Out-of-Distribution (OOD) classification is a domain generalization task in computer vision. Deep learning models are typically developed and tested under the implicit assumption that training and test data are drawn independently and identically distributed (IID) from the same distribution. Overlooking OOD images can lead to poor performance under unseen or adverse viewing conditions, which are common in real-world scenarios. In this work, the proposed solution can be described as a data-driven approach to solve the OOD classification task in computer vision. The proposed approach consists of three stages, a training stage for exploiting labeled source data with different data augmentation strategies using powerful pretrained vision transformer models, an intermediate stage for weighted model ensemble and post-processing strategies, and finally an inference stage for exploiting unlabeled target data by using test-time learning. The proposed data-driven approach enhances the OOD generalization ability of deep models that withstand shifts in nuisances such as shape, pose, context, texture, occlusion, and weather in OOD or rare scenarios. Extensive data-augmentation strategies are used to improve the OOD generalization of deep models across various nuisances. The effectiveness of the proposed approach is evaluated using two standard computer vision benchmarks: ROBIN and a test set provided by the OOD-CV Challenge 2023. The experimental results show that the proposed approach demonstrates a performance improvement of 2.73% the ROBIN test set and achieves accuracy of 94.04% for the Challenge test set in terms of OOD robustness evaluation with classification accuracy. Furthermore, the proposed solution has secured a position within the top three OOD-based rankings on the OOD-CV Challenge Image Classification Leaderboard, 2023.

• Hierarchical supervision in DINOv2 improves the latent space for WBC classification. • Biologically informed hierarchy reduces error severity and aligns features. • The clustering is improved even for non-leukocytic out-of-domain datasets. • Hierarchy supports labels with different degrees of precision. • Hierarchical supervision in DINOv2 training improves generalizability on white blood cell images. The microscopic observation of blood cells is a crucial step in diagnosing pathologies such as leukemia. DINOv2 models have been employed to extract features from blood cell images, but they do not include biological knowledge, nor do they allow multi-granular labels. To enhance the representation of these cells, we propose leveraging a biologically informed hierarchy of white blood cell types. We train a DINOv2-based foundation model with a semi-supervised framework that uses hierarchical supervision. It enables using datasets with varying levels of label precision within a structure that represents the process of cell differentiation. To support multi-level label precision, we modify the original hierarchical loss function, allowing any hierarchy level to serve as a ground truth class. We evaluate our model on three external datasets, including an out-of-domain set of cervical cells. Our approach improves generalization of the model to new datasets, improving by 1 percentage point the balanced accuracy on the two blood cell external datasets, and by 2.5 percentage point the balanced accuracy on the out-of-domain dataset. In addition the proposed strategy better aligns the model’s latent space with biological properties, leading to more acceptable misclassifications

A probabilistic graphical model is proposed, modeling the joint model parameter and multiplier evolution, with a hypervolume based likelihood, promoting multi-objective descent in structural risk minimization. We address multi-objective model parameter optimization via a surrogate single objective penalty loss with time-varying multipliers, equivalent to online scheduling of loss landscape. The multi-objective descent goal is dispatched hierarchically into a series of constraint optimization sub-problems with shrinking bounds according to Pareto dominance. The bound serves as setpoint for the low-level multiplier controller to schedule loss landscapes via output feedback of each loss term. Our method forms closed loop of model parameter dynamic, circumvents excessive memory requirements and extra computational burden of existing multi-objective deep learning methods, and is robust against controller hyperparameter variation, demonstrated on domain generalization tasks.

Feature importance metrics are critical for interpreting machine learning models and understanding the relevance of individual features. However, real-world data often exhibit missingness, thereby complicating how feature importance should be evaluated. We introduce the distinction between two evaluation frameworks under missing data: (1) feature importance under the full data, as if every feature had been fully measured, and (2) feature importance under the observed data, where missingness is governed by the current measurement policy. While the full data perspective offers insights into the data generating process, it often relies on unrealistic assumptions and cannot guide decisions when missingness persists at model deployment. Since neither framework directly informs improvements in data collection, we additionally introduce the feature measurement importance gradient (FMIG), a novel, model-agnostic metric that identifies features that should be measured more frequently to enhance predictive performance. Using synthetic data, we illustrate key differences between these metrics and the risks of conflating them.

Cells must continuously adjust to changing environments and, thus, have evolved mechanisms allowing them to respond to repeated stimuli. While faster gene induction upon a repeated stimulus is known as reinduction memory, responses to repeated repression have been less studied so far. Here, we studied gene repression across repeated carbon source shifts in over 1,500 single Saccharomyces cerevisiae cells. By monitoring the expression of a carbon source-responsive gene, galactokinase 1 (Gal1), and fitting a mathematical model to the single-cell data, we observed a faster response upon repeated repressions at the population level. Exploiting our single-cell data and quantitative modeling approach, we discovered that the faster response is mediated by a shortened repression response delay, the estimated time between carbon source shift and Gal1 protein production termination. Interestingly, we can exclude two alternative hypotheses, i) stronger dilution because of e.g., increased proliferation, and ii) a larger fraction of repressing cells upon repeated repressions. Collectively, our study provides a quantitative description of repression kinetics in single cells and allows us to pinpoint potential mechanisms underlying a faster response upon repeated repression. The computational results of our study can serve as the starting point for experimental follow-up studies.

BACKGROUND: Deep learning contributes to uncovering molecular and cellular processes with highly performant algorithms. Convolutional neural networks have become the state-of-the-art tool to provide accurate and fast image data processing. However, published algorithms mostly solve only one specific problem and they typically require a considerable coding effort and machine learning background for their application. RESULTS: We have thus developed InstantDL, a deep learning pipeline for four common image processing tasks: semantic segmentation, instance segmentation, pixel-wise regression and classification. InstantDL enables researchers with a basic computational background to apply debugged and benchmarked state-of-the-art deep learning algorithms to their own data with minimal effort. To make the pipeline robust, we have automated and standardized workflows and extensively tested it in different scenarios. Moreover, it allows assessing the uncertainty of predictions. We have benchmarked InstantDL on seven publicly available datasets achieving competitive performance without any parameter tuning. For customization of the pipeline to specific tasks, all code is easily accessible and well documented. CONCLUSIONS: With InstantDL, we hope to empower biomedical researchers to conduct reproducible image processing with a convenient and easy-to-use pipeline.

MOTIVATION: Dimensionality reduction is a key step in the analysis of single-cell RNA-sequencing data. It produces a low-dimensional embedding for visualization and as a calculation base for downstream analysis. Nonlinear techniques are most suitable to handle the intrinsic complexity of large, heterogeneous single-cell data. However, with no linear relation between gene and embedding coordinate, there is no way to extract the identity of genes driving any cell's position in the low-dimensional embedding, making it difficult to characterize the underlying biological processes. RESULTS: In this article, we introduce the concepts of local and global gene relevance to compute an equivalent of principal component analysis loadings for non-linear low-dimensional embeddings. Global gene relevance identifies drivers of the overall embedding, while local gene relevance identifies those of a defined sub-region. We apply our method to single-cell RNA-seq datasets from different experimental protocols and to different low-dimensional embedding techniques. This shows our method's versatility to identify key genes for a variety of biological processes. AVAILABILITY AND IMPLEMENTATION: To ensure reproducibility and ease of use, our method is released as part of destiny 3.0, a popular R package for building diffusion maps from single-cell transcriptomic data. It is readily available through Bioconductor. SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

MOTIVATION: Quantitative large-scale cell microscopy is widely used in biological and medical research. Such experiments produce huge amounts of image data and thus require automated analysis. However, automated detection of cell outlines (cell segmentation) is typically challenging due to, e.g. high cell densities, cell-to-cell variability and low signal-to-noise ratios. RESULTS: Here, we evaluate accuracy and speed of various state-of-the-art approaches for cell segmentation in light microscopy images using challenging real and synthetic image data. The results vary between datasets and show that the tested tools are either not robust enough or computationally expensive, thus limiting their application to large-scale experiments. We therefore developed fastER, a trainable tool that is orders of magnitude faster while producing state-of-the-art segmentation quality. It supports various cell types and image acquisition modalities, but is easy-to-use even for non-experts: it has no parameters and can be adapted to specific image sets by interactively labelling cells for training. As a proof of concept, we segment and count cells in over 200 000 brightfield images (1388 × 1040 pixels each) from a six day time-lapse microscopy experiment; identification of over 46 000 000 single cells requires only about two and a half hours on a desktop computer. AVAILABILITY AND IMPLEMENTATION: C ++ code, binaries and data at https://www.bsse.ethz.ch/csd/software/faster.html . CONTACT: [email protected] or [email protected]. SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

UNLABELLED: : Diffusion maps are a spectral method for non-linear dimension reduction and have recently been adapted for the visualization of single-cell expression data. Here we present destiny, an efficient R implementation of the diffusion map algorithm. Our package includes a single-cell specific noise model allowing for missing and censored values. In contrast to previous implementations, we further present an efficient nearest-neighbour approximation that allows for the processing of hundreds of thousands of cells and a functionality for projecting new data on existing diffusion maps. We exemplarily apply destiny to a recent time-resolved mass cytometry dataset of cellular reprogramming. AVAILABILITY AND IMPLEMENTATION: destiny is an open-source R/Bioconductor package "bioconductor.org/packages/destiny" also available at www.helmholtz-muenchen.de/icb/destiny A detailed vignette describing functions and workflows is provided with the package. CONTACT: [email protected] or [email protected] SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

BACKGROUND: Biological data often originate from samples containing mixtures of subpopulations, corresponding e.g. to distinct cellular phenotypes. However, identification of distinct subpopulations may be difficult if biological measurements yield distributions that are not easily separable. RESULTS: We present Multiresolution Correlation Analysis (MCA), a method for visually identifying subpopulations based on the local pairwise correlation between covariates, without needing to define an a priori interaction scale. We demonstrate that MCA facilitates the identification of differentially regulated subpopulations in simulated data from a small gene regulatory network, followed by application to previously published single-cell qPCR data from mouse embryonic stem cells. We show that MCA recovers previously identified subpopulations, provides additional insight into the underlying correlation structure, reveals potentially spurious compartmentalizations, and provides insight into novel subpopulations. CONCLUSIONS: MCA is a useful method for the identification of subpopulations in low-dimensional expression data, as emerging from qPCR or FACS measurements. With MCA it is possible to investigate the robustness of covariate correlations with respect subpopulations, graphically identify outliers, and identify factors contributing to differential regulation between pairs of covariates. MCA thus provides a framework for investigation of expression correlations for genes of interests and biological hypothesis generation.

BACKGROUND: In recent years, high-throughput microscopy has emerged as a powerful tool to analyze cellular dynamics in an unprecedentedly high resolved manner. The amount of data that is generated, for example in long-term time-lapse microscopy experiments, requires automated methods for processing and analysis. Available software frameworks are well suited for high-throughput processing of fluorescence images, but they often do not perform well on bright field image data that varies considerably between laboratories, setups, and even single experiments. RESULTS: In this contribution, we present a fully automated image processing pipeline that is able to robustly segment and analyze cells with ellipsoid morphology from bright field microscopy in a high-throughput, yet time efficient manner. The pipeline comprises two steps: (i) Image acquisition is adjusted to obtain optimal bright field image quality for automatic processing. (ii) A concatenation of fast performing image processing algorithms robustly identifies single cells in each image. We applied the method to a time-lapse movie consisting of ∼315,000 images of differentiating hematopoietic stem cells over 6 days. We evaluated the accuracy of our method by comparing the number of identified cells with manual counts. Our method is able to segment images with varying cell density and different cell types without parameter adjustment and clearly outperforms a standard approach. By computing population doubling times, we were able to identify three growth phases in the stem cell population throughout the whole movie, and validated our result with cell cycle times from single cell tracking. CONCLUSIONS: Our method allows fully automated processing and analysis of high-throughput bright field microscopy data. The robustness of cell detection and fast computation time will support the analysis of high-content screening experiments, on-line analysis of time-lapse experiments as well as development of methods to automatically track single-cell genealogies.

The set of regulatory interactions between genes, mediated by transcription factors, forms a species' transcriptional regulatory network (TRN). By comparing this network with measured gene expression data, one can identify functional properties of the TRN and gain general insight into transcriptional control. We define the subnet of a node as the subgraph consisting of all nodes topologically downstream of the node, including itself. Using a large set of microarray expression data of the bacterium Escherichia coli, we find that the gene expression in different subnets exhibits a structured pattern in response to environmental changes and genotypic mutation. Subnets with fewer changes in their expression pattern have a higher fraction of feed-forward loop motifs and a lower fraction of small RNA targets within them. Our study implies that the TRN consists of several scales of regulatory organization: (1) subnets with more varying gene expression controlled by both transcription factors and post-transcriptional RNA regulation and (2) subnets with less varying gene expression having more feed-forward loops and less post-transcriptional RNA regulation.