2026-10-10
Swap-seq quantified 117 endogenous deletions at PPIF. AlphaGenome correlates at Pearson r = 0.76 within 10 kb of the TSS and falls to -0.12 beyond it.
Sequence-to-function models now take in 100 kb to 1 Mb of DNA. What they still lack is a clean measurement of what happens to a nearby gene after a precise deletion in the endogenous genome. Reporter assays such as MPRA and STARR-seq can nominate enhancers, but they do not name the target gene and they do not measure endogenous expression. CRISPRi silences through KRAB and tends to miss enhancers inside the gene body, CTCF sites, and silencers. Dual-guide Cas9 deletions leave indels and inversions, and a pooled screen cannot check what each pair actually did.
PPIF, a regulator of mitochondrial permeability in a disease-associated locus, is estimated to sit among the top 10% of genes for enhancer complexity. Earlier CRISPRi-FlowFISH at this locus in THP-1 monocytes had found five elements.
Swap-seq uses twin prime editing. A pair of pegRNAs deletes 50–250 bp and writes in a 71 bp cassette with an edit-specific 8 bp barcode. The edit does not cut both DNA strands, so unwanted indels are rarer. The effect is read from the barcode sitting in the edited allele, not inferred from guide abundance.
The library covers the PPIF promoter, the first splice junction, and 24 candidate elements: distal enhancers previously hit by CRISPRi, accessible peaks with strong H3K27ac (three of them inside the gene), five CTCF sites, and other accessible regions. Promoters of the neighboring genes ZMIZ1 and ZCCHC24 test for trans effects. Controls are accessible sites on other chromosomes, plus constructs that carry only the barcode cassette and no pegRNA. That is 208 deletions, at least four barcodes each, 860 pegRNA pairs.
Barcodes were filtered with ChromBPNet, a model that predicts THP-1 chromatin accessibility from sequence, so the cassette itself would not accidentally create transcription-factor sites. THP-1 cells with doxycycline-inducible PE2 were infected at low MOI, edited for 14 days, then cultured 7 days without doxycycline so PE2 could degrade. RNA FlowFISH stained PPIF mRNA, cells were sorted into four expression bins, and barcode counts across bins were turned into an effect size by maximum likelihood, with editing efficiency included. In a three-locus pilot, intended deletions of 71–215 bp edited at 3.6–71.5%.
117 of 208 edits passed the frequency filter. Biological replicates correlated at Pearson r = 0.97, technical replicates at 0.90–0.97, and the four barcodes of one edit at r = 0.95. Of 860 barcodes, 505 exceeded 0.001% allele frequency, roughly 1,000 cells each. Promoter deletions cut PPIF expression by 69–90%. Forty-eight edits were significant at BH-corrected p < 0.001, in the promoter, the splice junction, and 16 other elements. Barcode-only controls did nothing. qPCR on homozygous clones matched the pooled readout.
Against CRISPRi at the same locus, effect sizes correlated at r = 0.75 across 18 elements. Swap-seq recovered the four known distal enhancers and added elements CRISPRi had missed, including intronic enhancers g1 and g2. CRISPRi at the gene body generally suppresses transcription, so those enhancers are invisible to it. Homozygous deletion of the CTCF site c2 lowered PPIF by 14.4%, weakened insulation, and reduced contact between the promoter and distal enhancers e1 and e2.
Silencer s1 sits in a ZCCHC24 intron, 57 kb from the PPIF promoter. All five tiled deletions raised PPIF by 25–29%. CRISPRi at the same element raised it by about 8%. In homozygous clones, ZCCHC24 rose 360%. That does not fit a trans story in which s1 acts only by changing ZCCHC24: perturbing the ZCCHC24 promoter also raises PPIF. s1 lacks strong promoter contact and lacks clear H3K27me3 or H3K9me3. In a reporter, s1 repressed the PPIF, EF1A, and FTH1 promoters in either orientation, by up to about 25%. The class IIa HDAC inhibitor TMP269 derepressed the silencer reporter 1.8-fold versus vehicle; an EZH2 inhibitor did not. After endogenous deletion, H3K27ac rose at accessible peaks within 30 kb on either side, while accessibility and H3K27me3 did not change significantly outside the cut. s1 represses nearby enhancers through HDACs, and those enhancers are what reach the distal genes.
Disrupting RUNX and CEBPA sites with 5 bp edits raised PPIF by 3.7–10.5%. Two adjacent 5 bp deletions in a ZNF324-like sequence raised it by up to 17.4%. ChromBPNet had not prioritized that site. Genome-wide, 160 high-accessibility, low-H3K27ac peaks carry at least three of RUNX, GABPA, CEBPA, and ZNF324. Genes next to silencer-like peaks are expressed lower than genes next to enhancer-like peaks (Mann-Whitney p = 2.2×10^-214). Those 160 peaks are nominations, not deletion-tested.
Forty-eight significant edits in 18 elements were scored with AlphaGenome, Enformer, and Borzoi. Overall Pearson r was 0.64–0.78, driven by elements within 10 kb of the TSS.
| Setting | Track | Pearson r |
| <10 kb from TSS, n=20 | AlphaGenome, THP-1 CAGE | 0.76 |
| >10 kb from TSS, n=28 | AlphaGenome, THP-1 CAGE | -0.12 |
| Best distal correlation | Enformer, CD14+ monocyte CAGE | 0.27 |
AlphaGenome CD14+ RNA-seq and Borzoi CD14+ CAGE fall in the same distal band, from -0.12 to 0.27. Placed close to a promoter in a reporter, AlphaGenome captured activation by enhancer e2. Its calls on silencer s1 flipped between repression and activation depending on the promoter and sequence context, and it did not pass either element's effect to the distal endogenous target.
A 1 Mb context window is not evidence that the model uses regulation beyond 10 kb. Pooling proximal edits into one correlation hides that failure. Benchmarks of genomic sequence models need a separate line for distal elements, silencers, and CTCF sites.
On the experimental side, CRISPRi remains the more scalable tool for enhancers and promoters. Deletion is what reaches element classes its mechanism cannot see. s1 would be missed by a search that requires H3K27me3 or H3K9me3. A FlowFISH run still reads one gene, and editing rates can differ by more than 100-fold. Swap-seq is a way to build gold-standard data, not yet a genome-wide screen you run casually.
The paper states three engineering limits. FlowFISH measures one target gene per experiment, though the same edit-and-barcode design can be pointed at sortable phenotypes such as growth or differentiation. Editing efficiency depends on accessibility and histone marks, and the method is unproven in heterochromatin. Efficiency also varies widely; this work used inducible PE2. At a 5% average edit rate and about 20% of cells surviving selection at MOI 0.3, a 200-edit library means infecting at least 200 million cells per biological replicate.
The model result is one locus. The correlation of -0.12 is 28 distal edits at PPIF. AlphaGenome was scored on a THP-1 CAGE track, Enformer and Borzoi on CD14+ monocyte tracks, so the cross-model comparison is not the same output head. The "broader class" of silencers stops at chromatin ratios and motif enrichment. The preprint was posted on 9 October 2026 and has not been peer reviewed.