Skip to content

PerturbFate — mapping convergent regulators of drug resistance

Adjacent methods resource for the FM-to-Virtual-Cells talk. Xu, Lu, Ugurbil et al., "Mapping convergent regulators of melanoma drug resistance by PerturbFate", Nature 2026.

What it is

PerturbFate is a perturbation-screening method that asks which gene perturbations change a cell's fate under drug pressure — specifically, which regulators drive melanoma cells into a drug-resistant state. It pairs a pooled genetic screen with a readout of cell state/fate, so the output is not just "this perturbation kills cells" but "this perturbation routes cells toward resistance." Applied to melanoma drug resistance, it surfaces a set of convergent regulators — distinct perturbations that funnel into the same resistant phenotype.

How it connects to the talk

The talk's whole Act 1 is about one task: perturbation prediction. The 2025 reckoning is the finding that single-cell FMs do not yet beat linear baselines at predicting what a perturbation does. PerturbFate is the same question from the other side of the bench — a wet-lab + computational measurement method, not a model that predicts.

That makes it adjacent in three useful ways:

  1. It is the kind of data the FMs are supposed to learn from. Convergent-regulator maps are exactly the causal structure the reckoning says current sc-FMs fail to encode (Simon & Zou's SAEs find "minimal regulatory logic"). PerturbFate-style screens are a source of ground truth for that structure.
  2. "Convergent regulators" is a causal-transportability story. Multiple perturbations reaching one resistant state is a statement about which causal pathways matter — the framing Virtual Cells Need Context makes theoretically.
  3. It is a cancer paper. For an AACR audience, it is a concrete oncology instance of the perturbation-prediction problem the talk argues FMs have not yet earned.

Where it sits

Not an FM paper — there is no pretraining-once-adapt-many model here — so it is not in the model glossary or the paper map. It belongs on the shelf as the assay-side counterpart to the prediction-side reckoning: if you want to argue an sc-FM is good at perturbation prediction, PerturbFate-style maps are what you would validate against.

Read more

Last updated 2026-05-14.