Predicting drug response from single cells: PERCEPTION

S3 Virtual Workshop Series: Drug Response and Resistance

NCI Scientific Events and Resources · 2026-04-23 · Source

NCI’s English-language “Drug Response and Resistance” workshop recording discusses the temporal and spatial biology of cancer drug response and resistance. The event took place on 20 May 2025; the YouTube publication is dated 23 April 2026. It provides general research context, rather than a study validating PERCEPTION’s clinical benefit or a news report specific to this article.

What problem do single-cell data address?

A tumor can contain cell subgroups that respond differently to drugs. Bulk RNA measurement combines this diversity into an average signal; single-cell RNA sequencing helps examine differences between cells. NCI notes that these measurements are more expensive and not widely available in clinical settings.

How does PERCEPTION learn?

In the research published on 18 April 2024, the model first learns from bulk gene-expression and drug-screening data in cell lines, then is tuned with single-cell data. This is transfer learning: applying knowledge learned from one data type to another related type. The study obtained predictive models meeting its criterion for 44 of 133 drugs; these counts are not a patient success rate.

Which patient data were evaluated?

According to NCI, previously published data included 41 people with multiple myeloma receiving a four-drug combination and 33 people with breast cancer receiving a two-drug combination. Resistance development was also investigated in data from 24 people with non-small cell lung cancer receiving targeted therapies. These are different disease and treatment contexts and should not be reduced to one general accuracy percentage.

Why separate prediction, clinical benefit and authorization?

Finding patterns associated with response in these analyses does not prove that model-guided treatment improves outcomes in a prospective comparison. Describing drugs as FDA-approved in the source also does not mean that the AI method or every use it proposes is authorized. Research findings, clinical-use decisions and regulatory authorization require separate assessment.

What should further evidence examine?

Editorial assessment: independent validation at new centers and in different patient groups, the effects of measurement quality and prospective evaluation of clinical benefit matter. Similar RNA profiles do not guarantee the same personal outcome. This explanation does not recommend starting, changing or stopping a drug and establishes no effectiveness claim for RAI products.