A team led by Megan D. Schertzer and Stella H. Park of the New York Genome Center and Columbia University, together with Jiayu Su, Fairlie Reese, Gloria M. Sheynkman and senior author David A. Knowles, believes the field has been answering that question in the wrong way. In a study published in Genome Biology, the researchers propose replacing the crude filtering steps that dominate current pipelines with a single, principled number borrowed from information theory: perplexity. Their central claim is provocative. Instead of discarding rare isoforms as noise, every transcript a gene produces, no matter how scarce, should contribute proportionally to that gene’s measured diversity.

To understand why this matters, it helps to look at how isoform analysis currently works. Long-read sequencing experiments inevitably generate some technical artifacts, spurious transcript models that do not correspond to real RNA molecules. Pipelines such as SQANTI3, which the new study uses to classify transcript structures into quality categories, help researchers separate genuine isoforms from these errors. But after artifact removal, most workflows then apply an expression threshold, keeping only isoforms that exceed some minimum abundance, typically measured in transcripts per million. Those thresholds are, as the authors point out, arbitrary. An isoform sitting just below the cutoff is thrown away, even though it may be a bona fide transcript with a real biological function. The result is a systematic undercounting of diversity that varies from study to study depending on where each team happens to set its cutoff, undermining reproducibility across labs and datasets.

Perplexity offers a way out of this arbitrariness. The concept originates in Shannon entropy, the foundational measure of uncertainty in information theory. If you treat the isoforms of a gene as the possible outcomes of a random draw, weighted by their relative abundances, the Shannon entropy quantifies how uncertain you are about which isoform you will draw next. Exponentiating that entropy yields perplexity, which has an elegant and intuitive interpretation: it is the effective number of isoforms a gene produces. A gene that expresses one dominant transcript with only trace alternatives has a perplexity close to one. A gene that spreads its expression evenly across ten isoforms has a perplexity approaching ten. The metric is closely related to the Hill numbers that ecologists have long used to measure biodiversity, a theoretical connection the authors develop in their supplementary materials, drawing on the framework formalized by Tom Leinster.

The beauty of this formulation is that no isoform is ever discarded. A transcript representing one percent of a gene’s output contributes to the perplexity calculation in proportion to its abundance, exactly as its share of the expression distribution dictates. Rare isoforms nudge the effective count upward; abundant ones dominate it. The measurement is continuous, interpretable and independent of any researcher-chosen threshold. In effect, perplexity converts the messy, long-tailed distribution of isoform abundances that long-read sequencers reveal into a single number that means the same thing in every dataset, every tissue and every laboratory.

To demonstrate that the metric delivers on this promise, the team applied it to a substantial trove of public data: 124 PacBio long-read RNA sequencing datasets from the ENCODE4 project, spanning 55 human cell types. This resource, generated by the Encyclopedia of DNA Elements consortium, captures transcriptomes across a remarkable range of biological contexts, from immune cells to neuronal lineages. After processing the data through their custom pipeline, which includes artifact filtering with SQANTI3 and annotation of open reading frames, the authors computed perplexity for every gene in every sample and asked whether the resulting diversity measurements were interpretable and reproducible.

The answer, according to the study, is yes. Perplexity proved to be a stable and meaningful quantity across genes, regulatory levels and tissues. Genes whose isoform repertoires are governed by well-characterized regulatory mechanisms showed diversity profiles consistent with known biology, while the metric distinguished cell types in ways that reflect their underlying splicing and transcription programs. Because the calculation uses all detected isoforms rather than a thresholded subset, repeated measurements of the same biological condition converge far more reliably than threshold-based counts, addressing one of the most persistent complaints about long-read transcriptomics: that different pipelines, and even slightly different parameters within a single pipeline, can yield noticeably different catalogs of isoforms.

The work also extends beyond simply counting transcripts. The authors analyzed their detected isoforms at the level of open reading frames, the stretches of RNA that are actually translated into protein, and their supplementary analyses include quadrant classifications of open reading frames and tissue-specific assignments. This reflects a growing recognition in the field that isoform diversity matters not merely as a cataloging exercise but because different isoforms can carry different protein-coding potential, potentially producing distinct protein products from the same gene. A diversity metric that faithfully captures the full isoform landscape therefore has direct implications for proteomics and for interpreting the functional consequences of genetic variation.

The implications reach into medicine as well. Alternative splicing is increasingly implicated in cancer, neurological disease and genetic disorders, and dysregulated isoform production is a hallmark of many tumors. Gloria Sheynkman, whose laboratory at the University of Virginia focuses on connecting transcriptomics to proteomics, and David Knowles, whose group at Columbia and the New York Genome Center develops computational methods for genome interpretation, have both founded or advised companies in this space, underscoring the translational interest in reliable isoform quantification. A reproducible, threshold-free measure of isoform diversity could improve the statistical power of studies that compare diseased and healthy tissue, where subtle shifts in splicing regimes may be drowned out by the noise introduced by arbitrary filtering.

There is also a satisfying intellectual symmetry in the approach. Ecologists counting species in a rainforest, linguists measuring the richness of a vocabulary and now genomicists quantifying the diversity of a transcriptome are all grappling with the same mathematical structure: a distribution over many categories, where rare categories are real but hard to handle. Shannon entropy and its exponential, perplexity, provide a common language for all of these problems. By importing that language into transcriptomics, Schertzer, Park and their colleagues have turned a long-standing methodological weakness into an opportunity, giving the field a way to embrace the full, messy richness of the human transcriptome rather than trimming it to fit a threshold. As long-read sequencing continues its march toward routine clinical and research use, metrics like perplexity may well become as standard a vocabulary for RNA biology as expression levels are today.

Subject of Research: An information-theoretic metric for quantifying isoform diversity in the human transcriptome

Article Title: Perplexity as a metric for isoform diversity in the human transcriptome

Article References: Schertzer, M. D., Park, S. H., Su, J., Reese, F., Sheynkman, G. M., & Knowles, D. A. (2026). Perplexity as a metric for isoform diversity in the human transcriptome. Genome Biology. https://doi.org/10.1186/s13059-026-04260-2

Image Credits: AI Generated

DOI: 10.1186/s13059-026-04260-2

Keywords: perplexity, isoform diversity, long-read RNA sequencing, Shannon entropy, Hill numbers, transcriptome, ENCODE4, PacBio, alternative splicing, SQANTI3, open reading frames, Genome Biology

The problem that GLASS addresses arises after the initial alignment step. Long reads, particularly those produced by Oxford Nanopore sequencers, carry relatively high error rates compared with short reads. When these noisy reads are aligned to a genome, the aligner must decide whether apparent gaps represent genuine introns, consistent with known splice junctions, or simply sequencing artifacts. Misjudgments at this stage propagate downstream: transcript assemblers rely on splice junction signals within alignments to reconstruct full-length isoforms, and a single unreliable junction can corrupt an entire reconstructed transcript. The result is a catalog of predicted isoforms that mixes genuine biology with computational noise, undermining studies of gene regulation, disease mechanisms, and genome annotation.

GLASS, which stands for a graph learning algorithm for screening splice-aware alignments, takes aim at this problem with a post-alignment filtering strategy. Rather than modifying the aligners themselves, the method operates on their output, screening the alignments for reads whose splice patterns look suspicious. The central insight of the work is that the relationships between sequencing reads and splicing events form a rich network structure, and that this structure can be exploited by modern graph-based machine learning to distinguish trustworthy alignments from unreliable ones.

Technically, the method constructs what the authors call a Read–AS bipartite graph. In this graph, one set of nodes represents sequencing reads, and the other set represents alternative-splicing events, with edges connecting reads to the splicing events they support. This bipartite structure encodes a powerful form of collective evidence: a read that supports a splicing event also backed by many other reads gains credibility, while a read whose junctions appear in isolation, or in patterns inconsistent with the surrounding network, becomes suspect. Alternative splicing, the process by which a single gene can produce multiple RNA isoforms by joining exons in different combinations, generates exactly the kind of interconnected event landscape that such a graph can capture.

On top of this graph, GLASS deploys a BipartiteGCN model, a graph convolutional neural network adapted for bipartite structures. Graph convolutional networks work by allowing each node to aggregate information from its neighbors, layer by layer, so that a read node gradually accumulates a representation of its local splicing context. After several rounds of message passing, the network produces a classification for each read, flagging those with potentially unreliable or annotation-inconsistent splice patterns. This is a notable departure from conventional filtering approaches, which typically rely on simple per-read heuristics such as alignment scores or the number of supporting bases around a junction. By learning from the graph-wide context, GLASS can detect subtle inconsistencies that local rules would miss.

The authors evaluated GLASS across multiple datasets and several transcript assemblers, the software tools that convert aligned reads into reconstructed transcript models. The results showed a consistent pattern: applying GLASS reduced the burden of annotation-inconsistent splice signals in the alignments and improved the precision of conservative transcript reconstruction. In practical terms, the transcripts that survived GLASS filtering were more likely to be genuine, at the cost of a modest reduction in sensitivity, meaning some true but harder-to-confirm isoforms were also filtered out. This precision-oriented trade-off is a deliberate design choice, reflecting the view that for many applications, a shorter but cleaner catalog of isoforms is more valuable than a longer but noisier one.

Robustness checks formed an important part of the study. The team repeated their analyses with updated assemblers, with alternative genome annotations, and with alignments preprocessed by complementary tools including minisplice and 2passtools, which are designed to improve splice junction detection in long-read data. They also tested the approach on PacBio datasets, confirming that the benefits were not specific to one sequencing platform or preprocessing pipeline. Across these varied conditions, the same trend held: GLASS filtering pushed reconstruction results toward higher precision, supporting the claim that the method captures a general property of alignment quality rather than an artifact of any particular toolchain.

One of the most scientifically careful aspects of the paper is the authors’ explicit clarification of what annotation-inconsistent splice patterns actually mean. A splice junction that does not appear in a reference annotation is not necessarily an error. Long-read sequencing frequently uncovers genuine but previously unannotated isoforms, and some of the inconsistent signals flagged by GLASS may represent real biology that current annotation databases simply do not record. The authors therefore position GLASS not as an oracle that separates truth from falsehood, but as a complementary filtering strategy aimed specifically at conservative transcript reconstruction, where the goal is to build a high-confidence catalog of isoforms rather than to exhaustively discover every possible splice event.

This framing has practical consequences for the field. Genome annotation projects, clinical transcriptomics, and studies of disease-associated isoforms all depend on transcript reconstructions that researchers can trust. False splice junctions embedded in reconstructed transcripts can lead to phantom proteins being predicted, misestimated gene structures, and wasted experimental effort chasing artifacts. A filtering layer like GLASS, which slots in after alignment and before assembly, offers a way to raise the quality floor of these pipelines without requiring users to abandon their preferred aligners or assemblers. Because it is precision-oriented, it is best suited to applications where confidence in each reported isoform matters more than completeness.

The work also illustrates a broader trend in computational genomics: the migration of graph neural networks from social network analysis and recommendation systems into the life sciences. Biological data is inherently relational, whether it involves protein interactions, regulatory networks, or, as here, the web of evidence connecting sequencing reads to splicing events. The Read–AS bipartite graph at the heart of GLASS is a clean example of how a domain-specific structure can be translated into a representation that modern learning algorithms handle naturally. The BipartiteGCN model learns which patterns of shared support and shared inconsistency tend to characterize reliable alignments, effectively distilling the collective judgment of the entire dataset into a per-read quality assessment.

Funded by China’s National Key R&D Program and the National Natural Science Foundation of China, the study comes from a mathematics-oriented research group, underscoring how interdisciplinary the field of genomics has become. As long-read sequencing continues to drop in cost and expand into clinical and population-scale projects, the volume of RNA-seq data demanding careful splice-aware processing will only grow. Methods like GLASS suggest a future in which the noisy, error-prone raw signal of long reads can be progressively refined by layers of intelligent filtering, yielding transcript catalogs accurate enough to support everything from basic gene discovery to the interpretation of disease variants. For now, the authors’ measured conclusion stands: GLASS is a complementary tool, one more safeguard in the long chain of computation that turns raw molecular signals into biological knowledge.

Subject of Research: Graph-learning-based filtering of splice-aware alignments in long-read RNA-seq transcriptomics

Article Title: GLASS: a graph learning algorithm for screening splice-aware alignments of long-read RNA-seq data

Article References: GLASS: a graph learning algorithm for screening splice-aware alignments of long-read RNA-seq data. (n.d.). https://doi.org/10.1186/s12864-026-13334-1

Image Credits: AI Generated

DOI: 10.1186/s12864-026-13334-1

Keywords: long-read RNA-seq, alternative splicing, graph neural network, transcript reconstruction, splice-aware alignment, BipartiteGCN, read classification, genome annotation, Oxford Nanopore, PacBio, bioinformatics, machine learning

The core problem the researchers set out to solve is one that has dogged the genome-editing field for years. When Cas9 is introduced into a fertilized egg at the single-cell stage, as is routine in zebrafish studies, the enzyme may continue to cut DNA after the embryo begins to divide. Each subsequent cut-and-repair event creates a genetically distinct cell lineage, so the animal that develops is not a uniform carrier of one intended edit but a patchwork of different alleles distributed across its tissues. This phenomenon, known as mosaicism, is well recognized, yet characterizing it thoroughly has been difficult, particularly when the events involved are structural variants occurring at low frequency. Short-read sequencing struggles to resolve large insertions, deletions, and rearrangements, and amplification-based methods can distort the true proportions of variant molecules in a sample.

To overcome these limitations, the team turned to amplification-free targeted sequencing on the PacBio platform using the PureTarget approach, which enriches predefined genomic regions without polymerase chain reaction. Thirty samples from pooled zebrafish larvae and individual adult fish were sequenced successfully, yielding an average coverage of more than 1,100-fold across the target loci. Critically, the reads achieved a consensus accuracy of QV39, meaning fewer than one error per billion bases, which the authors describe as near-perfect. Every read originated from a unique DNA molecule, so the depth of coverage translated directly into an unbiased census of the alleles present. The two haplotypes of each target locus showed balanced depth of coverage, a sign that the method captures both maternal and paternal chromosomes without preference.

The contrast with conventional approaches was striking. When the researchers applied long-range PCR to the same samples before sequencing, the resulting data were skewed, with certain alleles overrepresented and others underrepresented relative to the amplification-free measurements. This matters because many laboratories still rely on PCR-based amplicon sequencing to verify their editing experiments, and the new results indicate that such workflows can systematically misrepresent the complexity of what is actually present in an edited animal. By removing the amplification step entirely and reading native DNA molecules directly, PureTarget preserves the true variant allele frequencies, allowing even rare structural variants to be detected and quantified with confidence.

The biological findings were equally significant. Analysis of individual founder fish, the F0 animals injected at the single-cell embryo stage, revealed extensive mosaicism. A single adult founder carried up to 18 distinct on-target editing events at the intended locus, and up to 11 different off-target events elsewhere in the genome. In other words, one fish that might superficially appear to carry a simple knockout mutation was in fact a living mosaic of dozens of different indels, deletions, inversions, and other rearrangements, each present in a different fraction of its cells. Many of these events would be invisible to standard genotyping PCR or short-read sequencing, which typically only confirms the presence or absence of the intended edit.

Perhaps the most consequential discovery concerned inheritance. When the founders were bred, several editing outcomes, including large structural variants and off-target mutations, were transmitted to the F1 generation. Among juvenile siblings derived from a single founder pair, the researchers identified as many as seven unique editing events. This confirms that mosaicism was present not only in the somatic tissues of the founders but also in their germ cells, meaning that different gametes carried different edited alleles. The practical implication is profound: some consequences of CRISPR-Cas9 editing may only become apparent in the second generation, so a founder that screens negative for off-target mutations in a blood or fin biopsy could still be silently transmitting unexpected variants to its offspring.

For the zebrafish community, which generates thousands of edited lines each year for studies of development, disease modeling, and gene function, these findings argue for more rigorous genotyping strategies. A line established from a mosaic founder may harbor multiple distinct alleles that segregate in subsequent generations, confounding phenotypic interpretation if researchers assume all carriers share an identical mutation. Structural variants that remove or invert large genomic segments can affect neighboring genes, and off-target mutations that hitchhike through the germline could introduce unrelated phenotypes. The study suggests that careful outcrossing and thorough molecular characterization of early generations remain essential, and that the depth and accuracy of the method used for that characterization determine what can actually be seen.

Beyond counting mutations, the platform also allowed the team to interrogate epigenetic signals embedded in the same sequencing data. PacBio reads carry kinetic information that can be used to infer DNA methylation, and the researchers analyzed 5-methylcytosine levels at CpG dinucleotides across the targeted regions in their edited samples. Notably, they did not observe altered 5mC CpG levels in the genome-edited fish compared with controls, providing some reassurance that Cas9 cutting at these loci did not perturb local methylation patterns. This ability to profile both genetic mosaicism and DNA methylation from a single amplification-free dataset illustrates the versatility of the approach for pre-defined genomic regions of interest.

The technical achievement rests on a combination of factors: amplification-free library preparation that avoids PCR-induced bias, targeted enrichment that concentrates sequencing capacity on the loci that matter, and the high intrinsic accuracy of PacBio HiFi reads, here pushed to exceptional consensus quality over the target regions. The authors note that the work was supported by a technology development project grant from SciLifeLab’s National Genomics Infrastructure, with library preparation performed in Uppsala and sequencing carried out at Clinical Genomics Uppsala. Some reagents were provided by PacBio, and one author received conference travel support from the company, disclosures that are stated in the paper. All zebrafish experiments were approved by the Uppsala University Ethical Committee for Animal Research.

The broader message for the genome-editing field is that the true complexity of CRISPR outcomes has likely been underestimated, not because the events are rare but because the tools used to detect them were not up to the task. As CRISPR-based therapies advance through clinical trials and edited animal models underpin countless studies, the ability to comprehensively profile mosaicism, structural variation, and off-target effects with single-molecule accuracy becomes a matter of scientific validity and safety. This study demonstrates that near-perfect long-read sequencing of targeted regions can provide that visibility, and it warns that what happens in the cutting bowl does not stay in the founder: some of CRISPR’s unintended edits can echo quietly into the next generation before anyone notices.

Subject of Research: Characterization of CRISPR-Cas9 editing outcomes and genetic mosaicism using near-perfect targeted long-read sequencing in zebrafish

Article Title: Accurate characterization of CRISPR-Cas9 genome editing outcomes and mosaicism with near-perfect long reads

Article References: Höijer, I., van Schendel, R., Emmanouilidou, A., Östlund, R., Bunikis, I., Tijsterman, M., den Hoed, M., & Ameur, A. (2026). Accurate characterization of CRISPR-Cas9 genome editing outcomes and mosaicism with near-perfect long reads. Genome Medicine, 18(1), Article 135. https://doi.org/10.1186/s13073-026-01772-1

Image Credits: AI Generated

DOI: 10.1186/s13073-026-01772-1

Keywords: CRISPR-Cas9, genome editing, genetic mosaicism, off-target mutations, structural variation, long-read sequencing, PacBio, PureTarget, zebrafish, DNA methylation, targeted sequencing, Genome Medicine