The adoption of Vision Transformers (ViTs) in resource-constrained applications necessitates improvements in inference throughput. To this end several token pruning and merging approaches have been proposed that improve efficiency by successively reducing the number of tokens. However, it remains an open problem to design a token reduction method that is fast, maintains high performance, and is applicable to various vision tasks. In this work, we present a token pruner that uses auxiliary prediction heads that learn to select tokens end-to-end based on task relevance. These auxiliary heads can be removed after training, leading to throughput close to that of a random pruner. We evaluate our method on image classification, semantic segmentation, object detection, and instance segmentation, and show speedups of 1.5 – 4x with small drops in performance. As a best case, on the ADE20k semantic segmentation benchmark, we observe a 2x speedup relative to the no-pruning baseline, with a negligible performance penalty of 0.1 median mIoU across 5 seeds.

Machine learning has been widely adopted across sectors, yet its application in medical imaging remains challenging due to distribution shifts in real-world data. Deployed models often encounter samples that differ from the training dataset, particularly in the health domain, leading to performance issues. This limitation hinders the expressiveness and reliability of deep learning models in health applications. Thus, it becomes crucial to identify methods capable of producing reliable uncertainty estimation in the context of distribution shifts in the health sector. In this paper, we explore the feasibility of using cutting-edge Bayesian and non-Bayesian methods to detect distributionally shifted samples, aiming to achieve reliable and trustworthy diagnostic predictions in segmentation task. Specifically, we compare three distinct uncertainty estimation methods, each designed to capture either unimodal or multimodal aspects in the posterior distribution. Our findings demonstrate that methods capable of addressing multimodal characteristics in the posterior distribution, offer more dependable uncertainty estimates. This research contributes to enhancing the utility of deep learning in healthcare, making diagnostic predictions more robust and trustworthy.

Abstract Conformal prediction provides machine learning models with prediction sets that offer theoretical guarantees, but the underlying assumption of exchangeability limits its applicability to time series data. Furthermore, existing approaches struggle to handle multi-step ahead prediction tasks, where uncertainty estimates across multiple future time points are crucial. We propose JANET (Joint Adaptive predictioN-region Estimation for Time-series), a novel framework for constructing conformal prediction regions that are valid for both univariate and multivariate time series. JANET generalises the inductive conformal framework and efficiently produces joint prediction regions with controlled K-familywise error rates, enabling flexible adaptation to specific application needs. Our empirical evaluation demonstrates JANET’s superior performance in multi-step prediction tasks across diverse time series datasets, highlighting its potential for reliable and interpretable uncertainty quantification in sequential data.

Normalising Flows are non-parametric statistical models known for their dual capabilities of density estimation and generation. They are distinguished by their inherently invertible architecture. However, the requirement of invertibility imposes constraints on their expressiveness, necessitating a large number of parameters and innovative architectural designs to achieve satisfactory outcomes. Whilst flow-based models predominantly rely on neural-network-based transformations for expressive designs, alternative transformation methods have received limited attention. In this work, we present Ferumal flow, a novel kernelised normalising flow paradigm that integrates kernels into the framework. Our results demonstrate that a kernelised flow can yield competitive or superior results compared to neural network-based flows whilst maintaining parameter efficiency. Kernelised flows excel especially in the low-data regime, enabling flexible non-parametric density estimation in applications with sparse data availability.

With the development of high-throughput technologies, genomics datasets rapidly grow in size, including functional genomics data. This has allowed the training of large Deep Learning (DL) models to predict epigenetic readouts, such as protein binding or histone modifications, from genome sequences. However, large dataset sizes come at a price of data consistency, often aggregating results from a large number of studies, conducted under varying experimental conditions. While data from large-scale consortia are useful as they allow studying the effects of different biological conditions, they can also contain unwanted biases from confounding experimental factors. Here, we introduce Metadata-guided Feature Disentanglement (MFD)-an approach that allows disentangling biologically relevant features from potential technical biases. MFD incorporates target metadata into model training, by conditioning weights of the model output layer on different experimental factors. It then separates the factors into disjoint groups and enforces independence of the corresponding feature subspaces with an adversarially learned penalty. We show that the metadata-driven disentanglement approach allows for better model introspection, by connecting latent features to experimental factors, without compromising, or even improving performance in downstream tasks, such as enhancer prediction, or genetic variant discovery. The code will be made available at https://github.com/HealthML/MFD.

Normalizing flows are powerful non-parametric statistical models that function as a hybrid between density estimators and generative models. Current learning algorithms for normalizing flows assume that data points are sampled independently, an assumption that is frequently violated in practice, which may lead to erroneous density estimation and data generation. We propose a likelihood objective of normalizing flows incorporating dependencies between the data points, for which we derive a flexible and efficient learning algorithm suitable for different dependency structures. We show that respecting dependencies between observations can improve empirical results on both synthetic and real-world data, and leads to higher statistical power in a downstream application to genome-wide association studies.

MOTIVATION: Existing methods for simulating synthetic genotype and phenotype datasets have limited scalability, constraining their usability for large-scale analyses. Moreover, a systematic approach for evaluating synthetic data quality and a benchmark synthetic dataset for developing and evaluating methods for polygenic risk scores are lacking. RESULTS: We present HAPNEST, a novel approach for efficiently generating diverse individual-level genotypic and phenotypic data. In comparison to alternative methods, HAPNEST shows faster computational speed and a lower degree of relatedness with reference panels, while generating datasets that preserve key statistical properties of real data. These desirable synthetic data properties enabled us to generate 6.8 million common variants and nine phenotypes with varying degrees of heritability and polygenicity across 1 million individuals. We demonstrate how HAPNEST can facilitate biobank-scale analyses through the comparison of seven methods to generate polygenic risk scoring across multiple ancestry groups and different genetic architectures. AVAILABILITY AND IMPLEMENTATION: A synthetic dataset of 1 008 000 individuals and nine traits for 6.8 million common variants is available at https://www.ebi.ac.uk/biostudies/studies/S-BSST936. The HAPNEST software for generating synthetic datasets is available as Docker/Singularity containers and open source Julia and C code at https://github.com/intervene-EU-H2020/synthetic_data.

Machine-learning-based stress detection systems differ with respect to the ground truth used for training the algorithms. It is unclear how models trained on different facets of the stress reaction (e.g., biological, psychological, social) can be compared, interpreted and applied. In this study, we investigate the influence of the stress label on the performance of machine learning models trained on either vocal characteristics or facial expressions extracted from videos. We collected videos from 40 male participants while being exposed to the Trier Social Stress Test (TSST) and assessed self-reported, live observed, video-annotated and neuro-endocrinological stress levels. We train three standard machine learning models to separately predict different stress labels using either voice or facial cues. Analyzing the relationships of different stress facets we found that observers' annotations were significantly positively associated (live vs. video annotated,$\boldsymbol{\rho}_{\mathbf{s}}\ =\ . 53$). Similarly, the neuro-endocrinological stress indices correlated with each other (cortisol vs. sAA,$\boldsymbol{\rho}_{\mathbf{s}}$=.39). Machine learning experiments resulted in predictions that were positively associated with panel-annotated stress levels showing significantly stronger correlations in voice-based models$(\boldsymbol{\rho}_{\mathbf{s}=}.54\ \mathbf{v}\mathbf{s}. \boldsymbol{\rho}_{\mathbf{S}}=.30)$. Predictions of self-reported stress were positively related to ground truth values for face-based$(\boldsymbol{\rho}_{\mathbf{s}}$=.24) but not for voice-based models. There was no evidence for successful predictions of video-annotations or endocrinological stress levels in both settings. We provide evidence that machine learning models trained on different stress assessments perform differently and should be interpreted and applied accordingly. Implications and recommendations for future work on video-based stress detection are discussed.

High annotation costs are a substantial bottleneck in applying modern deep learning architectures to clinically relevant medical use cases, substantiating the need for novel algorithms to learn from unlabeled data. In this work, we propose ContIG, a self-supervised method that can learn from large datasets of unlabeled medical images and genetic data. Our approach aligns images and several genetic modalities in the feature space using a contrastive loss. We design our method to integrate multiple modalities of each individual person in the same model end-to-end, even when the available modalities vary across individuals. Our procedure outperforms state-of-the-art self-supervised methods on all evaluated downstream benchmark tasks. We also adapt gradient-based explainability algorithms to better understand the learned cross-modal associations between the images and genetic modalities. Finally, we perform genome-wide association studies on the features learned by our models, uncovering interesting relationships between images and genetic data.11Source code at: https://github.com/HealthML/ContIG

Gaussian process state-space models describe time series data in a probabilistic and non-parametric manner using a Gaussian process transition function. As inference is intractable, recent methods use variational inference and either rely on simplifying independence assumptions on the approximate posterior or learn the temporal states iteratively. The latter hampers optimization since the posterior over the presence can only be learned once the posterior governing the past has converged. We present a novel inference scheme that applies stochastic variational inference for the Gaussian process posterior and the Laplace approximation on the temporal states. This approach respects the conditional dependencies in the model and, through the Laplace approximation, treats the temporal states jointly, thereby avoiding their sequential learning. Our method is computationally efficient and leads to better calibrated predictions compared to state-of-the art alternatives on synthetic data and on a range of benchmark datasets.

MOTIVATION: Medical images can provide rich information about diseases and their biology. However, investigating their association with genetic variation requires non-standard methods. We propose transferGWAS, a novel approach to perform genome-wide association studies directly on full medical images. First, we learn semantically meaningful representations of the images based on a transfer learning task, during which a deep neural network is trained on independent but similar data. Then, we perform genetic association tests with these representations. RESULTS: We validate the type I error rates and power of transferGWAS in simulation studies of synthetic images. Then we apply transferGWAS in a genome-wide association study of retinal fundus images from the UK Biobank. This first-of-a-kind GWAS of full imaging data yielded 60 genomic regions associated with retinal fundus images, of which 7 are novel candidate loci for eye-related traits and diseases. AVAILABILITY AND IMPLEMENTATION: Our method is implemented in Python and available at https://github.com/mkirchler/transferGWAS/. SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

We propose a two-sample testing procedure based on learned deep neural network representations. To this end, we define two test statistics that perform an asymptotic location test on data samples mapped onto a hidden layer. The tests are consistent and asymptotically control the type-1 error rate. Their test statistics can be evaluated in linear time (in the sample size). Suitable data representations are obtained in a data-driven way, by solving a supervised or unsupervised transfer-learning task on an auxiliary (potentially distinct) data set. If no auxiliary data is available, we split the data into two chunks: one for learning representations and one for computing the test statistic. In experiments on audio samples, natural images and three-dimensional neuroimaging data our tests yield significant decreases in type-2 error rate (up to 35 percentage points) compared to state-of-the-art two-sample tests such as kernel-methods and classifier two-sample tests.

Deep Gaussian Processes learn probabilistic data representations for supervised learning by cascading multiple Gaussian Processes. While this model family promises flexible predictive distributions, exact inference is not tractable. Approximate inference techniques trade off the ability to closely resemble the posterior distribution against speed of convergence and computational efficiency. We propose a novel Gaussian variational family that allows for retaining covariances between latent processes while achieving fast convergence by marginalising out all global latent variables. After providing a proof of how this marginalisation can be done for general covariances, we restrict them to the ones we empirically found to be most important in order to also achieve computational efficiency. We provide an efficient implementation of our new approach and apply it to several benchmark datasets. It yields excellent results and strikes a better balance between accuracy and calibrated uncertainty estimates than its state-of-the-art alternatives.

Self-supervised learning methods have witnessed a recent surge of interest after proving successful in multiple application fields. In this work, we leverage these techniques, and we propose 3D versions for five different self-supervised methods, in the form of proxy tasks. Our methods facilitate neural network feature learning from unlabeled 3D images, aiming to reduce the required cost for expert annotation. The developed algorithms are 3D Contrastive Predictive Coding, 3D Rotation prediction, 3D Jigsaw puzzles, Relative 3D patch location, and 3D Exemplar networks. Our experiments show that pretraining models with our 3D tasks yields more powerful semantic representations, and enables solving downstream tasks more accurately and efficiently, compared to training the models from scratch and to pretraining them on 2D slices. We demonstrate the effectiveness of our methods on three downstream tasks from the medical imaging domain: i) Brain Tumor Segmentation from 3D MRI, ii) Pancreas Tumor Segmentation from 3D CT, and iii) Diabetic Retinopathy Detection from 2D Fundus images. In each task, we assess the gains in data-efficiency, performance, and speed of convergence. Interestingly, we also find gains when transferring the learned representations, by our methods, from a large unlabeled 3D corpus to a small downstream-specific dataset. We achieve results competitive to state-of-the-art solutions at a fraction of the computational expense. We publish our implementations for the developed algorithms (both 3D and 2D versions) as an open-source library, in an effort to allow other researchers to apply and extend our methods on their datasets.

In order to solve large-scale lasso problems, screening algorithms have been developed that discard features with zero coefficients based on a computationally efficient screening rule. Most existing screening rules were developed from a spherical constraint and half-space constraints on a dual optimal solution. However, existing rules admit at most two half-space constraints due to the computational cost incurred by the half-spaces, even though additional constraints may be useful to discard more features. In this paper, we present AdaScreen, an adaptive lasso screening rule ensemble, which allows to combine any one sphere with multiple half-space constraints on a dual optimal solution. Thanks to geometrical considerations that lead to a simple closed form solution for AdaScreen, we can incorporate multiple half-space constraints at small computational cost. In our experiments, we show that AdaScreen with multiple half-space constraints simultaneously improves screening performance and speeds up lasso solvers.

MOTIVATION: Set-based variance component tests have been identified as a way to increase power in association studies by aggregating weak individual effects. However, the choice of test statistic has been largely ignored even though it may play an important role in obtaining optimal power. We compared a standard statistical test-a score test-with a recently developed likelihood ratio (LR) test. Further, when correction for hidden structure is needed, or gene-gene interactions are sought, state-of-the art algorithms for both the score and LR tests can be computationally impractical. Thus we develop new computationally efficient methods. RESULTS: After reviewing theoretical differences in performance between the score and LR tests, we find empirically on real data that the LR test generally has more power. In particular, on 15 of 17 real datasets, the LR test yielded at least as many associations as the score test-up to 23 more associations-whereas the score test yielded at most one more association than the LR test in the two remaining datasets. On synthetic data, we find that the LR test yielded up to 12% more associations, consistent with our results on real data, but also observe a regime of extremely small signal where the score test yielded up to 25% more associations than the LR test, consistent with theory. Finally, our computational speedups now enable (i) efficient LR testing when the background kernel is full rank, and (ii) efficient score testing when the background kernel changes with each test, as for gene-gene interaction tests. The latter yielded a factor of 2000 speedup on a cohort of size 13 500. AVAILABILITY: Software available at http://research.microsoft.com/en-us/um/redmond/projects/MSCompBio/Fastlmm/. CONTACT: [email protected] SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

Multi-task prediction models are widely being used to couple regressors or classification models by sharing information across related tasks. A common pitfall of these models is that they assume that the output tasks are independent conditioned on the inputs. Here, we propose a multi-task Gaussian process approach to model both the relatedness between regressors as well as the task correlations in the residuals, in order to more accurately identify true sharing between regressors. The resulting Gaussian model has a covariance term that is the sum of Kronecker products, for which efficient parameter inference and out of sample prediction are feasible. On both synthetic examples and applications to phenotype prediction in genetics, we find substantial benefits of modeling structured noise compared to established alternatives.

MOTIVATION: Genomic studies have revealed a substantial heritable component of the transcriptional state of the cell. To fully understand the genetic regulation of gene expression variability, it is important to study the effect of genotype in the context of external factors such as alternative environmental conditions. In model systems, explicit environmental perturbations have been considered for this purpose, allowing to directly test for environment-specific genetic effects. However, such experiments are limited to species that can be profiled in controlled environments, hampering their use in important systems such as human. Moreover, even in seemingly tightly regulated experimental conditions, subtle environmental perturbations cannot be ruled out, and hence unknown environmental influences are frequent. Here, we propose a model-based approach to simultaneously infer unmeasured environmental factors from gene expression profiles and use them in genetic analyses, identifying environment-specific associations between polymorphic loci and individual gene expression traits. RESULTS: In extensive simulation studies, we show that our method is able to accurately reconstruct environmental factors and their interactions with genotype in a variety of settings. We further illustrate the use of our model in a real-world dataset in which one environmental factor has been explicitly experimentally controlled. Our method is able to accurately reconstruct the true underlying environmental factor even if it is not given as an input, allowing to detect genuine genotype-environment interactions. In addition to the known environmental factor, we find unmeasured factors involved in novel genotype-environment interactions. Our results suggest that interactions with both known and unknown environmental factors significantly contribute to gene expression variability. AVAILABILITY: and implementation: Software available at http://pmbio.github.io/envGPLVM/. SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

MOTIVATION: Approaches for testing sets of variants, such as a set of rare or common variants within a gene or pathway, for association with complex traits are important. In particular, set tests allow for aggregation of weak signal within a set, can capture interplay among variants and reduce the burden of multiple hypothesis testing. Until now, these approaches did not address confounding by family relatedness and population structure, a problem that is becoming more important as larger datasets are used to increase power. RESULTS: We introduce a new approach for set tests that handles confounders. Our model is based on the linear mixed model and uses two random effects-one to capture the set association signal and one to capture confounders. We also introduce a computational speedup for two random-effects models that makes this approach feasible even for extremely large cohorts. Using this model with both the likelihood ratio test and score test, we find that the former yields more power while controlling type I error. Application of our approach to richly structured Genetic Analysis Workshop 14 data demonstrates that our method successfully corrects for population structure and family relatedness, whereas application of our method to a 15 000 individual Crohn's disease case-control cohort demonstrates that it additionally recovers genes not recoverable by univariate analysis. AVAILABILITY: A Python-based library implementing our approach is available at http://mscompbio.codeplex.com.

MOTIVATION: Exploring the genetic basis of heritable traits remains one of the central challenges in biomedical research. In traits with simple Mendelian architectures, single polymorphic loci explain a significant fraction of the phenotypic variability. However, many traits of interest seem to be subject to multifactorial control by groups of genetic loci. Accurate detection of such multivariate associations is non-trivial and often compromised by limited statistical power. At the same time, confounding influences, such as population structure, cause spurious association signals that result in false-positive findings. RESULTS: We propose linear mixed models LMM-Lasso, a mixed model that allows for both multi-locus mapping and correction for confounding effects. Our approach is simple and free of tuning parameters; it effectively controls for population structure and scales to genome-wide datasets. LMM-Lasso simultaneously discovers likely causal variants and allows for multi-marker-based phenotype prediction from genotype. We demonstrate the practical use of LMM-Lasso in genome-wide association studies in Arabidopsis thaliana and linkage mapping in mouse, where our method achieves significantly more accurate phenotype prediction for 91% of the considered phenotypes. At the same time, our model dissects the phenotypic variability into components that result from individual single nucleotide polymorphism effects and population structure. Enrichment of known candidate genes suggests that the individual associations retrieved by LMM-Lasso are likely to be genuine. AVAILABILITY: Code available under http://webdav.tuebingen. mpg.de/u/karsten/Forschung/research.html. CONTACT: [email protected], [email protected] or [email protected] SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

Inference in matrix-variate Gaussian models has major applications for multi- output prediction and joint learning of row and column covariances from matrix- variate data. Here, we discuss an approach for efficient inference in such models that explicitly account for iid observation noise. Computational tractability can be retained by exploiting the Kronecker product between row and column covariance matrices. Using this framework, we show how to generalize the Graphical Lasso in order to learn a sparse inverse covariance between features while accounting for a low-rank confounding covariance between samples. We show practical utility on applications to biology, where we model covariances with more than 100,000 di- mensions. We find greater accuracy in recovering biological network structures and are able to better reconstruct the confounders.

MOTIVATION: Synthetic lethal interactions represent pairs of genes whose individual mutations are not lethal, while the double mutation of both genes does incur lethality. Several studies have shown a correlation between functional similarity of genes and their distances in networks based on synthetic lethal interactions. However, there is a lack of algorithms for predicting gene function from synthetic lethality interaction networks. RESULTS: In this article, we present a novel technique called kernelROD for gene function prediction from synthetic lethal interaction networks based on kernel machines. We apply our novel algorithm to Gene Ontology functional annotation prediction in yeast. Our experiments show that our method leads to improved gene function prediction compared with state-of-the-art competitors and that combining genetic and congruence networks leads to a further improvement in prediction accuracy.