Skip Alignment — Infer Straight From the Raw Reads

sc-SPLASH extends reference-free, statistics-first inference to 10x Chromium and Visium data. Rather than aligning to a reference genome, it extracts k-mer pairs from the raw reads and tests whether their distribution depends on the sample. The price is giving up direct gene-level interpretation; what it buys is the ability to see what a reference-based pipeline erases — alleles, paralogues, repeats, and genes that are not in the annotation at all.

~20×faster than Cell Ranger, and that includes the statistics
50×speed-up of BKC preprocessing over UMI-tools
8 GBmemory used, against 70 GB for Cell Ranger
400K+single cells analysed in Tabula Sapiens

Context · Motivation and method · Content 1 / 3

Why scRNA-seq Still Mostly Measures Gene Expression four limits of the existing tools

Alignment bias is hard to avoid, and worst for species without a good reference.
Tools do not talk to each other — splicing, V(D)J and editing each have their own, and the compute adds up.
Barcoded droplet data is enormous — millions of cells share one library, and barcodes and UMIs are error-prone.
The blind spot: alleles, paralogues and repeats — real biological diversity — are erased by a reference-based pipeline.

The Intuition: Anchors and Targets no reference required

The method takes k-mer pairs straight from the raw reads: a fixed k-mer — the anchor — followed by a varying one, the target. For each anchor it builds a table of target counts across samples, then tests whether that distribution depends on the sample. SPLASH had already shown this works on bulk and Smart-seq2 data; this study extends it to barcoded platforms.

Effect size, 0 to 1How separable the two target distributions are.
A closed-form p-valueThe OASIS test gives a closed-form p-value, which keeps multiple testing under control.

A Three-Stage Pipeline from barcoded FASTQ to significant anchors

01BKCbarcode extraction, UMI deduplication
02Contingency tablesparse matrices merge the counts
03OASISp-values with BY correction

The significance threshold is an effect size above 0.2 with p below 0.05, Benjamini–Yekutieli corrected. Post-processing is optional: build an extendor and compare against a reference, Pfam, BLAST or IgBLAST, all automatable from a JSON configuration.

Pivot · Performance and human data · Content 2 / 3

Performance: BKC Preprocessing a parallel C++ rewrite replacing single-threaded Python

Measure	BKC	UMI-tools
Average runtime	165 s	9,272 s, for whitelist and extract only — deduplication not included
Memory	7 GB	—
Extras	Single-base barcode correction, 3-to-1 base packing, SATC output	—

Performance: Against Cell Ranger and STARsolo Tabula Sapiens muscle, donor 1 · 16 threads

Tool	Runtime	Memory	Note
Cell Ranger v8.0.1	2,540 s	70 GB	Alignment only
STARsolo v2.7.10b	491 s	35 GB	Alignment only
sc-SPLASH	106 s	8 GB	Includes the statistical inference

Read that comparison carefully

Cell Ranger and STARsolo only align; Seurat or Scanpy is still needed afterwards for differential expression. sc-SPLASH runs one pipeline through to significant anchors. So the gap in speed is in fact wider than the table shows — but the two are also not producing the same kind of output.

Tabula Sapiens: 400,000 Human Cells across 16 tissues and two donors

Supervised inference with an L1-regularised GLM over cell-type metadata found 555 cell-type-specific genes, including known splicing targets such as RPS24 and MYL6. The same anchors recurred across donors and tissues — RPS24 37 times, MYL6 28. SPLASH is robust to batch effects because target variation is tested as a conditional probability given the anchor: anchor clusters overlapped between donors in the same tissue far more than chance would give, at p below 2.2 × 10⁻¹⁶ for both lung and muscle.

V(D)J: The Highest-Entropy Pfam Domain diversity is exactly what alignment cannot see

Measure	Value	What it means
Entropy	2.16	V-set has the highest entropy of any Pfam domain, far above keratin, histone and the rest
Effect size	0.90	The two target distributions are almost entirely separable — cell specificity is unambiguous
Unalignable	35 / 121	28.9% could not be aligned to the genome — high diversity being precisely the blind spot of reference-based tools
IgBLAST	60,697	in-frame V(D)J sequences, from plasma and B cells across 16 tissues

Resolution · Spatial data, other species, new genes · Content 3 / 3

Visium: A Mitochondrial Double Mutation in Squamous Cell Carcinoma the same pipeline, unmodified

In Visium spatial data from a human cutaneous squamous cell carcinoma, the strongest signal was a CC → TT double mutation in MT-ND4 (effect size 0.757, ChrM 11,413–11,466), expressed mainly but not exclusively in the tumour region — consistent with an early spontaneous mutation in the carcinoma lineage. The second strongest distinguished the keratin paralogues KRT16 and KRT17 (effect size 0.348): the tumour region tends to express KRT17 and normal epithelium KRT16, both being commonly upregulated in this cancer.

Spot resolutionVisium works at spot resolution, but the barcoded framework applies without modification.

A Splicing Difference Conserved Across Species: RPS24 Exon 6 human fetal tissue and the electric eel

Species	Observation
Human	Fetal small intestine, Visium, highest effect size: epithelial cells include the 3-nt microexon (exon 5), stromal cells exclude it.
Electric eel	Main electric organ: electrocytes include exon 6, while the stromal cells of the insulating septa exclude it (Chr11 7,503,567–7,506,064).

Why the correspondence holds

Electrocytes derive from the skeletal muscle lineage — one of the few human cell types that also includes exon 6. The RPS24 exons are homologous between the two, so this is not coincidence but a cell-type distinction conserved across species.

A Non-Model Organism: The Sponge's granny Anchor a sequence that is in no reference

Measure	Value	Note
Entropy	6.2	The highest in the entire Spongilla lacustris dataset
Distinct targets	667	The number of different sequences following one anchor
BLASTn and reference hits	0	None in odSpoLacu1.1 or NCBI — the sequence is in no annotation at all
Cell-type specificity	73	Of the cells expressing granny, 47 (64%) are granulocytes and 12 (16%) amoebocytes

HCR RNA-FISH confirmed that 88% of Acp5⁺ cells are also granny⁺. Manual assembly from PacBio HiFi long reads revealed five Granrep genes — secreted repeat proteins sharing one structure: signal peptide, then 30-bp granny repeats, then a lysine-rich region, then 18-bp C-terminal repeats, with the granny repeats likely O-glycosylated. Twelve days after treatment with LPS or cGAMP, total Granrep expression rose about 1.4-fold, in step with granulocyte markers such as Acp5 — which hints at a role in immune defence.

A Parallel in Another Phylum: the Sea Squirt's YYD Repeat Ciona robusta haemocytes

Ciona's 24-bp tandem repeat anchor is the only repeat anchor with entropy above 4 among aquatic invertebrates, ranging 4.80–5.97 across subsets. The gene consists largely of that 24-bp repeat, carries only a signal peptide, and encodes a secreted protein; most homologues share a conserved YYD motif, presumed to be tyrosine sulfation as in cionin — a clear parallel with the sponge's Granrep. RNA-FISH shows expression mainly in circulating haemocytes in juveniles, peaking at metamorphosis, twelve hours after hatching.

2 + 43Two homologues in the HT genome, plus 43 TBLASTn hits across C. intestinalis and C. savignyi.

What sc-SPLASH Delivers five conclusions

New genes without a reference — Granrep is absent from the Spongilla annotation, and YYD is highly polymorphic across sea squirt species.
One pipeline for droplet and spatial data — 10x Chromium and Visium alike, with no cell labels required.
BKC stands alone — UMI deduplication 50 times faster, droppable into an existing pipeline.
JSON-driven post-processing — extendor through STAR, Bowtie2, Pfam and IgBLAST, fully automated.
Suited to non-model organisms — repeat-rich, highly polymorphic, reference-poor work on immunity and diversity.