Cell biology gets foundation models: predicting perturbations, reading RNA structure, and ranking reprogramming regulators

Predicting a cell before perturbing it

In a preprint, GENESIS, short for Generative ENErgy-based Single-cell Intervention Simulator, addresses a central problem in single-cell biology: predicting how a cell will respond to a perturbation before running the experiment. The model outperforms existing generative models and statistical baselines in predicting single-cell response distributions, according to the authors, and reaches predictive accuracy approaching the reproducibility of repeated experimental measurements. That last comparison is the honest yardstick: a prediction cannot be expected to beat the noise between two runs of the same experiment. m:…2506

The preprint also reports a counterintuitive data lesson: broader, lower-depth measurements are more effective for training GENESIS than deep profiling of fewer experimental conditions. If that finding holds, it changes how labs should allocate sequencing budgets, favoring coverage across conditions over depth within a few. The practical implication is a shift in experimental strategy, from measuring a few conditions exhaustively toward measuring many conditions well enough for a model to learn the response surface. m:…2506

Reading RNA structure at genome scale

Two preprints attack RNA structure from complementary directions. In a preprint, ScanFold3 provides a faster way to identify functional RNA secondary structures across genomes and transcriptomes, achieving a 38-fold to 129-fold increase in speed over ScanFold2 for sequences up to 128 kb. The speedup comes from reframing consensus structure generation as a minimum perfect weighted matching problem on a graph, and the authors report that this framing also improves predictive accuracy. m:…7725

In a preprint, LIMBER takes the foundation-model route instead: a hybrid RNA model that integrates bidirectional self-attention with selective state-space modeling for RNA sequence analysis. The authors report leading performance on secondary-structure prediction and non-coding RNA classification benchmarks. They also report that RNA foundation models show a hierarchical, task-dependent organization of biological information across model layers, meaning different layers carry different kinds of biological signal depending on the task. For researchers choosing between a specialized tool and a general model, the two preprints frame the trade-off: ScanFold3 optimizes a specific job, while LIMBER aims to learn representations useful across many RNA tasks. m:…2838

Ranking the regulators of reprogramming

In a preprint, FateDriver addresses cellular reprogramming, where the challenge is identifying which endogenous regulators drive a specific transition from one cell state to another. The model uses a dual-view graph neural network that treats delivered reprogramming factors as sparse gene-level labels, which allows transition-specific prioritization of candidate regulators rather than a single generic ranking. m:…2162

The authors report that FateDriver outperforms six existing computational methods in recovering reference genes for cellular reprogramming across OSKM and GMT datasets. It also identifies a DDX58-interferon axis in cells lacking endogenous pluripotency markers during reprogramming, a concrete biological hypothesis that experimentalists can test. That is the right way to read this kind of preprint: not as a discovery, but as a prioritized list of candidates with a stated reason to check them first. m:…2162

What to watch

Across all four preprints, the pattern is prediction before experiment: a model proposes a response, a structure, or a regulator ranking, and the lab decides where to spend its next assay. None of these preprints has yet passed peer review, and several note the absence of public code or model weights in their abstracts, so independent verification is still open. The claims that deserve the closest scrutiny are the strongest ones: accuracy near experimental reproducibility for GENESIS, and the speed and accuracy gains for ScanFold3. If those hold, the bottleneck in single-cell and RNA work shifts from measurement to deciding which predictions are worth testing. m:…2506 m:…7725 m:…2162