Human Splice-Site Prediction with Deep Neural Networks

An automated framework for evaluation of deep learning models for splice site predictions

A novel framework for the automated evaluation of various deep learning-based splice site detectors is presented. The framework eliminates time-consuming development and experimenting activities for different codebases, architectures, and configurations to obtain the best models for a given RNA splice site dataset. RNA splicing is a cellular process in which pre-mRNAs are processed into mature mRNAs and used to produce multiple mRNA transcripts from a single gene sequence. Since the advancement of sequencing technologies, many splice site variants have been identified and associated with the diseases. So, RNA splice site prediction is essential for gene finding, genome annotation, disease-causing variants, and identification of potential biomarkers. Recently, deep learning models performed highly accurately for classifying genomic signals. Convolutional Neural Network (CNN), Long Short-Term Memory (LSTM) and its bidirectional version (BLSTM), Gated Recurrent Unit (GRU), and its bidirectional version (BGRU) are promising models. During genomic data analysis, CNN's locality feature helps where each nucleotide correlates with other bases in its vicinity. In contrast, BLSTM can be trained bidirectionally, allowing sequential data to be processed from forward and reverse directions. Therefore, it can process 1-D encoded genomic data effectively. Even though both methods have been used in the literature, a performance comparison was missing. To compare selected models under similar conditions, we have created a blueprint for a series of networks with five different levels. As a case study, we compared CNN and BLSTM models' learning capabilities as building blocks for RNA splice site prediction in two different datasets. Overall, CNN performed better with [Formula: see text] accuracy ([Formula: see text] improvement), [Formula: see text] F1 score ([Formula: see text] improvement), and [Formula: see text] AUC-PR ([Formula: see text] improvement) in human splice site prediction. Likewise, an outperforming performance with [Formula: see text] accuracy ([Formula: see text] improvement), [Formula: see text] F1 score ([Formula: see text] improvement), and [Formula: see text] AUC-PR ([Formula: see text] improvement) is achieved in C. elegans splice site prediction. Overall, our results showed that CNN learns faster than BLSTM and BGRU. Moreover, CNN performs better at extracting sequence patterns than BLSTM and BGRU. To our knowledge, no other framework is developed explicitly for evaluating splice detection models to decide the best possible model in an automated manner. So, the proposed framework and the blueprint would help selecting different deep learning models, such as CNN vs. BLSTM and BGRU, for splice site analysis or similar classification tasks and in different problems.

Oligonucleotide correction of an intronic TIMMDC1 variant in cells of patients with severe neurodegenerative disorder

TIMMDC1 encodes the Translocase of Inner Mitochondrial Membrane Domain-Containing protein 1 (TIMMDC1) subunit of complex I of the electron transport chain responsible for ATP production. We studied a consanguineous family with two affected children, now deceased, who presented with failure to thrive in the early postnatal period, poor feeding, hypotonia, peripheral neuropathy and drug-resistant epilepsy. Genome sequencing data revealed a known, deep intronic pathogenic variant TIMMDC1 c.597-1340A>G, also present in gnomAD (~1/5000 frequency), that enhances aberrant splicing. Using RNA and protein analysis we show almost complete loss of TIMMDC1 protein and compromised mitochondrial complex I function. We have designed and applied two different splice-switching antisense oligonucleotides (SSO) to restore normal TIMMDC1 mRNA processing and protein levels in patients' cells. Quantitative proteomics and real-time metabolic analysis of mitochondrial function on patient fibroblasts treated with SSOs showed restoration of complex I subunit abundance and function. SSO-mediated therapy of this inevitably fatal TIMMDC1 neurologic disorder is an attractive possibility.

Principles and correction of 5'-splice site selection

In Eukarya, immature mRNA transcripts (pre-mRNA) often contain coding sequences, or exons, interleaved by non-coding sequences, or introns. Introns are removed upon splicing, and further regulation of the retained exons leads to alternatively spliced mRNA. The splicing reaction requires the stepwise assembly of the spliceosome, a macromolecular machine composed of small nuclear ribonucleoproteins (snRNPs). This review focuses on the early stage of spliceosome assembly, when U1 snRNP defines each intron 5’-splice site (5ʹss) in the pre-mRNA. We first introduce the splicing reaction and the impact of alternative splicing on gene expression regulation. Thereafter, we extensively discuss splicing descriptors that influence the 5ʹss selection by U1 snRNP, such as sequence determinants, and interactions mediated by U1-specific proteins or U1 small nuclear RNA (U1 snRNA). We also include examples of diseases that affect the 5ʹss selection by U1 snRNP, and discuss recent therapeutic advances that manipulate U1 snRNP 5ʹss selectivity with antisense oligonucleotides and small-molecule splicing switches.

Spliceator: multi-species splice site prediction using convolutional neural networks

Background

Ab initio prediction of splice sites is an essential step in eukaryotic genome annotation. Recent predictors have exploited Deep Learning algorithms and reliable gene structures from model organisms. However, Deep Learning methods for non-model organisms are lacking.

Results

We developed Spliceator to predict splice sites in a wide range of species, including model and non-model organisms. Spliceator uses a convolutional neural network and is trained on carefully validated data from over 100 organisms. We show that Spliceator achieves consistently high accuracy (89–92%) compared to existing methods on independent benchmarks from human, fish, fly, worm, plant and protist organisms.

Conclusions

Spliceator is a new Deep Learning method trained on high-quality data, which can be used to predict splice sites in diverse organisms, ranging from human to protists, with consistently high accuracy.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12859-021-04471-3.

DASSI: differential architecture search for splice identification from DNA sequences

Background

The data explosion caused by unprecedented advancements in the field of genomics is constantly challenging the conventional methods used in the interpretation of the human genome. The demand for robust algorithms over the recent years has brought huge success in the field of Deep Learning (DL) in solving many difficult tasks in image, speech and natural language processing by automating the manual process of architecture design. This has been fueled through the development of new DL architectures. Yet genomics possesses unique challenges that requires customization and development of new DL models.

Methods

We proposed a new model, DASSI, by adapting a differential architecture search method and applying it to the Splice Site (SS) recognition task on DNA sequences to discover new high-performance convolutional architectures in an automated manner. We evaluated the discovered model against state-of-the-art tools to classify true and false SS in Homo sapiens (Human), Arabidopsis thaliana (Plant), Caenorhabditis elegans (Worm) and Drosophila melanogaster (Fly).

Results

Our experimental evaluation demonstrated that the discovered architecture outperformed baseline models and fixed architectures and showed competitive results against state-of-the-art models used in classification of splice sites. The proposed model - DASSI has a compact architecture and showed very good results on a transfer learning task. The benchmarking experiments of execution time and precision on architecture search and evaluation process showed better performance on recently available GPUs making it feasible to adopt architecture search based methods on large datasets.

Conclusions

We proposed the use of differential architecture search method (DASSI) to perform SS classification on raw DNA sequences, and discovered new neural network models with low number of tunable parameters and competitive performance compared with manually engineered architectures. We have extensively benchmarked DASSI model with other state-of-the-art models and assessed its computational efficiency. The results have shown a high potential of using automated architecture search mechanism for solving various problems in the field of genomics.

EDeepSSP: Explainable deep neural networks for exact splice sites prediction

Splice site prediction is crucial for understanding underlying gene regulation, gene function for better genome annotation. Many computational methods exist for recognizing the splice sites. Although most of the methods achieve a competent performance, their interpretability remains challenging. Moreover, all traditional machine learning methods manually extract features, which is tedious job. To address these challenges, we propose a deep learning-based approach (EDeepSSP) that employs convolutional neural networks (CNNs) architecture for automatic feature extraction and effectively predicts splice sites. Our model, EDeepSSP, divulges the opaque nature of CNN by extracting significant motifs and explains why these motifs are vital for predicting splice sites. In this study, experiments have been conducted on six benchmark acceptors and donor datasets of humans, cress, and fly. The results show that EDeepSSP has outperformed many state-of-the-art approaches. EDeepSSP achieves the highest area under the receiver operating characteristic curve (AUC_ROC) and area under the precision-recall curve (AUC_PR) of 99.32% and 99.26% on human donor datasets, respectively. We also analyze various filter activities, feature activations, and extracted significant motifs responsible for the splice site prediction. Further, we validate the learned motifs of our model against known motifs of JASPAR splice site database.

Complex Data Imputation by Auto-Encoders and Convolutional Neural Networks—A Case Study on Genome Gap-Filling

Missing data imputation has been a hot topic in the past decade, and many state-of-the-art works have been presented to propose novel, interesting solutions that have been applied in a variety of fields. In the past decade, the successful results achieved by deep learning techniques have opened the way to their application for solving difficult problems where human skill is not able to provide a reliable solution. Not surprisingly, some deep learners, mainly exploiting encoder-decoder architectures, have also been designed and applied to the task of missing data imputation. However, most of the proposed imputation techniques have not been designed to tackle “complex data”, that is high dimensional data belonging to datasets with huge cardinality and describing complex problems. Precisely, they often need critical parameters to be manually set or exploit complex architecture and/or training phases that make their computational load impracticable. In this paper, after clustering the state-of-the-art imputation techniques into three broad categories, we briefly review the most representative methods and then describe our data imputation proposals, which exploit deep learning techniques specifically designed to handle complex data. Comparative tests on genome sequences show that our deep learning imputers outperform the state-of-the-art KNN-imputation method when filling gaps in human genome sequences.

Splice sites detection using chaos game representation and neural network

A novel method is proposed to detect the acceptor and donor splice sites using chaos game representation and artificial neural network. In order to achieve high accuracy, inputs to the neural network, or feature vector, shall reflect the true nature of the DNA segments. Therefore it is important to have one-to-one numerical representation, i.e. a feature vector should be able to represent the original data. Chaos game representation (CGR) is an iterative mapping technique that assigns each nucleotide in a DNA sequence to a respective position on the plane in a one-to-one manner. Using CGR, a DNA sequence can be mapped to a numerical sequence that reflects the true nature of the original sequence. In this research, we propose to use CGR as feature input to a neural network to detect splice sites on the NN269 dataset. Computational experiments indicate that this approach gives good accuracy while being simpler than other methods in the literature, with only one neural network component. The code and data for our method can be accessed from this link: https://github.com/thoang3/portfolio/tree/SpliceSites_ANN_CGR.

Predicting the impact of single nucleotide variants on splicing via sequence-based deep neural networks and genomic features

Single nucleotide mutations in exonic regions can significantly affect gene function through a disruption of splicing, and various computational methods have been developed to predict the splicing-related effects of a single nucleotide mutation. We implemented a new method using ensemble learning that combines two types of predictive models: (a) base sequence-based deep neural networks (DNNs) and (b) machine learning models based on genomic attributes. This method was applied to the Massively Parallel Splicing Assay challenge of the Fifth Critical Assessment of Genome Interpretation, in which challenge participants predicted various experimentally-defined exonic splicing mutations, and achieved a promising result. We successfully revealed that combining different predictive models based upon the stacked generalization method led to significant improvement in prediction performance. In addition, whereas most of the genomic features adopted in constructing machine learning models were previously reported, feature values generated with DSSP, a DNN-based splice site prediction tool, were novel and helpful for the prediction. Learning the sequence patterns associated with normal splicing and the change in splicing site probabilities caused by a mutation was presumed to be helpful in predicting splicing disruption.

Assessing predictions of the impact of variants on splicing in CAGI5

Precision medicine and sequence-based clinical diagnostics seek to predict disease risk or to identify causative variants from sequencing data. The Critical Assessment of Genome Interpretation (CAGI) is a community experiment consisting of genotype-phenotype prediction challenges; participants build models, undergo assessment, and share key findings. In the past, few CAGI challenges have addressed the impact of sequence variants on splicing. In CAGI5, two challenges (Vex-seq and MaPSY) involved prediction of the effect of variants, primarily single-nucleotide changes, on splicing. Although there are significant differences between these two challenges, both involved prediction of results from high-throughput exon inclusion assays. Here, we discuss the methods used to predict the impact of these variants on splicing, their performance, strengths, and weaknesses, and prospects for predicting the impact of sequence variation on splicing and disease phenotypes.

Hybrid model for efficient prediction of poly(A) signals in human genomic DNA

Polyadenylation signals (PAS) are found in most protein-coding and some non-coding genes in eukaryotes. Their accurate recognition improves understanding gene regulation mechanisms and recognition of the 3'-end of transcribed gene regions where premature or alternate transcription ends may lead to various diseases. Although different methods and tools for in-silico prediction of genomic signals have been proposed, the correct identification of PAS in genomic DNA remains challenging due to a vast number of non-relevant hexamers identical to PAS hexamers. In this study, we developed a novel method for PAS recognition. The method is implemented in a hybrid PAS recognition model (HybPAS), which is based on deep neural networks (DNNs) and logistic regression models (LRMs). One of such models is developed for each of the 12 most frequent human PAS hexamers. DNN models appeared the best for eight PAS types (including the two most frequent PAS hexamers), while LRM appeared best for the remaining four PAS types. The new models use different combinations of signal processing-based, statistical, and sequence-based features as input. The results obtained on human genomic data show that HybPAS outperforms the well-tuned state-of-the-art Omni-PolyA models, reducing the classification error for different PAS hexamers by up to 57.35% for 10 out of 12 PAS types, with Omni-PolyA models being better for two PAS types. For the most frequent PAS types, 'AATAAA' and 'ATTAAA', HybPAS reduced the error rate by 35.14% and 34.48%, respectively. On average, HybPAS reduces the error by 30.29%. HybPAS is implemented partly in Python and in MATLAB available at https://github.com/EMANG-KAUST/PolyA_Prediction_LRM_DNN.