Post-hoc explanation methods attempt to make the inner workings of deep neural networks more comprehensible and trustworthy, which otherwise act as black box models. However, since a ground truth is in general lacking, local post-hoc explanation methods, which assign importance scores to input features, are challenging to evaluate. One of the most popular evaluation frameworks is to perturb features deemed important by an explanation and to measure the change in prediction accuracy. Intuitively, a large decrease in prediction accuracy would indicate that the explanation has correctly quantified the importance of features with respect to the prediction outcome (e.g., logits). However, the change in the prediction outcome may stem from perturbation artifacts, since perturbed samples in the test dataset are out of distribution (OOD) compared to the training dataset and can therefore potentially disturb the model in an unexpected manner. To overcome this challenge, we propose feature perturbation augmentation (FPA) which creates and adds perturbed images during the model training. Our computational experiments suggest that FPA makes the considered models more robust against perturbations. Overall, FPA is an intuitive and straightforward data augmentation technique that renders the evaluation of post-hoc explanations more trustworthy Reproducible codes and pre-trained models with FPA are available on Github: https://github.com/lenbrocki/Feature-Perturbation-Augmentation.

MOTIVATION: Single-cell RNA-sequencing (scRNA-seq) allows us to dissect transcriptional heterogeneity arising from cellular types, spatio-temporal contexts and environmental stimuli. Transcriptional heterogeneity may reflect phenotypes and molecular signatures that are often unmeasured or unknown a priori. Cell identities of samples derived from heterogeneous subpopulations are then determined by clustering of scRNA-seq data. These cell identities are used in downstream analyses. How can we examine if cell identities are accurately inferred? Unlike external measurements or labels for single cells, using clustering-based cell identities result in spurious signals and false discoveries. RESULTS: We introduce non-parametric methods to evaluate cell identities by testing cluster memberships in an unsupervised manner. Diverse simulation studies demonstrate accuracy of the jackstraw test for cluster membership. We propose a posterior probability that a cell should be included in that clustering-based subpopulation. Posterior inclusion probabilities (PIPs) for cluster memberships can be used to select and visualize samples relevant to subpopulations. The proposed methods are applied on three scRNA-seq datasets. First, a mixture of Jurkat and 293T cell lines provides two distinct cellular populations. Second, Cell Hashing yields cell identities corresponding to eight donors which are independently analyzed by the jackstraw. Third, peripheral blood mononuclear cells are used to explore heterogeneous immune populations. The proposed P-values and PIPs lead to probabilistic feature selection of single cells that can be visualized using principal component analysis (PCA), t-distributed stochastic neighbor embedding (t-SNE) and others. By learning uncertainty in clustering high-dimensional data, the proposed methods enable unsupervised evaluation of cluster membership. AVAILABILITY AND IMPLEMENTATION: https://cran.r-project.org/package=jackstraw. SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

Evaluating, explaining, and visualizing high-level concepts in generative models, such as variational autoencoders (VAEs), is challenging in part due to a lack of known prediction classes that are required to generate saliency maps in supervised learning. While saliency maps may help identify relevant features (e.g., pixels) in the input for classification tasks of deep neural networks, similar frameworks are understudied in unsupervised learning. Therefore, we introduce a new method of obtaining saliency maps for latent representations of known or novel high-level concepts, often called concept vectors in generative models. Concept scores, analogous to class scores in classification tasks, are defined as dot products between concept vectors and encoded input data, which can be readily used to compute the gradients. The resulting concept saliency maps are shown to highlight input features deemed important for high-level concepts. Our method is applied to the VAE's latent space of CelebA dataset in which known attributes such as "smiles" and "hats" are used to elucidate relevant facial features. Furthermore, our application to spatial transcriptomic (ST) data of a mouse ol-factory bulb demonstrates the potential of latent representations of morphological layers and molecular features in advancing our understanding of complex biological systems. By extending the popular method of saliency maps to generative models, the proposed concept saliency maps help improve interpretability of latent variable models in deep learning.

BACKGROUND: A survey of presences and absences of specific species across multiple biogeographic units (or bioregions) are used in a broad area of biological studies from ecology to microbiology. Using binary presence-absence data, we evaluate species co-occurrences that help elucidate relationships among organisms and environments. To summarize similarity between occurrences of species, we routinely use the Jaccard/Tanimoto coefficient, which is the ratio of their intersection to their union. It is natural, then, to identify statistically significant Jaccard/Tanimoto coefficients, which suggest non-random co-occurrences of species. However, statistical hypothesis testing using this similarity coefficient has been seldom used or studied. RESULTS: We introduce a hypothesis test for similarity for biological presence-absence data, using the Jaccard/Tanimoto coefficient. Several key improvements are presented including unbiased estimation of expectation and centered Jaccard/Tanimoto coefficients, that account for occurrence probabilities. The exact and asymptotic solutions are derived. To overcome a computational burden due to high-dimensionality, we propose the bootstrap and measurement concentration algorithms to efficiently estimate statistical significance of binary similarity. Comprehensive simulation studies demonstrate that our proposed methods produce accurate p-values and false discovery rates. The proposed estimation methods are orders of magnitude faster than the exact solution, particularly with an increasing dimensionality. We showcase their applications in evaluating co-occurrences of bird species in 28 islands of Vanuatu and fish species in 3347 freshwater habitats in France. The proposed methods are implemented in an open source R package called jaccard (https://cran.r-project.org/package=jaccard). CONCLUSION: We introduce a suite of statistical methods for the Jaccard/Tanimoto similarity coefficient for binary data, that enable straightforward incorporation of probabilistic measures in analysis for species co-occurrences. Due to their generality, the proposed methods and implementations are applicable to a wide range of binary data arising from genomics, biochemistry, and other areas of science.

MOTIVATION: There are a number of well-established methods such as principal component analysis (PCA) for automatically capturing systematic variation due to latent variables in large-scale genomic data. PCA and related methods may directly provide a quantitative characterization of a complex biological variable that is otherwise difficult to precisely define or model. An unsolved problem in this context is how to systematically identify the genomic variables that are drivers of systematic variation captured by PCA. Principal components (PCs) (and other estimates of systematic variation) are directly constructed from the genomic variables themselves, making measures of statistical significance artificially inflated when using conventional methods due to over-fitting. RESULTS: We introduce a new approach called the jackstraw that allows one to accurately identify genomic variables that are statistically significantly associated with any subset or linear combination of PCs. The proposed method can greatly simplify complex significance testing problems encountered in genomics and can be used to identify the genomic variables significantly associated with latent variables. Using simulation, we demonstrate that our method attains accurate measures of statistical significance over a range of relevant scenarios. We consider yeast cell-cycle gene expression data, and show that the proposed method can be used to straightforwardly identify genes that are cell-cycle regulated with an accurate measure of statistical significance. We also analyze gene expression data from post-trauma patients, allowing the gene expression data to provide a molecularly driven phenotype. Using our method, we find a greater enrichment for inflammatory-related gene sets compared to the original analysis that uses a clinically defined, although likely imprecise, phenotype. The proposed method provides a useful bridge between large-scale quantifications of systematic variation and gene-level significance analyses. AVAILABILITY AND IMPLEMENTATION: An R software package, called jackstraw, is available in CRAN. CONTACT: [email protected].