Alternative splicing could reveal proteins’ hidden functions

A new computational study is turning the spotlight on one of biology’s most difficult questions: why can two protein isoforms produced by the same gene behave so differently? The research, published in Computational Biomedicine, introduces SpliceEM, a machine-learning framework designed to predict isoform-specific functions by treating alternative splicing as a central source of biological information rather than as a secondary transcript-level detail.

Human genes are remarkably flexible. Through alternative splicing, a single gene can generate multiple messenger RNA molecules, each potentially encoding a distinct protein isoform. These variants may differ by only a short sequence segment, yet those small changes can alter where a protein travels inside a cell, which molecules it interacts with, or whether it activates or suppresses a signaling pathway. More than 95 percent of human multi-exon genes are estimated to undergo alternative splicing, making isoform diversity a defining feature of the human proteome.

This complexity has created a major challenge for functional genomics. Sequencing projects have identified millions of transcript isoforms, but experimentally testing the function of every variant would require enormous amounts of time and resources. Many existing computational tools attempt to infer function from protein sequence similarity, gene-level annotations, or known molecular interactions. While these approaches can be powerful, they may overlook the precise splicing events that distinguish one isoform from another, especially when closely related proteins have highly similar sequences.

SpliceEM was developed to address that limitation by combining several biological data types within a single computational framework. The system integrates protein and transcript sequences, Gene Ontology annotations, alternative splicing events, and molecular interaction information. Instead of representing an isoform as an isolated sequence, it places the isoform within a heterogeneous biological graph, a network in which different kinds of entities and relationships can be modeled simultaneously. Isoforms, genes, functional terms, and splicing events can therefore contribute to the same prediction process.

At the core of the framework is a Heterogeneous Graph Transformer, or HGT. This type of neural network is designed to pass information across networks containing multiple node and edge types. In SpliceEM, message passing allows the model to learn how sequence features relate to splicing patterns, how splicing events connect to functional annotations, and how isoforms fit into broader molecular interaction networks. The result is an embedding, or numerical representation, intended to capture the biological characteristics that make one isoform functionally distinct from another.

The researchers also incorporated a teacher-student optimization strategy based on exponential moving averages and multiple-instance learning. In this arrangement, a teacher model provides stable guidance to a student model as the system learns from partially labeled biological data. Multiple-instance learning is particularly useful when annotations are available at the gene level but not for every individual isoform. SpliceEM uses gene-masked asymmetric loss to reduce the risk that the model simply copies gene-level labels onto all transcripts. Hierarchical constraints further encourage predictions to remain consistent with the structured relationships among biological functions.

Across several benchmark datasets, the study reports that adding alternative splicing information substantially improved protein isoform function prediction. The gains were especially notable for functions with limited experimental annotations, a setting in which conventional prediction methods often struggle. By explicitly incorporating splicing events, the framework was better able to separate isoforms originating from the same gene and identify functional differences that may be invisible to sequence-similarity approaches alone.

The model’s learned representations also offered clues about the biology of isoform diversification. Skipped exons and alternative first exons emerged as especially influential event types. Exon skipping can remove a segment of the resulting protein and potentially alter its interaction surfaces or regulatory properties. Alternative first exons may change the beginning of a transcript, affecting protein targeting, cellular localization, or the regulation of transcription itself. In the study, these events were linked to functional divergence and showed strong associations with cancer-related signaling systems, including the MAPK and JAK–STAT pathways.

Those pathways are central to how cells respond to growth signals, inflammation, stress, and developmental cues. Disruptions in their regulation are common in cancer, where abnormal signaling can promote uncontrolled proliferation or help tumor cells evade normal cellular restraints. The findings suggest that relatively localized changes in RNA processing may contribute to much broader changes in cellular behavior. They also reinforce the idea that disease-associated biology may be missed when researchers analyze genes as single units rather than examining the individual isoforms a gene produces.

SpliceEM remains a computational framework, and its newly predicted functions will require experimental validation before they can be considered established biological facts. Nevertheless, the study points toward a more detailed approach to genome interpretation—one that connects RNA processing directly to protein function. As long-read sequencing, single-cell transcriptomics, and disease genomics continue to reveal increasingly complex patterns of isoform usage, tools capable of interpreting this diversity could become valuable for studying disease mechanisms, discovering biomarkers, and prioritizing targets for laboratory investigation. The broader message is clear: alternative splicing does not merely expand the number of transcripts in a cell; it may provide a functional language that helps explain how one gene can support many distinct biological outcomes.

Web References: https://doi.org/10.70401/cbm.2026.0019; Computational Biomedicine

References: Gu T and Wang J, “Isoform function prediction via knowledge distillation from alternative splicing,” Computational Biomedicine, DOI: 10.70401/cbm.2026.0019.