Saliency methods are used extensively to highlight the importance of input features in model predictions. These methods are mostly used in vision and language tasks, and their applications to time series data is relatively unexplored. In this paper, we set out to extensively compare the performance of various saliency-based interpretability methods across diverse neural architectures, including Recurrent Neural Network, Temporal Convolutional Networks, and Transformers in a new benchmark of synthetic time series data. We propose and report multiple metrics to empirically evaluate the performance of saliency methods for detecting feature importance over time using both precision (i.e., whether identified features contain meaningful signals) and recall (i.e., the number of features with signal identified as important). Through several experiments, we show that (i) in general, network architectures and saliency methods fail to reliably and accurately identify feature importance over time in time series data, (ii) this failure is mainly due to the conflation of time and feature domains, and (iii) the quality of saliency maps can be improved substantially by using our proposed two-step temporal saliency rescaling (TSR) approach that first calculates the importance of each time step before calculating the importance of each feature at a time step.

Recent efforts to improve the interpretability of deep neural networks use saliency to characterize the importance of input features to predictions made by models. Work on interpretability using saliency-based methods on Recurrent Neural Networks (RNNs) has mostly targeted language tasks, and their applicability to time series data is less understood. In this work we analyze saliency-based methods for RNNs, both classical and gated cell architectures. We show that RNN saliency vanishes over time, biasing detection of salient features only to later time steps and are, therefore, incapable of reliably detecting important features at arbitrary time intervals. To address this vanishing saliency problem, we propose a novel RNN cell structure (input-cell attention), which can extend any RNN cell architecture. At each time step, instead of only looking at the current input vector, input-cell attention uses a fixed-size matrix embedding, each row of the matrix attending to different inputs from current or previous time steps. Using synthetic data, we show that the saliency map produced by the input-cell attention RNN is able to faithfully detect important features regardless of their occurrence in time. We also apply the input-cell attention RNN on a neuroscience task analyzing functional Magnetic Resonance Imaging (fMRI) data for human subjects performing a variety of tasks. In this case, we use saliency to characterize brain regions (input features) for which activity is important to distinguish between tasks. We show that standard RNN architectures are only capable of detecting important brain regions in the last few time steps of the fMRI data, while the input-cell attention model is able to detect important brain region activity across time without latter time step biases.

BACKGROUND: Ultra-fast pseudo-alignment approaches are the tool of choice in transcript-level RNA sequencing (RNA-seq) analyses. Unfortunately, these methods couple the tasks of pseudo-alignment and transcript quantification. This coupling precludes the direct usage of pseudo-alignment to other expression analyses, including alternative splicing or differential gene expression analysis, without including a non-essential transcript quantification step. RESULTS: In this paper, we introduce a transcriptome segmentation approach to decouple these two tasks. We propose an efficient algorithm to generate maximal disjoint segments given a transcriptome reference library on which ultra-fast pseudo-alignment can be used to produce per-sample segment counts. We show how to apply these maximally unambiguous count statistics in two specific expression analyses - alternative splicing and gene differential expression - without the need of a transcript quantification step. Our experiments based on simulated and experimental data showed that the use of segment counts, like other methods that rely on local coverage statistics, provides an advantage over approaches that rely on transcript quantification in detecting and correctly estimating local splicing in the case of incomplete transcript annotations. CONCLUSIONS: The transcriptome segmentation approach implemented in Yanagi exploits the computational and space efficiency of pseudo-alignment approaches. It significantly expands their applicability and interpretability in a variety of RNA-seq analyses by providing the means to model and capture local coverage variation in these analyses.

Analysis of differential alternative splicing from RNA-seq data is complicated by the fact that many RNA-seq reads map to multiple transcripts, and that annotated transcripts from a given gene are often a small subset of many possible complete transcripts for that gene. Here we describe Yanagi, a tool which segments a transcriptome into disjoint regions to create a segments library from a complete transcriptome annotation that preserves all of its consecutive regions of a given length L while distinguishing annotated alternative splicing events in the transcriptome. In this paper, we formalize this concept of transcriptome segmentation and propose an efficient algorithm for generating segment libraries based on a length parameter dependent on specific RNA-Seq library construction. The resulting segment sequences can be used with pseudo-alignment tools to quantify expression at the segment level. We characterize the segment libraries for the reference transcriptomes of Drosophila melanogaster and Homo sapiens. Finally, we demonstrate the utility of quantification using a segment library based on an analysis of differential exon skipping in Drosophila melanogaster and Homo sapiens. The notion of transcript segmentation as introduced here and implemented in Yanagi will open the door for the application of lightweight, ultra-fast pseudo-alignment algorithms in a wide variety of analyses of transcription variation.