RNA molecules do not always remain in the same form after they are transcribed from DNA. Before messenger RNA (mRNA) is used to produce proteins, cells can rearrange segments of RNA through a process called alternative splicing. This allows a single gene to generate multiple transcript variants, known as isoforms, greatly increasing biological complexity.

Alternative splicing is essential for normal development and tissue function, but errors in this process are linked to many diseases, including cancer. Predicting how splicing changes across tissues and disease conditions has remained difficult because it depends on many factors working together.

Researchers at the Beijing Institute of Genomics have developed a new artificial intelligence framework called HELIX to better predict these complex splicing patterns.

Schematic overview of the HELIX framework

a, Schematic framework of HELIX. The HELIX model performs prediction at two distinct levels: the splice site level, which is decomposed as baseline and regulation splicing strength, and the transcript level. The baseline module predicts splice site classification (donor or acceptor) and baseline splicing strength of splice sites based on pre-mRNA sequence. The regulation module predicts sample-specific regulation of splicing strength by integrating both pre-mRNA sequence and RBP abundance. The transcript model, using a Bi-LSTM network, predicts isoform usage taking as input the isoform structure and splice site embedding generated by the splice site model. b, Training data attributes. Samples used for the splice site model training (N = 389) were sequenced with short-read technology, where each event corresponds to a splice site with a splicing strength estimation in a specific sample. Samples used for the transcript model training (N = 189) were sequenced in long-read technology, where each event corresponds to an isoform quantified by its relative abundance. The inner ring quantifies sample counts per tissue, while the outer ring represents the corresponding number of splicing or transcript events within each tissue.

HELIX uses hierarchical deep learning to combine two important sources of information: the RNA sequence itself and expression patterns of RNA-binding proteins (RBPs). RBPs help control when and how RNA molecules are processed, making them important regulators of splicing.

One notable feature of HELIX is that it combines information from both short-read and long-read RNA sequencing datasets during training.

Short-read RNA sequencing is widely used and provides large amounts of data, but it can struggle to fully reconstruct complete transcript isoforms. Long-read RNA sequencing captures full-length transcripts more directly, making it easier to identify isoforms and splicing events. By combining both approaches, HELIX achieved better predictive performance than previous methods.

The model showed improved accuracy in several areas, including:

• Predicting differential splicing events
• Estimating splice site strength at highly regulated locations
• Predicting transcript isoform usage
• Identifying tissue-specific splicing quantitative trait loci (sQTLs)

The researchers also applied HELIX to colon cancer patient datasets. The model was able to predict patient-specific splicing abnormalities and connect these changes to both genetic variants and altered RNA-binding protein expression.

Another interesting feature is the model’s ability to adapt to single-cell RNA sequencing data through transfer learning. This allowed HELIX to predict isoforms specific to individual cell types, potentially helping researchers better understand cellular diversity.

As RNA sequencing technologies continue to improve, computational tools like HELIX may become increasingly important for understanding disease mechanisms, identifying biomarkers, and supporting precision medicine approaches.

Availability – The code for using HELIX, as well as the parameters of trained models, are available via GitHub at https://github.com/gyjames/HELIX