Purpose To design a computational method for automatic brain glioma segmentation of multimodal MRI scans with high efficiency and accuracy. Materials and Methods The 2018 Multimodal Brain Tumor Segmentation Challenge (BraTS) dataset was used in this study, consisting of routine clinically acquired preoperative multimodal MRI scans. Three subregions of glioma-the necrotic and nonenhancing tumor core, the peritumoral edema, and the contrast-enhancing tumor-were manually labeled by experienced radiologists. Two-dimensional U-Net models were built using a three-plane-assembled approach to segment three subregions individually (three-region model) or to segment only the whole tumor (WT) region (WT-only model). The term three-plane-assembled means that coronal and sagittal images were generated by reformatting the original axial images. The model performance for each case was evaluated in three classes: enhancing tumor (ET), tumor core (TC), and WT. Results On the internal unseen testing dataset split from the 2018 BraTS training dataset, the proposed models achieved mean Sørensen-Dice scores of 0.80, 0.84, and 0.91, respectively, for ET, TC, and WT. On the BraTS validation dataset, the proposed models achieved mean 95% Hausdorff distances of 3.1 mm, 7.0 mm, and 5.0 mm, respectively, for ET, TC, and WT and mean Sørensen-Dice scores of 0.80, 0.83, and 0.91, respectively, for ET, TC, and WT. On the BraTS testing dataset, the proposed models ranked fourth out of 61 teams. The source code is available at https://github.com/GuanLab/Brain_Glioma. Conclusion This deep learning method consistently segmented subregions of brain glioma with high accuracy, efficiency, reliability, and generalization ability on screening images from a large population, and it can be efficiently implemented in clinical practice to assist neuro-oncologists or radiologists. Supplemental material is available for this article. © RSNA, 2020.

ABSTRACT Objective Type 2 diabetes (T2D) is a complex disease characterized by pancreatic islet dysfunction, insulin resistance, and disruption of blood glucose levels. Genome wide association studies (GWAS) have identified >400 independent signals that encode genetic predisposition. More than 90% of the associated single nucleotide polymorphisms (SNPs) localize to non-coding regions and are enriched in chromatin-defined islet enhancer elements, indicating a strong transcriptional regulatory component to disease susceptibility. Pancreatic islets are a mixture of cell types that express distinct hormonal programs, and so each cell type may contribute differentially to the underlying regulatory processes that modulate T2D-associated transcriptional circuits. Existing chromatin profiling methods such as ATAC-seq and DNase-seq, applied to islets in bulk, produce aggregate profiles that mask important cellular and regulatory heterogeneity. Methods We present genome-wide single cell chromatin accessibility profiles in >1,600 cells derived from a human pancreatic islet sample using single-cell-combinatorial-indexing ATAC-seq (sci-ATAC-seq). We also developed a deep learning model based on the U-Net architecture to accurately predict open chromatin peak calls in rare cell populations. Results We show that sci-ATAC-seq profiles allow us to deconvolve alpha, beta, and delta cell populations and identify cell-type-specific regulatory signatures underlying T2D. Particularly, we find that T2D GWAS SNPs are significantly enriched in beta cell-specific and cross cell-type shared islet open chromatin, but not in alpha or delta cell-specific open chromatin. We also demonstrate, using less abundant delta cells, that deep-learning models can improve signal recovery and feature reconstruction of rarer cell populations. Finally, we use co-accessibility measures to nominate the cell-specific target genes at 104 non-coding T2D GWAS signals. Conclusions Collectively, we identify the islet cell type of action across genetic signals of T2D predisposition and provide higher-resolution mechanistic insights into genetically encoded risk pathways.

ObjectiveAccurate prediction of treatment responses in rheumatoid arthritis (RA) patients can provide valuable information on effective drug selection. Anti–tumor necrosis factor (anti‐TNF) drugs are an important second‐line treatment after methotrexate, the classic first‐line treatment for RA. However, patient heterogeneity hinders identification of predictive biomarkers and accurate modeling of anti‐TNF drug responses. This study was undertaken to investigate the usefulness of machine learning to assist in developing predictive models for treatment response.MethodsUsing data on patient demographics, baseline disease assessment, treatment, and single‐nucleotide polymorphism (SNP) array from the Dialogue on Reverse Engineering Assessment and Methods (DREAM): Rheumatoid Arthritis Responder Challenge, we created a Gaussian process regression model to predict changes in the Disease Activity Score in 28 joints (DAS28) for the patients and to classify them into either the responder or the nonresponder group. This model was developed and cross‐validated using data from 1,892 RA patients. It was evaluated using an independent data set from 680 patients. We examined the effectiveness of the similarity modeling and the contribution of individual features.ResultsIn the cross‐validation tests, our method predicted changes in DAS28 (ΔDAS28), with a correlation coefficient of 0.405. It correctly classified responses from 78% of patients. In the independent test, this method achieved a Pearson's correlation coefficient of 0.393 in predicting ΔDAS28. Gaussian process regression effectively remapped the feature space and identified subpopulations that do not respond well to anti‐TNF treatments. Genetic SNP biomarkers showed small contributions in the prediction when added to the clinical models. This was the best‐performing model in the DREAM Challenge.ConclusionThe model described here shows promise in guiding treatment decisions in clinical practice, based primarily on clinical profiles with additional genetic information.

Due to the large numbers of transcription factors (TFs) and cell types, querying binding profiles of all TF/cell type pairs is not experimentally feasible, owing to constraints in time and resources. To address this issue, we developed a convolutional-recurrent neural network model, called FactorNet, to computationally impute the missing binding data. FactorNet trains on binding data from reference cell types to make accurate predictions on testing cell types by leveraging a variety of features, including genomic sequences, genome annotations, gene expression, and single-nucleotide resolution sequential signals, such as DNase I cleavage. To the best of our knowledge, this is the first deep learning method to study the rules governing TF binding at such a fine resolution. With FactorNet, a researcher can perform a single sequencing assay, such as DNase-seq, on a cell type and computationally impute dozens of TF binding profiles. This is an integral step for reconstructing the complex networks underlying gene regulation. While neural networks can be computationally expensive to train, we introduce several novel strategies to significantly reduce the overhead. By visualizing the neural network models, we can interpret how the model predicts binding which in turn reveals additional insights into regulatory grammar. We also investigate the variables that affect cross-cell type predictive performance to explain why the model performs better on some TF/cell types than others, and offer insights to improve upon this field. Our method ranked among the top four teams in the ENCODE-DREAM in vivo Transcription Factor Binding Site Prediction Challenge.

The ENCyclopedia of DNA Elements (ENCODE) consortium has generated transcription factor (TF) binding ChIP-seq data covering hundreds of TF proteins and cell types; however, due to limits on time and resources, only a small fraction of all possible TF-cell type pairs have been profiled. One solution is to build machine learning models trained on currently available epigenomic data sets that can be applied to the remaining missing pairs. A major challenge is that TF binding sites are cell-type–specific, which can be attributed to cellular contexts such as chromatin accessibility. Meanwhile, indirect TF-DNA binding and interactions between TFs complicate this regulatory process. Technical issues such as sequencing biases and batch effects render the prediction task even more challenging. Many pioneering efforts have been made to predict TF binding profiles based on DNA sequence and DNase-seq footprints, but to what extent a model can be generalized to completely untested cell conditions remains unknown. In this study, we describe our first place solution to the 2017 ENCODE-DREAM in vivo TF binding site prediction challenge. By carefully addressing multisource biases and information imbalance across cell types, we created a pipeline that significantly outperforms the current state-of-the-art methods. The proposed method is sufficiently complex enough to model nonlinear interactions between TF binding motifs and chromatin accessibility information up to 1500 bp from the genomic region of interest.

Motivation Motif discovery in large biopolymer sequence datasets can be computationally demanding, presenting significant challenges for discovery in omics research. MEME, arguably one of the most popular motif discovery software, takes quadratic time with respect to dataset size, leading to excessively long runtimes for large datasets. Therefore, there is a demand for fast programs that can generate results of the same quality as MEME. Results Here we describe YAMDA, a highly scalable motif discovery software package. It is built on Pytorch, a tensor computation deep learning library with strong GPU acceleration that is highly optimized for tensor operations that are also useful for motifs. YAMDA takes linear time to find motifs as accurately as MEME, completing in seconds or minutes, which translates to speedups over a thousandfold. Availability YAMDA is freely available on Github ( https://github.com/daquang/YAMDA ) Contact daquang@umich.edu

Combination therapies are commonly used to treat patients with complex diseases that respond poorly to single-agent therapies. In vitro high-throughput drug screening is a standard method for preclinical prioritization of synergistic drug combinations, but it can be impractical for large drug sets. Computational methods are thus being actively explored; however, most published methods were built on a limited size of cancer cell lines or drugs, and it remains a challenge to predict synergism at a large scale where the diversity within the data escalates the difficulty of prediction. Here, we present a state-of-the-field synergy prediction algorithm, which ranked first in all subchallenges in the AstraZeneca-Sanger Drug Combination Prediction DREAM Challenge. The model was built and evaluated using the largest drug combination screening dataset at the time of the competition, consisting of approximately 11,500 experimentally tested synergy scores of 118 drugs in 85 cancer cell lines. We developed a novel feature extraction strategy by integrating the cross-cell and cross-drug information with a novel network propagation method and then assembled the information in monotherapy and simulated molecular data to predict drug synergy. This represents a significant conceptual advancement of synergy prediction, using extracted features in the form of simulated posttreatment molecular profiles when only the pretreatment molecular profile is available. Our cross-tissue synergism prediction algorithm achieves promising accuracy comparable with the correlation between experimental replicates and can be applied to other cancer cell lines and drugs to guide therapeutic choices.Significance: This study presents a novel network propagation-based method that predicts anticancer drug synergy to the accuracy of experimental replicates, which establishes a state-of-the-field method as benchmarked by the pharmacogenomics research community involving models generated by 160 teams. Cancer Res; 78(18); 5446-57. ©2018 AACR.

Comparative genomics has been a powerful tool for identifying functional elements in the human genome. Millions of conserved elements have been discovered. However, understanding the functional roles of these elements still remain a challenge, especially in noncoding regions. In particular, it is still unclear why these elements are evolutionarily conserved and what kind of functional elements are encoded within these sequences. We present a deep learning framework, DeepCons, to further understand potential functional elements within conserved sequences. DeepCons is a convolutional neural net (CNN) that receives a short segment of DNA sequence as input and outputs the probability of the sequence of being evolutionary conserved. The CNN model utilizes hundreds of convolution kernels, which are analogous to sequence motifs, to extract features from DNA sequences during the training process. First, we train the model to discriminate 887,577 conserved elements from a matched number of nonconserved elements in the human genome. Then, we use visualization techniques to interpret how the model discriminates between the two classes of sequences, which provides indirect clues to the functional roles of conserved elements. Some kernels significantly match well-known regulatory motifs corresponding to transcription factors. Many kernels show positional biases relative to transcription start sites or transcription end sites. Most of the kernels do not correspond to any known functional element, suggesting that they might represent unknown categories of functional elements. We also utilize DeepCons to annotate how changes at individual nucleotides impact the conservation properties of the surrounding sequences, thereby providing an annotation of conserved sequences at an individual nucleotide level. The source code of DeepCons is publicly available at https://github.com/uci-cbcl/DeepCons.

Multidrug resistant bacteria are becoming a major threat to global public health. While there are many possible causes for this, there have so far been few adequate solutions to this problem. One of the major causes is a lack of clinical tools for efficient selection of an antibiotic in a reliable way. NastyBugs is a new program that can identify what type of antimicrobial resistance is most likely present in a metagenomic sample, which will allow for both smarter drug selection by clinicians and faster research in an academic environment.

Multidrug resistant bacteria are becoming a major threat to global public health. While there are many possible causes for this, there have so far been few adequate solutions to this problem. One of the major causes is a lack of clinical tools for efficient selection of an antibiotic in a reliable way. NastyBugs is a new program that can identify what type of antimicrobial resistance is most likely present in a metagenomic sample, which will allow for both smarter drug selection by clinicians and faster research in an academic environment.

Author(s): Quang, Daniel | Advisor(s): Xie, Xiaohui | Abstract: High-throughput sequencing (HTS) has led to many breakthroughs in basic and translational biology research. With this technology, researchers can interrogate whole genomes at single-nucleotide resolution. The large volume of data generated by HTS experiments necessitates the development of novel algorithms that can efficiently process these data. At the advent of HTS, several rudimentary methods were proposed. Often, these methods applied compromising strategies such as discarding a majority of the data or reducing the complexity of the models. This thesis focuses on the development of machine learning methods for efficiently capturing complex patterns from high volumes of HTS data.First, we focus on on de novo motif discovery, a popular sequence analysis method that predates HTS. Given multiple input sequences, the goal of motif discovery is to identify one or more candidate motifs, which are biopolymer sequence patterns that are conjectured to have biological significance. In the context of transcription factor (TF) binding, motifs may represent the sequence binding preference of proteins. Traditional motif discovery algorithms do not scale well with the number of input sequences, which can make motif discovery intractable for the volume of data generated by HTS experiments. One common solution is to only perform motif discovery on a small fraction of the sequences. Scalable algorithms that simplify the motif models are popular alternatives. Our approach is a stochastic method that is scalable and retains the modeling power of past methods. Second, we leverage deep learning methods to annotate the pathogenicity of genetic variants. Deep learning is a class of machine learning algorithms concerned with deep neural networks (DNNs). DNNs use a cascade of layers of nonlinear processing units for feature extraction and transformation. Each layer uses the output from the previous layer as its input. Similar to our novel motif discovery algorithm, artificial neural networks can be efficiently trained in a stochastic manner. Using a large labeled dataset comprised of tens of millions of pathogenic and benign genetic variants, we trained a deep neural network to discriminate between the two categories. Previous methods either focused only on variants lying in protein coding regions, which cover less than 2% of the human genome, or applied simpler models such as linear support vector machines, which can not usually capture non-linear patterns like deep neural networks can. Finally, we discuss convolutional (CNN) and recurrent (RNN) neural networks, variations of DNNs that are especially well-suited for studying sequential data. Specifically, we stacked a bidirectional recurrent layer on top of a convolutional layer to form a hybrid model. The model accepts raw DNA sequences as inputs and predicts chromatin markers, including histone modifications, open chromatin, and transcription factor binding. In this specific application, the convolutional kernels are analogous to motifs, hence the model learning is essentially also performing motif discovery. Compared to a pure convolutional model, the hybrid model requires fewer free parameters to achieve superior performance. We conjecture that the recurrent layer allows our model spatial and orientation dependencies among motifs better than a pure convolutional model can. With some modifications to this framework, the model can accept cell type-specific features, such as gene expression and open chromatin DNase I cleavage, to accurately predict transcription factor binding across cell types. We submitted our model to the ENCODE-DREAM in vivo Transcription Factor Binding Site Prediction Challenge, where it was among the top performing models. We implemented several novel heuristics, which significantly reduced the training time and the computational overhead. These heuristics were instrumental to meet the Challenge deadlines and to make the method more accessible for the research community.HTS has already transformed the landscape of basic and translational research, proving itself as a mainstay of modern biological research. As more data are generated and new assays are developed, there will be an increasing need for computational methods to integrate the data to yield new biological insights. We have only begun to scratch the surface of discovering what is possible from both an experimental and a computational perspective. Thus, further development of versatile and efficient statistical models is crucial to maintaining the momentum for new biological discoveries.

Modeling the properties and functions of DNA sequences is an important, but challenging task in the broad field of genomics. This task is particularly difficult for non-coding DNA, the vast majority of which is still poorly understood in terms of function. A powerful predictive model for the function of non-coding DNA can have enormous benefit for both basic science and translational research because over 98% of the human genome is non-coding and 93% of disease-associated variants lie in these regions. To address this need, we propose DanQ, a novel hybrid convolutional and bi-directional long short-term memory recurrent neural network framework for predicting non-coding function de novo from sequence. In the DanQ model, the convolution layer captures regulatory motifs, while the recurrent layer captures long-term dependencies between the motifs in order to learn a regulatory 'grammar' to improve predictions. DanQ improves considerably upon other models across several metrics. For some regulatory markers, DanQ can achieve over a 50% relative improvement in the area under the precision-recall curve metric compared to related models.

UNLABELLED:Annotating genetic variants, especially non-coding variants, for the purpose of identifying pathogenic variants remains a challenge. Combined annotation-dependent depletion (CADD) is an algorithm designed to annotate both coding and non-coding variants, and has been shown to outperform other annotation algorithms. CADD trains a linear kernel support vector machine (SVM) to differentiate evolutionarily derived, likely benign, alleles from simulated, likely deleterious, variants. However, SVMs cannot capture non-linear relationships among the features, which can limit performance. To address this issue, we have developed DANN. DANN uses the same feature set and training data as CADD to train a deep neural network (DNN). DNNs can capture non-linear relationships among features and are better suited than SVMs for problems with a large number of samples and features. We exploit Compute Unified Device Architecture-compatible graphics processing units and deep learning techniques such as dropout and momentum training to accelerate the DNN training. DANN achieves about a 19% relative reduction in the error rate and about a 14% relative increase in the area under the curve (AUC) metric over CADD's SVM methodology. AVAILABILITY AND IMPLEMENTATION:All data and source code are available at https://cbcl.ics.uci.edu/public_data/DANN/.

Background: Stretch enhancers (SEs) are large chromatin-defined regulatory elements that are at least 3,000 base pairs (bps) long, in contrast to the median enhancer length of 800 bps. SEs tend to be cell-type specific, regulate cell-type specific gene expression, and are enriched in disease-associated genetic variants in disease-relevant cell types. Transcription factors (TFs) can bind to enhancers to modulate enhancer activity, and their sequence specificity can be represented by motifs. We hypothesize motifs can provide a biological context for how genetic variants contribute to disease.Results: We integrated chromatin state, gene expression, and chromatin accessibility [measured as DNase I Hypersensitive Sites (DHSs)] maps across nine different cell types. Motif enrichment analyses of chromatin-defined enhancer sequences identify several known cell-type specific "master" factors. Furthermore, de novo motif discovery not only recovers many of these motifs, but also identifies novel non-canonical motifs, providing additional insight into TF binding preferences. Across the length of SEs, motifs are most enriched in DHSs, though relative enrichment is also observed outside of DHSs. Interestingly, we show that single nucleotide polymorphisms associated with diseases or quantitative traits significantly overlap motif occurrences located in SEs, but outside of DHSs.Conclusions: These results reinforce the role of SEs in influencing risk for diseases and suggest an expanded regulatory functional role for motifs that occur outside highly accessible chromatin. Furthermore, the motif signatures generated here expand our understanding of the binding preference of well-characterized TFs.

MOTIVATION Identifying regulatory elements is a fundamental problem in the field of gene transcription. Motif discovery-the task of identifying the sequence preference of transcription factor proteins, which bind to these elements-is an important step in this challenge. MEME is a popular motif discovery algorithm. Unfortunately, MEME's running time scales poorly with the size of the dataset. Experiments such as ChIP-Seq and DNase-Seq are providing a rich amount of information on the binding preference of transcription factors. MEME cannot discover motifs in data from these experiments in a practical amount of time without a compromising strategy such as discarding a majority of the sequences. RESULTS We present EXTREME, a motif discovery algorithm designed to find DNA-binding motifs in ChIP-Seq and DNase-Seq data. Unlike MEME, which uses the expectation-maximization algorithm for motif discovery, EXTREME uses the online expectation-maximization algorithm to discover motifs. EXTREME can discover motifs in large datasets in a practical amount of time without discarding any sequences. Using EXTREME on ChIP-Seq and DNase-Seq data, we discover many motifs, including some novel and infrequent motifs that can only be discovered by using the entire dataset. Conservation analysis of one of these novel infrequent motifs confirms that it is evolutionarily conserved and possibly functional. AVAILABILITY AND IMPLEMENTATION All source code is available at the Github repository http://github.com/uci-cbcl/EXTREME.

An appropriate structural superposition identifies similarities and differences between homologous proteins that are not evident from sequence alignments alone. We have coupled our Gaussian‐weighted RMSD (wRMSD) tool with a sequence aligner and seed extension (SE) algorithm to create a robust technique for overlaying structures and aligning sequences of homologous proteins (HwRMSD). HwRMSD overcomes errors in the initial sequence alignment that would normally propagate into a standard RMSD overlay. SE can generate a corrected sequence alignment from the improved structural superposition obtained by wRMSD. HwRMSD's robust performance and its superiority over standard RMSD are demonstrated over a range of homologous proteins. Its better overlay results in corrected sequence alignments with good agreement to HOMSTRAD. Finally, HwRMSD is compared to established structural alignment methods: FATCAT, secondary‐structure matching, combinatorial extension, and Dalilite. Most methods are comparable at placing residue pairs within 2 Å, but HwRMSD places many more residue pairs within 1 Å, providing a clear advantage. Such high accuracy is essential in drug design, where small distances can have a large impact on computational predictions. This level of accuracy is also needed to correct sequence alignments in an automated fashion, especially for omics‐scale analysis. HwRMSD can align homologs with low‐sequence identity and large conformational differences, cases where both sequence‐based and structural‐based methods may fail. The HwRMSD pipeline overcomes the dependency of structural overlays on initial sequence pairing and removes the need to determine the best sequence‐alignment method, substitution matrix, and gap parameters for each unique pair of homologs. Proteins 2012. © 2012 Wiley Periodicals, Inc.