Shift Bioscience has published research detailing a refined framework for calibrating evaluation metrics in its AI‑driven “virtual cell” models, which aim to predict cellular responses to genetic perturbations. According to the company, the work addresses concerns raised in prior studies regarding weak performance of perturbation models, which the company attributes to mis‑calibrated metrics rather than faulty modelling.
In the study, the company applied its framework across 14 Perturb‑seq datasets, identifying that commonly used metrics failed to reliably distinguish meaningful predictions from uninformative ones—especially in data with weak perturbations. The new approach emphasises rank‑based and differentially expressed gene (DEG)‑aware metrics, which the company says offer consistent calibration across datasets. When evaluated using these refined metrics, Shift’s virtual‑cell models “consistently outperform uninformative mean, control and linear baselines”, the company claimed.
Shift Bioscience said the findings challenge “prior reports that genetic perturbation models do not work” and support broader use of virtual‑cell modelling for target discovery. The company operates a platform combining machine learning and cell biology to drive cell‑rejuvenation therapeutics, with virtual‑cell modelling at its core.
The research is expected to strengthen Shift’s drug‑discovery pipeline by improving confidence in model‑derived gene‑target predictions and streamlining early stage identification of age‑driven therapeutic candidates.
