Everyone
Everyone builds bigger models to predict what a gene knockdown does.
Core trick
Separates per-perturbation effect size from sequencing noise using a reliability ceiling, then flags knockdowns whose signal falls below an SNR threshold.
Tags
Description
Everyone builds bigger models to predict what a gene knockdown does. I asked the prior question: how much is even predictable? Using a genome-scale CRISPRi Perturb-seq screen in primary human CD4⁺ T cells (~22M cells, 4 donors), I ran seven pre-registered investigations, each gated on CPU to fail cheaply before spending GPU.
Six failed — and the failures are the finding. Under honest, reliability-ceiling-calibrated measurement, the per-perturbation signal sits at the noise floor: only ~17% of knockdowns have an effect you can trust (SNR > 3), and reaching reliability would need ~12× more sequencing depth. The one method that worked — a causal do-operator (C2 +0.118/+0.162) — beats its twin only within-distribution; on external causal data it's correlational, not causal.
I wrapped all seven probes into a reliability-ceiling-calibrated predictability audit: a model-agnostic pre-flight scorecard, anchored by a positive control that proves it detects signal when present.
Why it matters: perturbation models exist to prioritize expensive bench experiments. This audit tells a lab what its data can and can't support — before a single experiment runs.
VibeCodingScientist/CD4-Perturbation-Predictability-Audit★ 10 forksPythonMITCreated Jul 8, 2026 · last push Jul 13, 2026
README excerpt
# CD4+ T Cell Perturbation Prediction — Causal + JEPA **Built with Claude: Life Sciences (2026).** This work predicts the transcriptional effect of a CRISPRi gene knockdown in primary human CD4+ T cells, including prediction within an activation state the model has never observed. > "Standard models treat a gene knockdown as an *observation*. This work treats it as an > *intervention*, and that distinction is what enables the model to predict a > knockdown's effect in an activation state it has never observed." This repository comprises a single pipeline whose experimental core is a 2×2 ablation (`JEPA-init × causal-mask`). The causal claim, the do-operator isolation, and the JEPA claim are three cells of that single matrix. Ridge, TabPFN, PseudoBulk-FCN, and Arc State serve as external reference points; a Value-of-Information (VOI) layer converts model disagreement into a sample-efficient experimental-design recommendation. The full technical specification is [`UNIFIED_BUILD_PLAN.md`](docs/UNIFIED_BUILD_PLAN.md). The pre-registered hypotheses are recorded in [`hypotheses.md`](hypotheses.md) and were committed *before* any model saw data. > ## Summary for reviewers > **Scope.** A dataset predictability audit for the Marson genome-scale CD4⁺ CRISPRi Perturb-seq: seven > pre-registered probes together with the do-operator positive control, assembled into a single scorecard that > addresses how much of this dataset is predictable at all, and by what kind of structure. Each