Stable-Shift: Biologically Structured Prediction of Transcriptional Responses to Unseen Gene Perturbations

A new computational method called Stable-Shift predicts how cells alter gene expression when genes are disrupted, including genes the model has never seen before, according to research posted to arXiv on June 22, 2026 [1]. The method, developed by Sajib Acharjee Dip, addresses a core challenge in functional genomics: experimentally testing every possible gene perturbation is prohibitively expensive. Stable-Shift learns a compact representation of expression changes from known perturbations and then maps biological context — such as protein interaction networks and functional annotations — to predict responses for unseen genes [1]. The approach aggregates single-cell measurements into perturbation-level expression shifts and fits a low-rank response basis using only training perturbations [2]. It then predicts an unseen gene’s coordinates in that basis by combining STRING interaction data, Node2Vec network structure, control-cell expression statistics, and Gene Ontology annotations, integrated through a graph convolution encoder [3]. On the K562 Perturb-seq benchmark, Stable-Shift achieved a cosine similarity of 0.592, compared with 0.569 for the existing GEARS model [1]. It also posted higher Spearman correlation and top-gene precision among evaluated methods [4]. Across five unseen-gene splits, the mean cosine similarity was 0.589 with a standard deviation of 0.008, indicating consistent performance [5]. The same performance ordering held in graph-aware, residualized, gene-space, and Norman-dataset comparisons [2]. When decoding predictions into the measured gene space, Stable-Shift reached a cosine similarity of 0.392 versus 0.375 for the CPA model — the highest in that comparison, though substantially below its latent-space scores [4]. The authors note that lower gene-space accuracy and sensitivity to sparse graph neighborhoods limit the scope of the present conclusions [1]. The work builds on decades of molecular biology research into how proteins are regulated. Protein phosphorylation, first reported in 1906, is a reversible modification that alters protein function and is central to the signaling pathways that transcriptional perturbation studies aim to map [6]. Approximately 13,000 human proteins have sites that can be phosphorylated, underscoring the complexity of the regulatory networks that models like Stable-Shift attempt to capture [6].