Shift published a calibration and benchmarking framework in Nature Biotechnology for deep learning models that predict transcriptomic responses to genetic perturbation, attributing earlier reports of poor model performance to miscalibrated evaluation metrics that failed to separate biological signal from technical artifact. The company intends to apply the framework to large-scale in silico and in vitro screens for inhibition targets in fibrosis and age-related disease, following its identification of SB-101 as a first dual-purpose target.
Key Points
- Calibration framework separates biological from technical signal in perturbation datasets
- Miscalibrated metrics cited as cause of prior model underperformance
- Framework directs fibrosis and rejuvenation target screens, starting with SB-101
Longevity Analysis
Fibrosis is a common endpoint of chronic tissue injury across the lung, liver, kidney and heart, where repair capacity gives way to scarring and functional decline. Computational models that predict how cells respond to a specific genetic intervention compress the search space for targets that restore regenerative capacity rather than merely suppress symptoms — but only if their evaluation metrics distinguish genuine transcriptomic response from batch effects and assay noise. Correcting measurement before scaling discovery is the operative principle here, and it applies as much to the instruments interpreting the body as to the models predicting it.
Original published by LT Wire.