Systematic gene perturbation experiments coupled with transcriptomic profiling reveal causal gene–gene relationships but remain expensive to perform across many biological contexts. To reduce experimental cost and to help prioritize experiments, computational methods that predict transcriptional responses to gene perturbations across conditions have emerged. The authors frame this need as a cross-condition prediction problem and note that prior methods began to address this task but that systematic characterization across distinct cross-condition scenarios is incomplete.
The study distinguishes two operational scenarios for cross-condition prediction. In the seen-perturbation scenario, the model has encountered the same perturbed gene during training but in different biological conditions; prediction requires transferring the perturbation effect to a new condition. In the unseen-perturbation scenario, the perturbed gene itself was not observed in any training condition, so the model must generalize effect predictions for a novel perturbation in a target condition. This distinction structures evaluation and reflects realistic experimental planning trade-offs when deciding whether to prioritize coverage of conditions or perturbation diversity.
To address both scenarios, the authors introduce TranScouter, a lightweight encoder–decoder framework. TranScouter represents the perturbed gene using text-derived embeddings produced by large language model (LLM) processing of gene summaries, rather than relying solely on numeric identifiers or precompiled effect vectors. For target conditions, the model encodes biological context using the transcriptomic profile of control cells from the condition of interest. By combining an LLM-derived text embedding for the perturbation with a condition-specific control transcriptome, TranScouter aims to predict the expected transcriptional response under the target condition.
The authors evaluated TranScouter across benchmarks designed to test both seen- and unseen-perturbation cross-condition tasks. Across these evaluations, TranScouter performed competitively relative to the methods included in the benchmarks. The source text reports that TranScouter is lightweight and capable in both scenarios, but the provided excerpt does not list specific datasets, numerical performance metrics, or the identities of comparator models. Those experimental details and quantitative results were not reported in the source excerpt.
Beyond reporting performance, the study presents empirical analyses to unpack what drives success or failure in cross-condition prediction. Two factors are highlighted: condition-space coverage and perturbation-effect transferability. Condition-space coverage refers to how well the training data spans the diversity of biological contexts that the model must generalize across. Perturbation-effect transferability describes the extent to which the transcriptional effect of a given perturbation is consistent or predictable across different conditions. The authors use these analyses to show how gaps in coverage or limited transferability constrain predictive accuracy, and conversely how broader coverage and more transferable perturbation effects improve cross-condition generalization.
The work positions in silico cross-condition prediction as a tool to prioritize experimental perturbations and reduce the resource burden of exhaustive perturbation profiling. By enabling either transfer of known perturbations to new conditions or prediction for unseen perturbations using gene text embeddings, approaches like TranScouter can guide which perturbations and conditions to test experimentally. The preprint reports no competing interests and is released under a CC-BY 4.0 license.
The source excerpt does not provide full experimental details: specific benchmark datasets, training procedures, model hyperparameters, quantitative performance comparisons, or ablation results were not included. Readers interested in those specifics should consult the full preprint and supplementary material referenced by the authors.
TranScouter combines LLM-derived gene embeddings with control condition transcriptomes in an encoder–decoder architecture to enable prediction of transcriptional responses across conditions. The framework is explicitly evaluated in both seen and unseen perturbation scenarios and is reported to be competitive across benchmarks. Empirical analyses highlight the importance of training coverage of condition space and the transferability of perturbation effects for cross-condition predictive performance.
The article is a bioRxiv preprint by Ouyang Zhu and Jun Li (University of Notre Dame) and was posted August 05, 2026. The preprint and supplementary materials are available via the cited DOI and the bioRxiv page for the manuscript; the excerpt provided here does not reproduce the full methods or numerical results.