic determinant in breast cancer.

High throughput biology enables the measurements of relative concentrations of thousands of biomolecules from e.g. tissue samples. The process leaves the investigator with the problem of how to best interpret the potentially large number of differences between samples. Many activities in a cell depend on ordered reactions involving multiple biomolecules, often referred to as pathways. It hence makes sense to study differences between samples in terms of altered pathway activity, using so-called pathway analysis. Traditional pathway analysis gives significance to differences in the pathway components' concentrations between sample groups, however, less frequently used methods for estimating individual samples' pathway activities have been suggested. Here we demonstrate that such a method can be used for pathway-based survival analysis. Specifically, we investigate the pathway activities' association with patients' survival time based on the transcription profiles of the METABRIC dataset. Our implementation shows that pathway activities are better prognostic markers for survival time in METABRIC than the individual transcripts. We also demonstrate that we can regress out the effect of individual pathways on other pathways, which allows us to estimate the other pathways' residual pathway activity on survival. Furthermore, we illustrate how one can visualize the often interdependent measures over hierarchical pathway databases using sunburst plots.

Prosit Transformer: A transformer for Prediction of MS2 Spectrum Intensities.

Machine learning has been an integral part of interpreting data from mass spectrometry (MS)-based proteomics for a long time. Relatively recently, a machine-learning structure appeared successful in other areas of bioinformatics, Transformers. Furthermore, the implementation of Transformers within bioinformatics has become relatively convenient due to transfer learning, i.e., adapting a network trained for other tasks to new functionality. Transfer learning makes these relatively large networks more accessible as it generally requires less data, and the training time improves substantially. We implemented a Transformer based on the pretrained model TAPE to predict MS2 intensities. TAPE is a general model trained to predict missing residues from protein sequences. Despite being trained for a different task, we could modify its behavior by adding a prediction head at the end of the TAPE model and fine-tune it using the spectrum intensity from the training set to the well-known predictor Prosit. We demonstrate that the predictor, which we call Prosit Transformer, outperforms the recurrent neural-network-based predictor Prosit, increasing the median angular similarity on its hold-out set from 0.908 to 0.929. We believe that Transformers will significantly increase prediction accuracy for other types of predictions within MS-based proteomics.

Interpretation of the DOME Recommendations for Machine Learning in Proteomics and Metabolomics.

Machine learning is increasingly applied in proteomics and metabolomics to predict molecular structure, function, and physicochemical properties, including behavior in chromatography, ion mobility, and tandem mass spectrometry. These must be described in sufficient detail to apply or evaluate the performance of trained models. Here we look at and interpret the recently published and general DOME (Data, Optimization, Model, Evaluation) recommendations for conducting and reporting on machine learning in the specific context of proteomics and metabolomics.

A Comprehensive Evaluation of Consensus Spectrum Generation Methods in Proteomics.

Spectrum clustering is a powerful strategy to minimize redundant mass spectra by grouping them based on similarity, with the aim of forming groups of mass spectra from the same repeatedly measured analytes. Each such group of near-identical spectra can be represented by its so-called consensus spectrum for downstream processing. Although several algorithms for spectrum clustering have been adequately benchmarked and tested, the influence of the consensus spectrum generation step is rarely evaluated. Here, we present an implementation and benchmark of common consensus spectrum algorithms, including spectrum averaging, spectrum binning, the most similar spectrum, and the best-identified spectrum. We have analyzed diverse public data sets using two different clustering algorithms (spectra-cluster and MaRaCluster) to evaluate how the consensus spectrum generation procedure influences downstream peptide identification. The BEST and BIN methods were found the most reliable methods for consensus spectrum generation, including for data sets with post-translational modifications (PTM) such as phosphorylation. All source code and data of the present study are freely available on GitHub at https://github.com/statisticalbiotechnology/representative-spectra-benchmark.

Putting Humpty Dumpty Back Together Again: What Does Protein Quantification Mean in Bottom-Up Proteomics?

Bottom-up proteomics provides peptide measurements and has been invaluable for moving proteomics into large-scale analyses. Commonly, a single quantitative value is reported for each protein-coding gene by aggregating peptide quantities into protein groups following protein inference or parsimony. However, given the complexity of both RNA splicing and post-translational protein modification, it is overly simplistic to assume that all peptides that map to a singular protein-coding gene will demonstrate the same quantitative response. By assuming that all peptides from a protein-coding sequence are representative of the same protein, we may miss the discovery of important biological differences. To capture the contributions of existing proteoforms, we need to reconsider the practice of aggregating protein values to a single quantity per protein-coding gene.

Parallelized calculation of permutation tests.

MOTIVATION: Permutation tests offer a straightforward framework to assess the significance of differences in sample statistics. A significant advantage of permutation tests are the relatively few assumptions about the distribution of the test statistic are needed, as they rely on the assumption of exchangeability of the group labels. They have great value, as they allow a sensitivity analysis to determine the extent to which the assumed broad sample distribution of the test statistic applies. However, in this situation, permutation tests are rarely applied because the running time of naïve implementations is too slow and grows exponentially with the sample size. Nevertheless, continued development in the 1980s introduced dynamic programming algorithms that compute exact permutation tests in polynomial time. Albeit this significant running time reduction, the exact test has not yet become one of the predominant statistical tests for medium sample size. Here, we propose a computational parallelization of one such dynamic programming-based permutation test, the Green algorithm, which makes the permutation test more attractive. RESULTS: Parallelization of the Green algorithm was found possible by non-trivial rearrangement of the structure of the algorithm. A speed-up-by orders of magnitude-is achievable by executing the parallelized algorithm on a GPU. We demonstrate that the execution time essentially becomes a non-issue for sample sizes, even as high as hundreds of samples. This improvement makes our method an attractive alternative to, e.g. the widely used asymptotic Mann-Whitney U-test. AVAILABILITYAND IMPLEMENTATION: In Python 3 code from the GitHub repository https://github.com/statisticalbiotechnology/parallelPermutationTest under an Apache 2.0 license. SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

Triqler for MaxQuant: Enhancing Results from MaxQuant by Bayesian Error Propagation and Integration.

Error estimation for differential protein quantification by label-free shotgun proteomics is challenging due to the multitude of error sources, each contributing uncertainty to the final results. We have previously designed a Bayesian model, Triqler, to combine such error terms into one combined quantification error. Here we present an interface for Triqler that takes MaxQuant results as input, allowing quick reanalysis of already processed data. We demonstrate that Triqler outperforms the original processing for a large set of both engineered and clinical/biological relevant data sets. Triqler and its interface to MaxQuant are available as a Python module under an Apache 2.0 license from https://pypi.org/project/triqler/.

Performing Selection on a Monotonic Function in Lieu of Sorting Using Layer-Ordered Heaps.

Nonparametric statistical tests are an integral part of scientific experiments in a diverse range of fields. When performing such tests, it is standard to sort values; however, this requires Ω(n log(n)) time to sort n values. Thus given enough data, sorting becomes the computational bottleneck, even with very optimized implementations such as the C++ standard library routine, std::sort. Frequently, a nonparametric statistical test is only used to partition values above and below a threshold in the sorted ordering, where the threshold corresponds to a significant statistical result. Linear-time selection and partitioning algorithms cannot be directly used because the selection and partitioning are performed on the transformed statistical significance values rather than on the sorted statistics. Usually, those transformed statistical significance values (e.g., the p value when investigating the family-wise error rate and q values when investigating the false discovery rate (FDR)) can only be computed at a threshold. Because this threshold is unknown, this leads to sorting the data. Layer-ordered heaps, which can be constructed in O(n), only partially sort values and thus can be used to get around the slow runtime required to fully sort. Here we introduce a layer-ordering-based method for selection and partitioning on the transformed values (e.g., p values or q values). We demonstrate the use of this method to partition peptides using an FDR threshold. This approach is applied to speed up Percolator, a postprocessing algorithm used in mass-spectrometry-based proteomics to evaluate the quality of peptide-spectrum matches (PSMs), by >70% on data sets with 100 million PSMs.

Integrated Identification and Quantification Error Probabilities for Shotgun Proteomics.

Protein quantification by label-free shotgun proteomics experiments is plagued by a multitude of error sources. Typical pipelines for identifying differential proteins use intermediate filters to control the error rate. However, they often ignore certain error sources and, moreover, regard filtered lists as completely correct in subsequent steps. These two indiscretions can easily lead to a loss of control of the false discovery rate (FDR). We propose a probabilistic graphical model, Triqler, that propagates error information through all steps, employing distributions in favor of point estimates, most notably for missing value imputation. The model outputs posterior probabilities for fold changes between treatment groups, highlighting uncertainty rather than hiding it. We analyzed 3 engineered data sets and achieved FDR control and high sensitivity, even for truly absent proteins. In a bladder cancer clinical data set we discovered 35 proteins at 5% FDR, whereas the original study discovered 1 and MaxQuant/Perseus 4 proteins at this threshold. Compellingly, these 35 proteins showed enrichment for functional annotation terms, whereas the top ranked proteins reported by MaxQuant/Perseus showed no enrichment. The model executes in minutes and is freely available at https://pypi.org/project/triqler/.

CoExpresso: assess the quantitative behavior of protein complexes in human cells.

BACKGROUND: Translational and post-translational control mechanisms in the cell result in widely observable differences between measured gene transcription and protein abundances. Herein, protein complexes are among the most tightly controlled entities by selective degradation of their individual proteins. They furthermore act as control hubs that regulate highly important processes in the cell and exhibit a high functional diversity due to their ability to change their composition and their structure. Better understanding and prediction of these functional states demands methods for the characterization of complex composition, behavior, and abundance across multiple cell states. Mass spectrometry provides an unbiased approach to directly determine protein abundances across different cell populations and thus to profile a comprehensive abundance map of proteins. RESULTS: We provide a tool to investigate the behavior of protein subunits in known complexes by comparing their abundance profiles across up to 140 cell types available in ProteomicsDB. Thorough assessment of different randomization methods and statistical scoring algorithms allows determining the significance of concurrent profiles within a complex, therefore providing insights into the conservation of their composition across human cell types as well as the identification of intrinsic structures in complex behavior to determine which proteins orchestrate complex function. This analysis can be extended to investigate common profiles within arbitrary protein groups. CoExpresso can be accessed through http://computproteomics.bmb.sdu.dk/Apps/CoExpresso . CONCLUSIONS: With the CoExpresso web service, we offer a potent scoring scheme to assess proteins for their co-regulation and thereby offer insight into their potential for forming functional groups like protein complexes.

Speeding Up Percolator.

The processing of peptide tandem mass spectrometry data involves matching observed spectra against a sequence database. The ranking and calibration of these peptide-spectrum matches can be improved substantially using a machine learning postprocessor. Here, we describe our efforts to speed up one widely used postprocessor, Percolator. The improved software is dramatically faster than the previous version of Percolator, even when using relatively few processors. We tested the new version of Percolator on a data set containing over 215 million spectra and recorded an overall reduction to 23% of the running time as compared to the unoptimized code. We also show that the memory footprint required by these speedups is modest relative to that of the original version of Percolator.

Covariation of Peptide Abundances Accurately Reflects Protein Concentration Differences.

Most implementations of mass spectrometry-based proteomics involve enzymatic digestion of proteins, expanding the analysis to multiple proteolytic peptides for each protein. Currently, there is no consensus of how to summarize peptides' abundances to protein concentrations, and such efforts are complicated by the fact that error control normally is applied to the identification process, and do not directly control errors linking peptide abundance measures to protein concentration. Peptides resulting from suboptimal digestion or being partially modified are not representative of the protein concentration. Without a mechanism to remove such unrepresentative peptides, their abundance adversely impacts the estimation of their protein's concentration. Here, we present a relative quantification approach, Diffacto, that applies factor analysis to extract the covariation of peptides' abundances. The method enables a weighted geometrical average summarization and automatic elimination of incoherent peptides. We demonstrate, based on a set of controlled label-free experiments using standard mixtures of proteins, that the covariation structure extracted by the factor analysis accurately reflects protein concentrations. In the 1% peptide-spectrum match-level FDR data set, as many as 11% of the peptides have abundance differences incoherent with the other peptides attributed to the same protein. If not controlled, such contradicting peptide abundance have a severe impact on protein quantifications. When adding the quantities of each protein's three most abundant peptides, we note as many as 14% of the proteins being estimated as having a negative correlation with their actual concentration differences between samples. Diffacto reduced the amount of such obviously incorrectly quantified proteins to 1.6%. Furthermore, by analyzing clinical data sets from two breast cancer studies, our method revealed the persistent proteomic signatures linked to three subtypes of breast cancer. We conclude that Diffacto can facilitate the interpretation and enhance the utility of most types of proteomics data.

Response to "Comparison and Evaluation of Clustering Algorithms for Tandem Mass Spectra".

In the recent benchmarking article entitled "Comparison and Evaluation of Clustering Algorithms for Tandem Mass Spectra", Rieder et al. compared several different approaches to cluster MS/MS spectra. While we certainly recognize the value of the manuscript, here, we report some shortcomings detected in the original analyses. For most analyses, the authors clustered only single MS/MS runs. In one of the reported analyses, three MS/MS runs were processed together, which already led to computational performance issues in many of the tested approaches. This fact highlights the difficulties of using many of the tested algorithms on the nowadays produced average proteomics data sets. Second, the authors only processed identified spectra when merging MS runs. Thereby, all unidentified spectra that are of lower quality were already removed from the data set and could not influence the clustering results. Next, we found that the authors did not analyze the effect of chimeric spectra on the clustering results. In our analysis, we found that 3% of the spectra in the used data sets were chimeric, and this had marked effects on the behavior of the different clustering algorithms tested. Finally, the authors' choice to evaluate the MS-Cluster and spectra-cluster algorithms using a precursor tolerance of 5 Da for high-resolution Orbitrap data only was, in our opinion, not adequate to assess the performance of MS/MS clustering approaches.