Identifying a small molecule from its mass spectrum is the primary open problem in computational metabolomics. This is typically cast as information retrieval: an unknown spectrum is matched against spectra predicted computationally from a large database of chemical structures. However, current approaches to spectrum prediction model the output space in ways that force a tradeoff between capturing high resolution mass information and tractable learning. We resolve this tradeoff by casting spectrum prediction as a mapping from an input molecular graph to a probability distribution over chemical formulas. We further discover that a large corpus of mass spectra can be closely approximated using a fixed vocabulary constituting only 2% of all observed formulas. This enables efficient spectrum prediction using an architecture similar to graph classification - GrAFF-MS - achieving significantly lower prediction error and greater retrieval accuracy than previous approaches.

Subseasonal forecasting of the weather two to six weeks in advance is critical for resource allocation and climate adaptation but poses many challenges for the forecasting community. At this forecast horizon, physics-based dynamical models have limited skill, and the targets for prediction depend in a complex manner on both local weather variables and global climate variables. Recently, machine learning methods have shown promise in advancing the state of the art but only at the cost of complex data curation, integrating expert knowledge with aggregation across multiple relevant data sources, file formats, and temporal and spatial resolutions.To streamline this process and accelerate future development, we introduce SubseasonalClimateUSA, a curated dataset for training and benchmarking subseasonal forecasting models in the United States. We use this dataset to benchmark a diverse suite of models, including operational dynamical models, classical meteorological baselines, and ten state-of-the-art machine learning and deep learning-based methods from the literature. Overall, our benchmarks suggest simple and effective ways to extend the accuracy of current operational models. SubseasonalClimateUSA is regularly updated and accessible via the https://github.com/microsoft/subseasonal_data/ Python package.

MOTIVATION: Advances in bioimaging now permit in situ proteomic characterization of cell-cell interactions in complex tissues, with important applications across a spectrum of biological problems from development to disease. These methods depend on selection of antibodies targeting proteins that are expressed specifically in particular cell types. Candidate marker proteins are often identified from single-cell transcriptomic data, with variable rates of success, in part due to divergence between expression levels of proteins and the genes that encode them. In principle, marker identification could be improved by using existing databases of immunohistochemistry for thousands of antibodies in human tissue, such as the Human Protein Atlas. However, these data lack detailed annotations of the types of cells in each image. RESULTS: We develop a method to predict cell type specificity of protein markers from unlabeled images. We train a convolutional neural network with a self-supervised objective to generate embeddings of the images. Using non-linear dimensionality reduction, we observe that the model clusters images according to cell types and anatomical regions for which the stained proteins are specific. We then use estimates of cell type specificity derived from an independent single-cell transcriptomics dataset to train an image classifier, without requiring any human labelling of images. Our scheme demonstrates superior classification of known proteomic markers in kidney compared to selection via single-cell transcriptomics. AVAILABILITY AND IMPLEMENTATION: Code and trained model are available at www.github.com/murphy17/HPA-SimCLR. SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

We propose a cloud-based multimodal dialog platform for the remote assessment and monitoring of Amyotrophic Lateral Sclerosis (ALS) at scale. This paper presents our vision, technology setup, and an initial investigation of the efficacy of the various acoustic and visual speech metrics automatically extracted by the platform. 82 healthy controls and 54 people with ALS (pALS) were instructed to interact with the platform and completed a battery of speaking tasks designed to probe the acoustic, articulatory, phonatory, and respiratory aspects of their speech. We find that multiple acoustic (rate, duration, voicing) and visual (higher order statistics of the jaw and lip) speech metrics show statistically significant differences between controls, bulbar symptomatic and bulbar pre-symptomatic patients. We report on the sensitivity and specificity of these metrics using five-fold cross-validation. We further conducted a LASSO-LARS regression analysis to uncover the relative contributions of various acoustic and visual features in predicting the severity of patients' ALS (as measured by their self-reported ALSFRS-R scores). Our results provide encouraging evidence of the utility of automatically extracted audiovisual analytics for scalable remote patient assessment and monitoring in ALS.

With the recent technological developments a vast amount of high-throughput data has been profiled to understand the mechanism of complex diseases. The current bioinformatics challenge is to interpret the data and underlying biology, where efficient algorithms for analyzing heterogeneous high-throughput data using biological networks are becoming increasingly valuable. In this paper, we propose a software package based on the Prize-collecting Steiner Forest graph optimization approach. The PCSF package performs fast and user-friendly network analysis of high-throughput data by mapping the data onto a biological networks such as protein-protein interaction, gene-gene interaction or any other correlation or coexpression based networks. Using the interaction networks as a template, it determines high-confidence subnetworks relevant to the data, which potentially leads to predictions of functional units. It also interactively visualizes the resulting subnetwork with functional enrichment analysis.

High-throughput, 'omic' methods provide sensitive measures of biological responses to perturbations. However, inherent biases in high-throughput assays make it difficult to interpret experiments in which more than one type of data is collected. In this work, we introduce Omics Integrator, a software package that takes a variety of 'omic' data as input and identifies putative underlying molecular pathways. The approach applies advanced network optimization algorithms to a network of thousands of molecular interactions to find high-confidence, interpretable subnetworks that best explain the data. These subnetworks connect changes observed in gene expression, protein abundance or other global assays to proteins that may not have been measured in the screens due to inherent bias or noise in measurement. This approach reveals unannotated molecular pathways that would not be detectable by searching pathway databases. Omics Integrator also provides an elegant framework to incorporate not only positive data, but also negative evidence. Incorporating negative evidence allows Omics Integrator to avoid unexpressed genes and avoid being biased toward highly-studied hub proteins, except when they are strongly implicated by the data. The software is comprised of two individual tools, Garnet and Forest, that can be run together or independently to allow a user to perform advanced integration of multiple types of high-throughput data as well as create condition-specific subnetworks of protein interactions that best connect the observed changes in various datasets. It is available at http://fraenkel.mit.edu/omicsintegrator and on GitHub at https://github.com/fraenkel-lab/OmicsIntegrator.

MOTIVATION: High-throughput datasets such as genetic screens, mRNA expression assays and global phospho-proteomic experiments are often difficult to interpret due to inherent noise in each experimental system. Computational tools have improved interpretation of these datasets by enabling the identification of biological processes and pathways that are most likely to explain the measured results. These tools are primarily designed to analyse data from a single experiment (e.g. drug treatment versus control), creating a need for computational algorithms that can handle heterogeneous datasets across multiple experimental conditions at once. SUMMARY: We introduce SAMNetWeb, a web-based tool that enables functional enrichment analysis and visualization of high-throughput datasets. SAMNetWeb can analyse two distinct data types (e.g. mRNA expression and global proteomics) simultaneously across multiple experimental systems to identify pathways activated in these experiments and then visualize the pathways in a single interaction network. Through the use of a multi-commodity flow based algorithm that requires each experiment 'share' underlying protein interactions, SAMNetWeb can identify distinct and common pathways across experiments. AVAILABILITY AND IMPLEMENTATION: SAMNetWeb is freely available at http://fraenkel.mit.edu/samnetweb.

Cellular signal transduction generally involves cascades of post-translational protein modifications that rapidly catalyze changes in protein-DNA interactions and gene expression. High-throughput measurements are improving our ability to study each of these stages individually, but do not capture the connections between them. Here we present an approach for building a network of physical links among these data that can be used to prioritize targets for pharmacological intervention. Our method recovers the critical missing links between proteomic and transcriptional data by relating changes in chromatin accessibility to changes in expression and then uses these links to connect proteomic and transcriptome data. We applied our approach to integrate epigenomic, phosphoproteomic and transcriptome changes induced by the variant III mutation of the epidermal growth factor receptor (EGFRvIII) in a cell line model of glioblastoma multiforme (GBM). To test the relevance of the network, we used small molecules to target highly connected nodes implicated by the network model that were not detected by the experimental data in isolation and we found that a large fraction of these agents alter cell viability. Among these are two compounds, ICG-001, targeting CREB binding protein (CREBBP), and PKF118-310, targeting β-catenin (CTNNB1), which have not been tested previously for effectiveness against GBM. At the level of transcriptional regulation, we used chromatin immunoprecipitation sequencing (ChIP-Seq) to experimentally determine the genome-wide binding locations of p300, a transcriptional co-regulator highly connected in the network. Analysis of p300 target genes suggested its role in tumorigenesis. We propose that this general method, in which experimental measurements are used as constraints for building regulatory networks from the interactome while taking into account noise and missing data, should be applicable to a wide range of high-throughput datasets.

Understanding the mechanistic basis of transcriptional regulation has been a central focus of molecular biology since its inception. New high-throughput chromatin immunoprecipitation experiments have revealed that most regulatory proteins bind thousands of sites in mammalian genomes. However, the functional significance of these binding sites remains unclear. We present a quantitative model of transcriptional regulation that suggests the contribution of each binding site to tissue-specific gene expression depends strongly on its position relative to the transcription start site. For three cell types, we show that, by considering binding position, it is possible to predict relative expression levels between cell types with an accuracy approaching the level of agreement between different experimental platforms. Our model suggests that, for the transcription factors profiled in these cell types, a regulatory site's influence on expression falls off almost linearly with distance from the transcription start site in a 10 kilobase range. Binding to both evolutionarily conserved and non-conserved sequences contributes significantly to transcriptional regulation. Our approach also reveals the quantitative, tissue-specific role of individual proteins in activating or repressing transcription. These results suggest that regulator binding position plays a previously unappreciated role in influencing expression and blurs the classical distinction between proximal promoter and distal binding events.

MOTIVATION: Genome-wide chromatin-immunoprecipitation (ChIP-chip) detects binding of transcriptional regulators to DNA in vivo at low resolution. Motif discovery algorithms can be used to discover sequence patterns in the bound regions that may be recognized by the immunoprecipitated protein. However, the discovered motifs often do not agree with the binding specificity of the protein, when it is known. RESULTS: We present a powerful approach to analyzing ChIP-chip data, called THEME, that tests hypotheses concerning the sequence specificity of a protein. Hypotheses are refined using constrained local optimization. Cross-validation provides a principled standard for selecting the optimal weighting of the hypothesis and the ChIP-chip data and for choosing the best refined hypothesis. We demonstrate how to derive hypotheses for proteins from 36 domain families. Using THEME together with these hypotheses, we analyze ChIP-chip datasets for 14 human and mouse proteins. In all the cases the identified motifs are consistent with the published data with regard to the binding specificity of the proteins.

BACKGROUND: The regulatory map of a genome consists of the binding sites for proteins that determine the transcription of nearby genes. An initial regulatory map for S. cerevisiae was recently published using six motif discovery programs to analyze genome-wide chromatin immunoprecipitation data for 203 transcription factors. The programs were used to identify sequence motifs that were likely to correspond to the DNA-binding specificity of the immunoprecipitated proteins. We report improved versions of two conservation-based motif discovery algorithms, PhyloCon and Converge. Using these programs, we create a refined regulatory map for S. cerevisiae by reanalyzing the same chromatin immunoprecipitation data. RESULTS: Applying the same conservative criteria that were applied in the original study, we find that PhyloCon and Converge each separately discover more known specificities than the combination of all six programs in the previous study. Combining the results of PhyloCon and Converge, we discover significant sequence motifs for 36 transcription factors that were previously missed. The new set of motifs identifies 636 more regulatory interactions than the previous one. The new network contains 28% more regulatory interactions among transcription factors, evidence of greater cross-talk between regulators. CONCLUSION: Combining two complementary computational strategies for conservation-based motif discovery improves the ability to identify the specificity of transcriptional regulators from genome-wide chromatin immunoprecipitation data. The increased sensitivity of these methods significantly expands the map of yeast regulatory sites without the need to alter any of the thresholds for statistical significance. The new map of regulatory sites reveals a more elaborate and complex view of the yeast genetic regulatory network than was observed previously.

Many functionally important regions of the genome can be recognized by searching for sequence patterns, or “motifs.” Aside from the genes themselves, examples include CpG islands, often present in promoter regions, and splice sites that denote intron/exon boundaries. Other motifs of great interest correspond to sites bound by regulatory proteins. Differential expression of genes in response to environmental and developmental cues depends on the action of these proteins, which are also known as transcription factors. Identifying the regulatory motifs bound by transcription factors can provide crucial insight into the mechanisms of transcriptional regulation. However, the search for these sites is challenging because a single regulatory protein will often recognize a variety of similar sequences. In this tutorial, we review computational techniques, termed “motif discovery,” to learn representations of regulatory motifs from sequence data. In Figure 1, we present an overview of the basic workflow in a motif discovery analysis and some practical strategies for successfully mining sequence data for biologically important regulatory motifs. In the remainder of this tutorial, we discuss the main challenges associated with motif discovery in detail, and we review recent developments for addressing these challenges. Figure 1 Motif Discovery Workflow