Shift Bioscience has published research in Nature Biotechnology describing a calibration framework designed to improve the assessment of deep learning models used as AI virtual cells.

The framework will support Shift’s planned large-scale in vitro and in silico screens to identify genetic inhibition targets for rejuvenation and age-related diseases, with an initial focus on fibrosis.

Framework aims to improve virtual cell benchmarking

Genetic perturbation models can predict how cells respond to interventions such as the activation or inhibition of specific genes at the transcriptomic level.

These AI virtual cell models could enable researchers to screen large numbers of potential targets in silico, but questions have remained around how reliably their performance is measured.

Shift’s study addresses this issue by establishing a benchmarking approach that accounts for both biological and technical signals within datasets used to assess model performance.

The company said previous benchmarks may have been affected by calibration issues that reduced their ability to distinguish genuine differences in model performance from limitations in the datasets or metrics being used.

The research builds on work previously reported by Shift in November 2025 and provides a framework for assessing whether virtual cell models generate biologically meaningful predictions.

AI models to support target screening

Shift plans to apply the framework to large-scale target discovery programmes combining in vitro and in silico screening.

The initial work will focus on fibrosis, which is associated with ageing and multiple age-related diseases. The company said the screens will seek inhibition targets with potential applications in both cellular rejuvenation and disease treatment.

The programme follows the discovery of SB-101, which Shift describes as its first dual-purpose target.

The distinction between model performance and target validation will be important as the programme progresses. While improved benchmarking could help researchers identify models that make more reliable predictions, targets identified through AI-based screening will still require experimental validation.

Dr Brendan Swain, CSO and founder of Shift Bioscience, said: “Our findings show that by using well-calibrated metrics and the right dataset, virtual cell models can generate biologically meaningful insights. As a result, we can use them with greater confidence to identify promising new targets that are relevant to aging and disease.”

Shift said it is applying the framework directly to its target identification programme, focusing on targets whose inhibition could have applications in both rejuvenation and age-related disease.

Virtual cells and drug discovery

AI virtual cells are being developed as computational tools for modelling how biological systems respond to genetic or other interventions.

For drug discovery, the potential value is the ability to evaluate large numbers of biological hypotheses before progressing selected targets into laboratory studies.

The Shift study therefore addresses an important part of the developing AI drug discovery field: not simply whether a model can make predictions, but whether the methods used to evaluate those predictions accurately reflect biological performance.

The company’s next stage will test the framework through target discovery programmes combining computational predictions with experimental screening.