Skip to main content
Back to timeline
arXivSource publication:

S1Pert stops gene-perturbation models from reciting response frequencies: macro-F1 up to 0.850 across four cell lines

Synopsis

In CRISPR perturbation screens, models can score well by memorizing whether a gene tends to rise or fall historically, without resolving which perturbed target caused it. S1Pert removes statistically inconclusive responses from supervision and caches frozen-encoder representations of targets, query genes and historical evidence offline, leaving only a lightweight decoder for online inference. Across HepG2, Jurkat, RPE1 and K562, its macro-F1 is 0.850, 0.698, 0.785 and 0.769, 5.8 to 6.4 percentage points above the strongest baseline.

Source-provided article image: S1Pert: Scalable Target-Specific Gene Response Prediction with System-One Intelligence
arXiv

Interpretation

S1Pert turns how the same query gene responds under different perturbed targets into a separately assessable prediction target, rather than only an aggregate classification score. Under target-held-out evaluation, a gene's own historical response tendency remains informative across unseen perturbations, so high aggregate metrics do not show that predictions depend on the perturbed target. Over five target-held-out splits per cell line, S1Pert reaches macro-F1 of 0.850, 0.698, 0.785 and 0.769, 5.8 to 6.4 percentage points above the strongest baseline in the table, and ranks first in F1 for all three response classes.

Training labels are statistically adjudicated first, separating verified invariance from insufficient evidence before they enter supervision and historical priors. Conventional differential-expression pipelines label a gene unchanged whenever directional tests miss significance, conflating low-depth sampling noise with biological invariance. On K562, cells of each target and untreated controls are split into two non-overlapping pools; the lower-depth Pool A directly detects only part of the reference differential-expression responses and leaves the rest underpowered, and S1Pert's equivalence adjudication isolates all of these as inconclusive without adding false invariance labels.

Online inference no longer re-encodes full text for every target–gene pair, reusing semantic landmarks cached offline. Encoding shared text pair by pair or generating autoregressive rationales incurs substantial repeated computation across millions of target–gene pairs. The frozen Jev backbone runs one forward pass over unique single-stream inputs, concatenates intermediate landmarks into a 36-token sequence for the upper eight layers, and caches both evidence-present and evidence-masked representations offline; online, a lightweight hierarchical cross-attention decoder produces the output in a single forward pass.

Cross-target pairwise ranking and synchronized historical-evidence dropout together reduce reliance on gene-level response frequencies. With pointwise fitting alone, the model can repeat each gene's historical propensity while ignoring the perturbed target. On K562, permuting the target at inference time lowers macro-F1 from 0.769 to 0.703 and downregulated F1 from 0.602 to 0.489, showing that predictions do use the upstream perturbation identity.

Perspective

The demonstrated scope is CRISPRi directional three-class prediction on four cell lines, HepG2, Jurkat, RPE1 and K562, each modeled independently with training, validation and test partitions split by held-out targets, and all methods using the same adjudicated labels and the same target–gene pairs.

Statistically inconclusive target–gene pairs are excluded from training supervision and primary evaluation, so reported performance covers only adjudicable responses; cross-cell-line cold start falls back to global class totals, whose evidential support is limited.