Icon Legend

This session is not in your schedule.

This session is in your schedule. Click again to remove it.

Bioinformatics, Computational and Systems Biology

Bioinformatics, Computational and Systems Biology - Poster Session B

Poster A8 - Dual Deep Neural Network-Based Tools for Genomic Splice Site Prediction Across Multiple Species

Thursday, October 24, 2024

2:45 PM - 3:45 PM EST

Location: Exhibit Hall E, F & G

Presenting Author(s)

Alan Mao (he/him/his)

Student Researcher
Johns Hopkins University
Baltimore, Maryland, United States

Co-Author(s)

Kuan-Hao Chao, MSE

Computer Science PhD Candidate
Johns Hopkins University, United States
MP

Mihaela Pertea, PhD

Associate Professor of Biomedical Engineering and Computer Science
Johns Hopkins University, United States

Primary Investigator(s)

SS

Steven L. Salzberg, PhD (he/him/his)

Bloomberg Distinguished Professor of Computer Science, Biomedical Engineering, and Biostatistics
Johns Hopkins University, United States

Introduction: Recognizing genomic splice sites is a crucial step in accurately assembling gene transcripts, annotating genomes, and identifying protein-coding gene segments. However, ever since the discovery of RNA splicing, our understanding of gene transcription and expression has grown steadily more complex. Alternative splicing (AS), where exons and introns are spliced in different combinations to create multiple gene isoforms, is crucially responsible for the diversity of gene function, but the cellular mechanisms by which such specificity is achieved are incompletely understood. Nonetheless, we know that AS is prevalent in the vast majority of multicellular eukaryotes.

Standard spliced alignment tools like TopHat, SpliceMap, and HISAT2 are capable of aligning RNA-seq reads to a reference genome, thereby revealing the locations where introns have been spliced out. However, they are not always reliable, possibly generating false positives due to alignment errors, as well as transcriptional and splicing noise.

In recent years, convolutional neural networks (CNNs) have emerged as promising tools for recognizing complex patterns such as gene splicing. Several new splice site predictors, including SpliceAI, SpliceFinder, and Spliceator, have adopted CNN-based architectures. Among these methods, SpliceAI is considered to represent the state-of-the-art because it has the highest accuracy. However, its accuracy is largely limited to canonical human transcripts.

We introduce two novel tools which improve upon SpliceAI’s performance across multiple species: (1) Splam, a lightweight and general-purpose deep CNN for fast and accurate prediction; and (2) OpenSpliceAI, an all-purpose PyTorch toolkit that offers versatile retraining and fine-tuning of SpliceAI’s standard model on any given dataset.

Materials and

Methods: Splam was trained using 228,038 splice junctions extracted from protein-coding genes with high alignment support in the MANE and RefSeq databases, and 684,114 false splice junctions created from the opposite DNA strand. In contrast to SpliceAI which uses a 10,000 base pair (bp) context in addition to the full intron, Splam looks at a relatively limited window of 400 bp flanking each splice site, motivated by the observation that the biological process of splicing relies primarily on signals within this window. Additionally, Splam introduces the idea of training the network on donor and acceptor pairs together, based on the principle that the splicing machinery recognizes both ends of each intron at once. Splam’s architecture leverages grouped convolution, consisting of 40 total convolutional layers (Fig. 1B). Splam was trained with batch size 100 for 15 epochs on an Apple M1 Pro chip using PyTorch “mps” mode. To examine generalization ability across both closely-related and distant species, we evaluated Splam and SpliceAI on human, chimpanzee, mouse, and Arabidopsis thaliana (thale cress) test datasets.

OpenSpliceAI features a suite of five modular functions enabling data creation, model training, fine-tuning, high-throughput prediction, and assessment of DNA mutations on cryptic splicing, built completely on the versatile PyTorch framework (as opposed to SpliceAI’s Keras framework). We retrained the SpliceAI model using 80, 400, 2000, and 10,000 bp input contexts, for human (MANE annotation), mouse, zebrafish, honeybee, and thale cress genomes, and compared its accuracy with SpliceAI’s models (trained with older GENCODE annotation).

Results, Conclusions, and Discussions: We find that Splam is consistently more accurate than SpliceAI, with an overall accuracy of 96% at predicting human splice junctions, an improvement of 22% despite SpliceAI having 14–20 times as many base pairs available to make predictions. Additionally, we note that Splam and SpliceAI both perform well on canonical MANE splice sites, but Splam is noticeably more accurate on alternative splice sites (Fig. 1A).

Furthermore, Splam generalizes to non-human species, even distant ones like the flowering plant Arabidopsis thaliana where it achieved a 0.996 AUROC score (Fig. 1C). Across all score thresholds and for all species examined, Splam outperformed SpliceAI in terms of accuracy, sensitivity, and precision.

After using OpenSpliceAI to train various models on mouse, A. thaliana, and human (MANE) datasets, we saw highly concordant performance on human datasets between OpenSpliceAI-trained and SpliceAI models, as expected. However, with the ability to retrain the model on species-specific datasets, OpenSpliceAI-trained models exhibited much improved precision and recall compared to SpliceAI on other species, including mouse (Fig. 1D) and A. thaliana (Fig. 1E).

Splam is a single, lightweight model which shows remarkable generalizability across both closely-related and distant species, and can be easily run on large datasets of aligned RNA-seq reads, such as the outputs from spliced read aligners like STAR and HISAT2. In experiments on these data, we have shown that using Splam to remove low-scoring splice junctions after HISAT2 alignment improved the precision of downstream StringTie assemblies by 5.5-6.5%, a novel application of splice site prediction.

OpenSpliceAI unravels SpliceAI’s model architecture, allowing users to custom train their own models using new data. Rather than scoring existing donor-acceptor pairs like Splam, OpenSpliceAI accepts entire input sequences and identifies donor and acceptor sites de novo using its “predict” tool. Additionally, using the “variant” subcommand, researchers can examine the effects of DNA mutations on splicing patterns.

In summary, Splam and OpenSpliceAI offer two different, but highly accurate and versatile approaches to splice site prediction, both of which improve upon the current state-of-the-art SpliceAI model.

Acknowledgements (Optional):