Randomized smoothing is a leading method for achieving certified robustness in deep classifiers against ℓp-norm adversarial perturbations. However, randomized smoothing requires expensive training procedures that tune large models for different Gaussian noise levels from scratch and thus cannot leverage high-performance pre-trained neural networks. In this work, we introduce the certifying adapters framework (CAF) that enables and enhances the certification of classifier adversarial robustness. Our approach makes few assumptions about the underlying training algorithm or feature extractor, and is thus broadly applicable to different feature extractor architectures (e.g., convolutional neural networks or vision transformers) and randomized smoothing algorithms. We show that CAF (a) enables certification in uncertified models pre-trained on clean datasets and (b) substantially improves the performance of classifiers certified using randomized smoothing and SmoothAdv at multiple radii in CIFAR-10 and ImageNet. Classifiers trained with CAF achieve substantially improved certified accuracies compared to random or denoised smoothing methods. Finally, we demonstrate that CAF is insensitive to hyperparameter settings and adapter ensembles enable a single pre-trained feature extractor to defend against a range of noise perturbation scales.

MOTIVATION: Characterizing the differential excision of introns is critical for understanding the functional complexity of a cell or tissue, from normal developmental processes to disease pathogenesis. Most transcript reconstruction methods infer full-length transcripts from high-throughput sequencing data. However, this is a challenging task due to incomplete annotations and the heterogeneous expression of transcripts across cell-types, tissues, and experimental conditions. Several recent methods circumvent these difficulties by considering local splicing events, but these methods lose transcript-level splicing information and may conflate similar, but distinct transcripts. RESULTS: In this work, we formalize a new transcript reconstruction problem that interpolates between the full-length and local splicing perspectives by considering sequences of exon-exon junctions (SEEJs) that co-occur in transcripts. We then present a hierarchical Bayesian admixture model and posterior inference algorithms for computing SEEJs (BSEEJ), and a generalized linear model for characterizing differential SEEJ usage based on model parameter estimates. We show that BSEEJ achieves high F1 score for reconstruction tasks and improved accuracy and sensitivity in differential splicing when compared with six transcript and local splicing methods on simulated data. Lastly, we evaluate BSEEJ on experimental data based on transcript reconstruction, novelty of transcripts produced, model sensitivity to hyperparameters, and a functional analysis of differentially expressed SEEJs. AVAILABILITY AND IMPLEMENTATION: BSEEJ is freely available at https://github.com/bayesomicslab/BSEEJ.

The bokeh effect is an artistic technique that blurs out-of-focus areas in a photograph and has gained interest due to recent developments in text-to-image synthesis and the ubiquity of smartphone cameras and photo sharing apps. Prior work on rendering bokeh effects have focused on manipulating photographs using classical computer graphics or neural rendering techniques, but have either depth discontinuity artifacts or are restricted to reproducing bokeh effects that are present in the training data. In this paper, we present generative bokeh with stage diffusion (GBSD), the first generative text-to-image model that synthesizes photorealistic images with a bokeh style. Motivated by how image synthesis occurs progressively in diffusion models, our approach combines latent diffusion models with a 2-stage conditioning algorithm to render bokeh effects on semantically defined objects. Since GBSD focuses the blurring effect on objects, this semantic bokeh effect is more versatile than classical rendering techniques. We evaluate GBSD both quantitatively and qualitatively and demonstrate its ability to be applied in both text-to-image and image-to-image settings.

Distilling the structured information captured in feature maps has contributed to improved results for object detection tasks, but requires careful selection of baseline architectures and substantial pre-training. Self-distillation addresses these limitations and has recently achieved state-of-the-art performance for object detection despite making several simplifying architectural assumptions. Building on this work, we propose Smooth and Stepwise Self-Distillation (SSSD) for object detection. Our SSSD architecture forms an implicit teacher from object labels and a feature pyramid network backbone to distill label-annotated feature maps using Jensen-Shannon distance, which is smoother than distillation losses used in prior work. We additionally add a distillation coefficient that is adaptively configured based on the learning rate. We extensively benchmark SSSD against a baseline and two state-of-the-art object detector architectures on the COCO dataset by varying the coefficients and backbone and detector networks. We demonstrate that SSSD achieves higher average precision in most experimental settings, is robust to a wide range of coefficients, and benefits from our stepwise distillation procedure.

MOTIVATION: Non-canonical (or non-B) DNA are genomic regions whose three-dimensional conformation deviates from the canonical double helix. Non-B DNA play an important role in basic cellular processes and are associated with genomic instability, gene regulation, and oncogenesis. Experimental methods are low-throughput and can detect only a limited set of non-B DNA structures, while computational methods rely on non-B DNA base motifs, which are necessary but not sufficient indicators of non-B structures. Oxford Nanopore sequencing is an efficient and low-cost platform, but it is currently unknown whether nanopore reads can be used for identifying non-B structures. RESULTS: We build the first computational pipeline to predict non-B DNA structures from nanopore sequencing. We formalize non-B detection as a novelty detection problem and develop the GoFAE-DND, an autoencoder that uses goodness-of-fit (GoF) tests as a regularizer. A discriminative loss encourages non-B DNA to be poorly reconstructed and optimizing Gaussian GoF tests allows for the computation of P-values that indicate non-B structures. Based on whole genome nanopore sequencing of NA12878, we show that there exist significant differences between the timing of DNA translocation for non-B DNA bases compared with B-DNA. We demonstrate the efficacy of our approach through comparisons with novelty detection methods using experimental data and data synthesized from a new translocation time simulator. Experimental validations suggest that reliable detection of non-B DNA from nanopore sequencing is achievable. AVAILABILITY AND IMPLEMENTATION: Source code is available at https://github.com/bayesomicslab/ONT-nonb-GoFAE-DND.

Lawyers regularly predict court outcomes to make strategic decisions, including when, if at all, to sue or settle, what to argue, and how to reduce their clients' liability risk. Yet, lawyer predictions tend to be poorly calibrated and biased, which exacerbate unjustifiable disparities in civil case outcomes. Current machine learning (ML) approaches for predicting court outcomes are typically constrained to final dispositions or are based on features unavailable in real-time during litigation, like judicial opinions. Here, we present the first ML-based methods to support lawyer and client decision making in real-time for motion filings in civil proceedings. Using the State of Connecticut Judicial Branch administrative data and court case documents, we trained six classifiers to predict motion to strike outcomes in tort and vehicular cases between July 1, 2004 and February 18, 2019. Integrating dense word embeddings from complaint documents, which contain information specific to the claims alleged, with the Judicial Branch data improved classification accuracy across all models. Subsequent models defined using a novel attorney case-entropy feature, dense word embeddings using corpus specific TF-IDF weightings, and algorithmic classification rules yielded the best predictor, Adaboost, with a classification accuracy of 64.4%. An analysis of feature importance weights confirmed the usefulness of incorporating attorney case-entropy and natural language features from complaint documents. Since all features used in model training are available during litigation, these methods will help lawyers make better predictions than they otherwise could given disparities in lawyer and client resources. All ML models, training code, and evaluation scripts are available at https://github.com/aguiarlab/motionpredict.

MOTIVATION: Genome-wide association studies (GWAS) have discovered thousands of significant genetic effects on disease phenotypes. By considering gene expression as the intermediary between genotype and disease phenotype, expression quantitative trait loci studies have interpreted many of these variants by their regulatory effects on gene expression. However, there remains a considerable gap between genotype-to-gene expression association and genotype-to-gene expression prediction. Accurate prediction of gene expression enables gene-based association studies to be performed post hoc for existing GWAS, reduces multiple testing burden, and can prioritize genes for subsequent experimental investigation. RESULTS: In this work, we develop gene expression prediction methods that relax the independence and additivity assumptions between genetic markers. First, we consider gene expression prediction from a regression perspective and develop the HAPLEXR algorithm which combines haplotype clusterings with allelic dosages. Second, we introduce the new gene expression classification problem, which focuses on identifying expression groups rather than continuous measurements; we formalize the selection of an appropriate number of expression groups using the principle of maximum entropy. Third, we develop the HAPLEXD algorithm that models haplotype sharing with a modified suffix tree data structure and computes expression groups by spectral clustering. In both models, we penalize model complexity by prioritizing genetic clusters that indicate significant effects on expression. We compare HAPLEXR and HAPLEXD with three state-of-the-art expression prediction methods and two novel logistic regression approaches across five GTEx v8 tissues. HAPLEXD exhibits significantly higher classification accuracy overall; HAPLEXR shows higher prediction accuracy on approximately half of the genes tested and the largest number of best predicted genes (r2>0.1) among all methods. We show that variant and haplotype features selected by HAPLEXR are smaller in size than competing methods (and thus more interpretable) and are significantly enriched in functional annotations related to gene regulation. These results demonstrate the importance of explicitly modeling non-dosage dependent and intragenic epistatic effects when predicting expression. AVAILABILITY AND IMPLEMENTATION: Source code and binaries are freely available at https://github.com/rapturous/HAPLEX. SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

MOTIVATION: Genome-wide haplotype reconstruction from sequence data, or haplotype assembly, is at the center of major challenges in molecular biology and life sciences. For complex eukaryotic organisms like humans, the genome is vast and the population samples are growing so rapidly that algorithms processing high-throughput sequencing data must scale favorably in terms of both accuracy and computational efficiency. Furthermore, current models and methodologies for haplotype assembly (i) do not consider individuals sharing haplotypes jointly, which reduces the size and accuracy of assembled haplotypes, and (ii) are unable to model genomes having more than two sets of homologous chromosomes (polyploidy). Polyploid organisms are increasingly becoming the target of many research groups interested in the genomics of disease, phylogenetics, botany and evolution but there is an absence of theory and methods for polyploid haplotype reconstruction. RESULTS: In this work, we present a number of results, extensions and generalizations of compass graphs and our HapCompass framework. We prove the theoretical complexity of two haplotype assembly optimizations, thereby motivating the use of heuristics. Furthermore, we present graph theory-based algorithms for the problem of haplotype assembly using our previously developed HapCompass framework for (i) novel implementations of haplotype assembly optimizations (minimum error correction), (ii) assembly of a pair of individuals sharing a haplotype tract identical by descent and (iii) assembly of polyploid genomes. We evaluate our methods on 1000 Genomes Project, Pacific Biosciences and simulated sequence data. AVAILABILITY AND IMPLEMENTATION: HapCompass is available for download at http://www.brown.edu/Research/Istrail_Lab/. SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

MOTIVATION: The understanding of the genetic determinants of complex disease is undergoing a paradigm shift. Genetic heterogeneity of rare mutations with deleterious effects is more commonly being viewed as a major component of disease. Autism is an excellent example where research is active in identifying matches between the phenotypic and genomic heterogeneities. A considerable portion of autism appears to be correlated with copy number variation, which is not directly probed by single nucleotide polymorphism (SNP) array or sequencing technologies. Identifying the genetic heterogeneity of small deletions remains a major unresolved computational problem partly due to the inability of algorithms to detect them. RESULTS: In this article, we present an algorithmic framework, which we term DELISHUS, that implements three exact algorithms for inferring regions of hemizygosity containing genomic deletions of all sizes and frequencies in SNP genotype data. We implement an efficient backtracking algorithm-that processes a 1 billion entry genome-wide association study SNP matrix in a few minutes-to compute all inherited deletions in a dataset. We further extend our model to give an efficient algorithm for detecting de novo deletions. Finally, given a set of called deletions, we also give a polynomial time algorithm for computing the critical regions of recurrent deletions. DELISHUS achieves significantly lower false-positive rates and higher power than previously published algorithms partly because it considers all individuals in the sample simultaneously. DELISHUS may be applied to SNP array or sequencing data to identify the deletion spectrum for family-based association studies. AVAILABILITY: DELISHUS is available at http://www.brown.edu/Research/Istrail_Lab/.