Explainable Deep Learning for Augmentation of Small RNA Expression Profiles

Abstract

The lack of well-structured metadata annotations complicates the re-usability and interpretation of the growing amount of publicly available RNA expression data. The machine learning-based prediction of metadata (data augmentation) can considerably improve the quality of expression data annotation. In this study, we systematically benchmark deep learning (DL) and random forest (RF)-based metadata augmentation of tissue, age, and sex using small RNA (sRNA) expression profiles. We use 4243 annotated sRNA-Seq samples from the small RNA expression atlas (SEA) database to train and test the augmentation performance. In general, the DL machine learner outperforms the RF method in almost all tested cases. The average crossvalidated prediction accuracy of the DL algorithm for tissues is 96.5%, for sex is 77%, and for age is 77.2%.The average tissue prediction accuracy for a completely new dataset is 83.1% (DL) and 80.8% (RF). To understand which sRNAs influence DL predictions, we employ backpropagation-based feature importance scores using the DeepLIFT method, which enable us to obtain information on biological relevnace of sRNAs.

Molecular Image-Based Prediction Models of Nuclear Receptor Agonists and Antagonists Using the DeepSnap-Deep Learning Approach with the Tox21 10K Library

“…These studies try to resolve the issue by calculating the gradient of the input image with respect to the output label and highlighting the target pixel as a recognition target when a slight change in a specific input pixel causes a large change in the output label. However, a simple calculation of the gradient generates a noisy highlight, so some improved methods have been proposed for sharpening [54][55][56][57][58][59]. In addition, in the DeepSnap-DL approach, the performance improves as data size increases, and performance deterioration is observed with insufficient data size or the presence of noise.…”

Interpretation of convolutional neural networks reveals crucial sequence features involving in transcription during fiber development

“…Aiming to interpret models with biological meanings, we applied a published package named DeepLIFT which was previously used in biology field to interpret the models [ 6 , 15 , 17 , 18 ]. We selected all true positive sequences (genes were correctly predicted as expressed) in three stages for interpretation.…”

Explainable deep learning in healthcare: A methodological survey from an attribution view

“…For static analysis (e.g., therapy recommendation based on a fixed set of features), fully connected neural networks are typically utilized for modeling and thus are the target to be interpreted. Commonly used interpretability methods include DeepLIFT (Fiosina et al, 2020), LRP (Li et al, 2018; Zihni et al, 2020), and so on. For time‐series analysis, besides being able to analyze which features are more important or relevant to the prediction among all features used (Yang et al, 2018), it is noteworthy that we can also analyze what temporal patterns are more influential to the final model decision (Mayampurath et al, 2019; Suresh et al, 2017).…”