What's Hiding in the Half of the Transcriptome That Single-Cell 3′ Sequencing Misses

First, a look at what these two labs have already published using full-length sequencing

The first paper captures a population of molecules that “never grew a tail.”

When a gene first starts transcribing, the nascent RNA hasn’t had time to acquire a poly(A) tail yet. Trying to observe nascent transcription at single-cell resolution means trying to catch a population of molecules that are tailless by nature — and mainstream single-cell technology works precisely by grabbing tails. In 2025, Tsinghua University’s Shen Xiaoxia lab published work in Cell proposing a full-length approach called scFLUENT-seq. According to the paper, this made “seeing nascent transcription at single-cell resolution” possible for the first time, and revealed the sparsity and plasticity of genome usage.

The second paper is about tumor subtyping.

The Xie Weiyan lab at Beijing Neurosurgical Institute swapped rulers: instead of expression level, they looked at alternative splicing. In work published in Nature Communications in 2025, they used full-length single-cell transcriptomics to build a validation cohort of 16 samples and 195,720 cells, systematically characterizing splicing dysregulation across every lineage. The result: tumor subtypes that stuck together at the expression level came apart cleanly under the lens of splicing diversity.

One paper is clinical subtyping, the other basic transcriptional biology — worlds apart as scientific questions. But underneath, both rest on the same judgment call:

There are answers hiding in the RNA that mainstream single-cell technology throws away.

The same gene, spliced differently, can produce proteins with completely different — even opposite — functions. A nascent RNA without a tail simply can’t be captured. Why has this information gone unused at single-cell resolution for so long?

The reason is simple: mainstream single-cell technology is blind to it by design.

A trade made a decade ago

In 2017, Zheng et al. published work in Nature Communications that brought single-cell sequencing into the high-throughput era. The design was elegant: use oligo-dT to grab the poly(A) tail of mRNA, sequence only a short stretch at the 3′ end, and de-duplicate with UMIs. From then on, measuring how many genes a single cell expresses became cheap and fast. [3]

But there’s no free lunch.

Poly(A) capture means recognizing only the tail; end sequencing means covering only about a tenth of the transcript. Throughput and cost came down — at the price of two things: the other nine-tenths of the transcript, and every RNA species that never had a poly(A) tail to begin with.

In that missing nine-tenths lives alternative splicing, fusion breakpoints, and mutation sites. In the tailless population lives lncRNA, histone mRNA, nascent transcripts, and viral RNA.

Put simply: what 3′-end sequencing sees is really “half a transcriptome” — and it’s the expression-level half.

The two papers above each land in one of these blind spots — one in the missing nine-tenths, one in the tailless population.

Full-length approaches have been tried before. Smart-seq2 set the gold standard for full-length single-cell sensitivity back in 2013 [4]; Smart-seq3 added UMIs [5]; VASA-seq brought total RNA into view [6]. Beautiful work, all of it — but full-length coverage, high throughput, and a simple workflow have never come together at once.

In 2023, the scFAST-seq methods paper [7] found an answer to the “can’t have all three” problem: semi-random priming.

One-sentence mechanism, effort hidden underneath

Swap oligo-dT for a 12N7K semi-random primer — instead of starting reverse transcription at the poly(A) tail, it starts anywhere along the transcript, paired with template switching to amplify full-length cDNA.

That’s the whole principle in one sentence. But making it actually work required solving three engineering problems.

Random priming doesn’t recognize tails, so naturally it also grabs ribosomal RNA — which makes up the vast majority of a cell’s RNA. The fix: targeted removal with an rRNA blocker.

Also, “detecting the full length” and “covering the full length evenly” are two different things. SeekGene’s own numbers: reads mapping to the gene body’s midsection make up over 55%, and ACTB coverage uniformity exceeds 70%.

And random capture can’t come at the cost of single-cell quantification. Doublet rate is held to 0.3%, cell capture rate reaches up to 65%, 500–12,000 cells per lane, library prep completed in 8 hours.

Data is more convincing. A 2026 head-to-head study in Scientific Reports [8] split the same batch of tumor cells between an international 3′ platform and full-length sequencing: median genes detected per cell, 848 vs. 1,358; median UMI count, 1,254 vs. 2,169; total genes detected, 26,350 vs. 30,664; fraction of reads in cells, 76.00% vs. 91.27%.

Everything an expression atlas needs is still there — full-length information comes as a bonus, from the same workflow.

One change, five kinds of research questions unlocked

Swapping the primer opens up five things that 3′-end sequencing can never deliver, in principle. And each one maps onto a category of research question that used to be off the table — and now isn’t.

1. Alternative splicing: from “is it expressed” to “which version is expressed”

You need full-length coverage to even talk about isoforms. The same gene can produce functionally distinct — sometimes opposite — proteins depending on how it’s spliced; 3′-end sequencing only covers the transcript’s tail end, so it can’t tell these versions apart.

The first paper above is a case in point: pituitary tumor subtypes that stayed lumped together at the expression level came apart under the ruler of splicing diversity. Expression level isn’t the only ruler for subtyping.

→ Solves: Tumor subtypes that won’t separate cleanly, drug-resistance differences between subtypes with no clear explanation — re-examine them through the lens of splicing.

→ Best for: Labs working in oncology, neuroscience, or developmental biology that need mechanistic depth to support high-impact papers or grant proposals.

Representative study: Xie Weiyan’s lab, Beijing Neurosurgical Institute [2].

2. Mutations, fusions, and copy-number variation: matching genotype to phenotype in the same cell

Mutation sites sit in the middle of the transcript: EGFR exon 19 deletions, KRAS G12 — textbook driver mutations — mostly aren’t near the tail, so 3′-end reads never pass through them.

This is why oncology research has long lived with a familiar split: DNA sequencing for mutations, single-cell transcriptomics for cell state, two experiments, two batches of cells. Saying a subpopulation is resistant “because of” a mutation is, strictly speaking, an inference — not a direct observation.

Full-length sequencing closes that gap: in the same batch of cells, which variant a cell carries and what transcriptional state it’s in show up in a single picture. In the head-to-head study, full-length data was used with Numbat to reconstruct a tumor evolution tree made up of four genotype subclones; allelic-imbalance amplifications on chromosomes 1, 5, 9, 12, and 13 were resolved simultaneously, and additional amplification and deletion events invisible to 3′-end data were also detected [8]. For finer mutation detection, targeted enrichment can be layered onto the full-length library to boost coverage depth at specific loci.

→ Solves: Clonal evolution, co-mutation, resistance tracing, minimal residual disease — research questions that technology used to force researchers to work around, now approachable head-on.

→ Best for: Translational research teams in oncology and hematology, hospital oncology/hematology departments, early-stage pharma R&D.

Representative studies: Trofimov et al. [8] — single-cell CNV and subclone analysis using full-length sequencing; Liu Feng’s lab, Shanghai Jiao Tong University [9] — combining full-length sequencing with a targeted panel to connect fusion events, differentiation hierarchy abnormalities, and the immune microenvironment in PML::RARA and CBFα/β fusion AML.

3. Non-poly(A) molecules: bringing “tailless” regulators back into view

lncRNA, histone mRNA, nascent RNA, repeat-element RNA — this tailless population has been a blind spot for poly(A) capture since day one, but makes no difference to random priming.

The first paper above already gave the most extreme example: nascent RNA doesn’t have a poly(A) tail yet, so tail-based capture is structurally powerless there — full-length sequencing let it be seen at single-cell resolution for the first time.

Repeat-element RNA benefits the same way. Work published in Cell Stem Cell in 2025 by Tongji University’s Gao Shaoguang lab [10] used semi-random priming to capture genic and LTR-repeat RNA simultaneously, and resolved how m6A modification on L1PA regulates totipotency.

→ Solves: Upstream regulatory questions about why a gene turns on or off.

→ Best for: Labs doing basic research in epigenetic regulation, transcriptional regulation, and developmental biology.

Representative studies: Tsinghua’s Shen Xiaoxia lab [1] — single-cell nascent transcription; Tongji’s Gao Shaoguang lab [10] — L1PA m6A regulation of totipotency.

4. Viral and prokaryotic transcripts: host–pathogen interaction visible at single-cell resolution

Most viral and bacterial transcripts lack a poly(A) tail. Full-length sequencing can directly answer two questions: which cell types are infected, and what state the infected cells are in.

Parvovirus B19 is a case in point: it was previously only suspected of being linked to hematopoietic failure. Full-length sequencing confirmed at single-cell resolution that it infects hematopoietic stem cells and triggers viral transcription there, driving increased apoptosis and differentiation blocks in hematopoietic stem/progenitor cells.

→ Solves: Viral infection mechanisms, tracing post-transplant complications, intracellular bacteria–host interactions.

→ Best for: Infectious disease, transplant, and hematology teams, and researchers studying host–pathogen interactions.

Representative study: Huang Xiaojun and Zhao Xiangyu’s labs, Peking University [11] — B19V targeting hematopoietic stem cells and disrupting multilineage differentiation.

5. Imperfect samples: nuclei, frozen tissue, and degraded material back in play

The poly(A) tail is one of the most fragile parts of a transcript — in frozen, nucleus-extracted, or slightly degraded samples, the tail goes first. Technology that only recognizes intact tails starts at a disadvantage with these samples; random priming, which can start anywhere along the transcript, is naturally tolerant of incomplete transcripts.

In third-party head-to-head tests on frozen, nucleus-extracted human brain organoids and on liver cancer samples, full-length sequencing delivered consistent results. A large body of “imperfect but precious” real-world clinical samples is back in the research pool.

→ Solves: Reopening the huge stock of imperfect-but-valuable clinical samples — biobanks stop being a gold mine no one can dig into.

→ Best for: Hospital research groups sitting on clinical sample resources: frozen tissue, FFPE pre-processed samples, rare-case samples.

Behind all five questions is the same underlying fact: once reads can start anywhere along the transcript, the information written in the middle, in the tailless RNAs, and in the difficult samples all comes back.

Where this technology is being used

As of July 2026, full-length transcriptomics has supported 16 published Online-first papers, spanning oncology, reproduction and development, neuroscience, and immunology.

Beyond the papers already discussed, a few more are worth noting: Nanjing Medical University’s Zheng Ke lab published work in Science in 2024 [12] mapping RNA-binding proteins across mammalian spermatogenesis — a study whose foundation is exactly this kind of complete transcript capture; Tsinghua University-affiliated Beijing Tsinghua Changgung Hospital’s Dong Jiahong lab published work in Signal Transduction and Targeted Therapy in 2025 [13] pinning down how AKR1B1-driven metabolic reprogramming relates to systemic-therapy resistance in liver cancer; Wuhan University’s Zhang Jinfang lab published work in Nature Cancer in 2026 [14] revealing how CDK10 suppresses antitumor immunity by inhibiting nucleic-acid sensing pathways; Kunming University of Science and Technology’s Ji Weizhi lab published work in Brain in 2026 [15] tracking the link between α-synuclein aggregation and brain atrophy in an SNCA-A53T transgenic monkey model.

Figure 10 | Representative published studies

Following these studies, the space full-length sequencing opens up is concrete: tumor heterogeneity can be dug one layer deeper, down to the isoform level; fusion-gene screening can move from bulk to single-cell; single-cell atlases of non-poly(A) molecules can finally be filled in; genotype and phenotype can be linked in the same cell; and nucleus-extracted, frozen, and lightly degraded samples are back on the table.

At the field level, the value of full-length information is also being validated by other technology routes. Several examples from 2025 alone: in Science Translational Medicine [16], researchers using long-read single-cell sequencing found splicing-dysregulation programs shared across cell states in heterogeneous leukemias, and inhibiting the corresponding splicing regulators rescued mis-splicing and suppressed leukemia cell growth; in Protein & Cell [17], systematic characterization of full-length single-cell isoforms in colorectal cancer pushed splicing regulation into the territory of subtyping and treatment targets; and in Nature Biotechnology, ScISOr-ATAC [18] jointly analyzed full-length single-cell transcriptomes and chromatin accessibility in the same cell, directly observing dual dysregulation of splicing and epigenetic regulation in Alzheimer’s disease.

Different diseases, different scientific questions — landing, in the end, on the same data requirement.

A practical question: what happens to old data

The biggest fear with switching platforms is that old data becomes worthless.

Full-length sequencing covers everything a standard 3′ transcriptome analysis does, and its expression atlas is highly consistent with 3′-end data — in the head-to-head study, cell-type annotations from the two datasets cross-validated each other after integration [8].

Previously generated 3′-end data can be integrated into the same analysis — every dollar of sequencing budget already spent stays useful.

Where the boundary sits: when to choose full-length, and when not to

Full-length sequencing isn’t a cure-all, and its limits deserve as clear an account as its capabilities.

First, the costs. Data volume is larger, cost is higher, and analysis is heavier. Random priming brings in ribosomal RNA background alongside the target RNA, which needs to be filtered out downstream; full-length assembly of low-abundance transcripts can be uneven in coverage; for fine detection of low-frequency mutations, layering on targeted enrichment is recommended. These are practical considerations to plan for at the experimental design and analysis stages.

As for how to choose: if the goal is building a cell atlas, screening differentially expressed genes, or identifying cell-type markers, the 3′-end approach remains the best value for money — no argument there.

But full-length is worth serious consideration if a project involves any of the following: fusion-gene and driver-mutation screening in tumor samples; dynamic regulation of alternative splicing and isoforms; non-poly(A) molecules such as lncRNA, histone mRNA, and nascent transcripts; viral infection and host–pathogen interaction; or clinical samples that are nucleus-extracted, frozen, or lightly degraded.

The principle behind the choice is simple: the research question determines the data requirement, and the data requirement determines the technology. There’s no “best” technology — only the “most suitable” one.

In closing

Looking back, full-length sequencing really does just one thing: it completes the single-cell transcriptome, from “an end” to “the full length.”

But the meaning of length is the meaning of information. Splicing, mutation, fusion, non-poly(A) molecules, viruses, difficult samples — these five capabilities aren’t five separate feature bullet points stacked together. They’re the natural result of one single change: once reads can start anywhere along the transcript, everything written in the middle, in the tailless molecules, and hidden in difficult samples, comes back.

Back to the two papers we opened with. A ruler made of splicing revealed tumor subtypes that had refused to separate. A capture of nascent transcription let tailless molecules be seen for the first time. They’re not exceptions — they’re two answers among a whole set of questions currently being resolved.

More answers are written in the middle of the transcript, waiting for data complete enough to read them.

References