RNA sequencing allows researchers to measure which genes are active in a biological sample, but looking only at genes can leave out an important layer of information. Many genes can produce several different RNA transcripts through a process called alternative splicing. These transcripts can differ in structure and function, even though they originate from the same gene.

Researchers therefore use differential transcript usage, or DTU, analysis to determine whether cells change which versions of a transcript they produce under different biological conditions. DTU analysis can provide information about RNA regulation that may not be visible when researchers examine gene expression alone.

As long-read RNA sequencing becomes more common in bulk samples, individual cells, and spatially resolved tissues, researchers need reliable ways to determine which DTU analysis methods work best for each type of data.

Chenxing Zhang of the State Key Laboratory of Holistic Integrative Management of Gastrointestinal Cancers and the Department of Immunology at Fourth Military Medical University in Xi’an, China, and colleagues developed a new framework for evaluating DTU methods using real biological RNA sequencing data rather than relying primarily on computer-generated datasets.

A gene is not always used to produce just one RNA molecule. Through alternative splicing, cells can combine different portions of an RNA transcript to create multiple versions, often called transcript isoforms.

The relative amounts of these transcripts can change between tissues, cell types, developmental stages, or disease conditions. Measuring those changes can help researchers understand how RNA processing is regulated.

Several computational methods have been developed to identify differential transcript usage. However, different methods use different statistical approaches, and they do not always identify the same transcripts.

This creates an important question for researchers, how can they determine which method is producing the most biologically meaningful results?

Moving beyond simulated RNA sequencing data

One common way to evaluate computational methods is to create simulated datasets in which researchers already know the expected answer. They can then determine how accurately each method recovers those expected results.

Simulation is useful because the underlying biological truth is known. However, simulated RNA sequencing data may not reproduce all of the complexity found in real biological samples.

Zhang and colleagues developed an alternative evaluation framework grounded in experimental data.

Their approach uses RNA sequencing experiments in which specific RNA-binding proteins, or RBPs, have been knocked out or reduced. RNA-binding proteins interact with RNA molecules and can regulate processes such as alternative splicing, RNA stability, transport, and translation.

Because experimentally validated interactions between particular RBPs and transcripts are available, these datasets provide a biological reference for evaluating whether DTU methods identify transcripts that are expected to change when an RBP is disrupted.

Creating a transcript-level benchmark

The researchers combined RBP knockout or knockdown RNA sequencing data with experimentally validated RBP-transcript interactions to create reference sets of transcripts.

They then introduced a measure called the transcript set enrichment score, or TSES.

Rather than simply asking whether a method identifies a specific transcript as statistically significant, TSES evaluates how highly biologically relevant transcripts appear within a method’s ranked results.

A DTU method that consistently places known RBP-associated transcripts near the top of its results would therefore receive stronger support than a method that places them much lower.

The researchers tested the framework across multiple real RNA sequencing datasets to evaluate its reliability, stability, effectiveness, robustness, and potential sources of bias.

Comparing ten DTU methods

After validating the framework, the researchers used it to evaluate ten representative DTU analysis methods.

The comparison included short-read and long-read RNA sequencing as well as bulk, single-cell, and spatial transcriptomic datasets.

Framework for evaluating DTU methods

(A) Collection of datasets and DTU methods. It includes short-read bulk RNA-seq datasets from RBP KO/KD experiments, as well as long-read bulk RNA-seq, long-read single-cell RNA-seq, and long-read spatial RNA-seq datasets without KO/KD. Ten representative DTU methods are evaluated. (B) Construction of a reference transcript set based on real biological data. It includes the common transcript set (left panel), the regulated transcript set (right panel), and the overlapped transcript set (middle panel). (C) Design of a DTU-specific evaluation metric. The transcript set enrichment score (TSES) is designed to evaluate the enrichment of the reference transcript set along a transcript list ranked by P-value in a DTU result. The score is then normalized using random permutations.

Long-read sequencing is particularly relevant to transcript analysis because it can often sequence much larger portions of RNA molecules than traditional short-read sequencing. This makes it easier to distinguish complete transcript isoforms produced by alternative splicing.

Single-cell RNA sequencing adds another layer by measuring RNA in individual cells rather than averaging signals across many cells. Spatial transcriptomics can go further by preserving information about where RNA molecules are located within a tissue.

The researchers found that DTU methods differed in performance depending on the type of RNA sequencing data being analyzed. Methods that performed well with one type of dataset did not necessarily perform equally well with another.

These results suggest that selecting a DTU method should take the characteristics of the sequencing data into account rather than assuming that one analytical approach will perform equally well across all experiments.

Using transcript changes to study RNA-binding proteins

The researchers also explored another potential application of their framework, identifying changes in RNA-binding protein activity.

Because RBPs help control transcript processing, disrupting an RBP can produce recognizable patterns of transcript-level changes.

The researchers used DTU results to determine whether these patterns could reveal which RBP had been experimentally disrupted. Their transcript-level approach recovered perturbed RBPs more consistently than strategies based only on gene-level differential expression.

This illustrates why transcript-level RNA sequencing analysis can provide information that may be missed when researchers examine genes as single units.

A gene might show relatively little change in its overall expression while still producing substantially different proportions of its individual transcripts.

Improving analysis as RNA sequencing becomes more complex

RNA sequencing is increasingly moving beyond measuring total gene expression. Long-read sequencing, single-cell sequencing, and spatial transcriptomics allow researchers to examine RNA at progressively greater levels of detail.

As those technologies develop, evaluating the computational tools used to interpret their data becomes increasingly important.

The framework developed by Zhang and colleagues provides a way to compare differential transcript usage methods using experimentally grounded biological information. Instead of relying only on simulated datasets, researchers can evaluate whether computational methods prioritize transcripts connected to known biological changes.

The approach could help researchers choose analytical methods that are better suited to particular RNA sequencing datasets while also supporting deeper investigation of alternative splicing, transcript regulation, and RNA-binding protein activity.