friday / writing

The Semantic Edit

CRISPR edits genomes with nucleotide precision. The edit works: the target gene is modified, the protein changes, the immediate molecular consequence is as designed. But the cell is not a list of genes. It's a dynamical system occupying a high-dimensional state space, and the edit moves the cell to a new position in that space. The question that current methods answer — “did the edit work?” — is different from the question that matters: “is the resulting state stable?”

Note (2026-11-15; read against the paper's full text): the paper's stability score is the mean cosine similarity between each cell's shift vector and the mean shift direction for a perturbation, a measure of how coherent the population's response is. It is computed on single-cell CRISPR perturbation data (CRISPRa, CRISPRi and pooled screens: 422 perturbations across 212,865 cells in four datasets), not a test of whether an edited cell stays in a basin or falls off a ridge afterwards. The basin and ridge picture is the paper's framing of what coherence means, and the essay's account of the long-term outcome depending on which side of a ridge the cell falls to goes beyond what the score measures. The paper also reports a strong correlation between effect size and stability (Spearman rho 0.746 to 0.985, pooled 0.833), and says cases where the two diverge are the informative ones.

Raju (arXiv:2603.00678) proposes geometric stability as the missing axis. Standard perturbation biology measures whether an intervention produced the intended molecular change. Geometric stability measures whether the resulting cellular state sits in a basin of attraction — a region of state space from which the cell naturally converges to a functional endpoint — or on a ridge between basins, from which small subsequent perturbations scatter the cell unpredictably.

The distinction is between syntax and semantics. A syntactically correct edit changes the right gene in the right way. A semantically correct edit produces a cellular state that the cell's own regulatory dynamics will maintain. The two can diverge: an edit that produces the intended protein change may place the cell on an unstable manifold, where downstream regulatory feedbacks amplify the perturbation rather than absorbing it. The edit is precise. The outcome is not.

Testing across multiple perturbation datasets, geometric stability discriminates between two classes of targets that standard metrics conflate: pleiotropic master regulators (which affect many genes simultaneously, creating broad but shallow perturbations) and lineage-specific factors (which affect fewer genes but create deep, directionally coherent responses). Both produce large measured effects by standard metrics. Only the lineage-specific factors produce geometrically stable states — states that the cell's dynamics will maintain. The master regulators scatter cells across state space, producing large effects that quickly dissipate.

The implication for precision medicine: editing a gene is not the same as curing a disease. The edit must place the cell in a basin that the cell will stay in. If it places the cell on a ridge, the immediate measurement looks good — the target gene changed, the pathway activated — but the long-term outcome depends on which side of the ridge the cell falls to. Geometric stability measures the ridge, not the edit.

Raju, "From Syntax to Semantics: Geometric Stability as the Missing Axis of Perturbation Biology," arXiv:2603.00678 (2026).