Perturbation-first target discovery

Knock out a regulator.Read what moved, base by base.

SeqletNet runs CRISPR knockouts in tumor-infiltrating CD8 T cells, measures ATAC and RNA in the same cell, and attributes every accessibility change to a specific transcription-factor motif. Our first system, Trim28, has no motif of its own.

See the Trim28 readout Read the preprint →
4,930cellsGSE334878
124,340consensus elementsmm10, 10x Multiome
6T cell statesOT-I CD8 TIL
1knockout, fully mappedTrim28

What a knockout looks like at base-pair resolution

Control above the axis, Trim28-KO mirrored below it, on one shared scale. Change the locus, switch to the difference track, or zoom to 120 bp to see the contribution the model assigns to individual bases.

Illustrative profile, not run output
chr11:69,588,120-69,590,168 · mm10 control 0 · knockout 0
scale 0 to 42 · group autoscale arrows scrub · [ ] locus · - = zoom · d difference
Show the numbers
offset (bp)controlknockoutΔ
Accessibility at the selected locus. Control reaches 42 and knockout reaches 11 at the central element, a 3.8-fold loss over a TCF and LEF motif that the model was never given.

The problem

Trim28 has no motif of its own. That is the whole problem.

Trim28 is recruited to DNA by KRAB zinc-finger proteins. It never binds a sequence directly, so it is invisible to any method that scans for a factor's own motif. Ask a motif-enrichment tool what Trim28 does and it returns nothing, because there is nothing of Trim28 to find.

The signal is entirely indirect: the repeats it stops silencing, and the downstream factors whose binding shifts once it is gone. Recovering that requires attributing each accessibility change to the sequence driving it, which is what a base-resolution model does and a peak-level enrichment cannot.

Why enrichment misses it

  • Enrichment asks which motifs are over-represented in a set of changed peaks. It cannot rank a motif inside a single peak.
  • A peak that stays open but swaps which factor holds it open looks unchanged.
  • Young repeat families are discarded by standard blacklists and unique-mapping filters before the analysis begins.
  • The knockout also shifts how many cells are in each state, so a population-level difference conflates composition with regulation.

Method

Perturb, measure, attribute

Four steps, each with a stated failure mode. Everything runs on mm10 against one frozen consensus peak set shared by both arms.

01

Perturb

CRISPR knockout in antigen-specific CD8 T cells inside the tumor, not in an immortalised line.

B16-OVA · OT-I

02

Measure

Single-cell multiome, so accessibility and expression come from the same cell and are never joined across experiments.

10x Multiome · ATAC + RNA

03

Model

A bias-factorised base-resolution model per condition, sharing one Tn5 bias model validated to contain no factor signal.

ChromBPNet · mm10

04

Attribute

De novo motif discovery from the contribution scores, then per-motif effect sizes compared across conditions.

TF-MoDISco · Fi-NeMo

Result

The factors that changed when Trim28 was removed

Motifs discovered from the data rather than matched from a database. Sort by effect to see what moved most. Every row carries the evidence class it is entitled to, and two are held back deliberately.

Illustrative profile, not run output
motiflogodirectioneffectevidence

Limitations

What we have not shown yet

  • One animal per arm. We report effect sizes and stability across model seeds, never a biological p-value, because a single sample cannot support one.
  • The knockout changes how many cells sit in each state. Until a within-state comparison holds composition fixed, a population difference is not evidence of rewiring.
  • Two factors on the list above sit inside the region our bias model still explains, so they are withheld rather than reported.
  • The repeat story requires quantification that keeps multi-mapping reads. Standard filters delete the evidence before the model sees it.
  • No therapeutic claim. This is target nomination, and nomination is not validation.

Reproducibility

You can rerun this

The dataset is public, the method is written down to the flag level, and the decisions are logged with the evidence for each. We would rather be checked than believed.

If you find something wrong in it, tell us and we will correct the record. That has already happened more than once.

Single-cell multiome of Trim28-knockout and control tumor-infiltrating OT-I CD8 T cellsGEO · GSE334878
Discovery of Tcf7 regulators with clonally-resolved CRISPR screens identifies Trim28 as a mediator of CD8 T cell differentiation in tumorsPreprint, not peer reviewed · bioRxiv 2026.06.14.732162
Method specification and decision ledgerEvery parameter, with its evidence and rigor label

Applications

What this is for

Choosing a target

Two genes give the same expression phenotype. Only one of them changes the regulatory program you care about. Attribution separates them before you commit a campaign.

Reading a failed screen

A hit with no motif of its own looks like noise to enrichment. Reading the sequence recovers what it was doing indirectly.

Designing the next perturbation

If a factor's motif carries the change, the follow-up is a base edit at that motif rather than another whole-gene knockout.

Team

Three people, in one room, in person

Everyone here has both run the experiment and written the model. We are not hiring a wet lab to hand data to a modelling team.

Name

Regulatory genomics

Built the base-resolution differential and the comparability backbone.

Name

Tumor immunology

Ran the clonally-resolved CRISPR screen the Trim28 hit came from.

Name

Machine learning

Attribution correction, motif discovery, and the calibration null.

Contact

Send us a knockout you care about

Tell us the gene and the setting. If reading its regulatory consequence is tractable, we will say so, and if it is not, we will say that instead.

Collaborate

You have a perturbation and chromatin data.

science@seqletnet.com

Join

Genomics, ML, or immunology. Say what you have built.

jobs@seqletnet.com

Invest

Pre-seed. We will show you the limitations first.

hello@seqletnet.com