Can genetic screens predict drug responses in an untested cell type?
An unreviewed model improved transcription-response correlations across ten cellular contexts, but every test replayed public data. It did not treat cells prospectively, measure toxicity or show that a candidate medicine works.
The short version
- What happened
- In five held-out LINCS cancer-cell contexts, GRAFT predicted 644 drugs across 976 genes and raised mean Top20 Pearson correlation from 0.493 for the source-cell average to 0.734; the strongest zero-drug-measurement comparator reached 0.689.
- Who it affects
- A reliable prioritisation tool could reduce the number of expensive experiments needed before a drug programme advances. A misleading one could concentrate laboratory budgets on familiar cell states and compound classes while overlooking unusual biology or toxicity.
- What remains unproven
- Unreviewed preprint; all performance tests replayed existing public perturbation data.
Keep in mind
Unreviewed preprint; all performance tests replayed existing public perturbation data.
The main benchmark used five LINCS cancer cell lines and transcriptional responses, not patient outcomes or organism-level efficacy.
External query panels varied from 4,530 drugs to only 86, and HAP1 showed weaker, platform-mismatched transfer.
The donor analysis reused genetic measurements supplied to the model and did not include independent drug-treatment experiments.
The study did not measure dose response, toxicity, pharmacokinetics or clinical benefit.
Source analysis · 1 linked source. Primary analysed source date: 1 November 2026.
Keep following the evidence
Join The Impact Brief founding list for selected AI reports, their effect on people and the evidence limits. Read a sample first.
Read the full report · 5 min The evidence, methods, implications and limitations
Genetic responses provided a map for unseen drug measurements
A new bioRxiv preprint from Tsinghua University and Chongqing General Hospital asks whether a limited genetic screen can help predict how drugs will change gene expression in a cellular context where those drugs have not been measured. GRAFT learns from genetic and chemical perturbations in source cells, then uses differences in the target cell's genetic responses to correct the average source-cell drug response. The output is a predicted transcriptional profile—not a prediction of clinical benefit.
The main benchmark used five LINCS cancer cell lines. Each was held out in turn while the other four supplied training data. For every fold, the model predicted 644 query drugs across 976 genes, with target-cell drug measurements reserved for evaluation. Mean correlation across the 20 strongest measured gene responses was 0.734, compared with 0.493 for simply averaging source-cell responses and 0.689 for PRnet, the strongest comparator that also used no target-drug measurements. About 93% of drug–cell-line pairs improved over the source mean.[1]
External contexts were broader, but still retrospective
The authors then evaluated five external contexts—HepG2, MCF10A, Jurkat, K562 and HAP1—with only 21 to 508 usable genetic perturbations. Query panels ranged from 4,530 drugs in HepG2 to 86 in HAP1. The equally weighted mean Top20 correlation was 0.606, versus 0.416 for the source mean and 0.575 for the strongest macro-average comparator. Performance varied: HAP1 had lower correlations, and its drug-anchor gains were not significant in paired tests, possibly because genetic and drug data came from different platforms and conditions.
A calibrated version, GRAFT-Adapt, used ten measured target-cell drugs as anchors and improved most reported reconstruction metrics. The team also generated GRAFT-Atlas: 102,128 predicted drug–context combinations spanning 13 contexts and 7,856 drug identifiers, including 65,703 combinations without measured responses. In 15 induced-pluripotent-stem-cell lines from ten donors, the model predicted 7,755 compounds without iPSC drug measurements. Selected donor rankings correlated with measured target-gene knockdown effects, but those genetic measurements were also model inputs, so the comparison is not independent validation.[1]
The distinction that matters
What the model knows—and what remains unmeasured
The pipeline transfers information between perturbation types, but each step remains upstream of a therapeutic claim.
Measured input
Genetic perturbations in the target cell plus drug responses in source cells
A context representation for prediction
Predicted output
Rank gene-expression changes for unmeasured drug–cell combinations
Higher retrospective correlation on public datasets
Missing test
Prospective treatment of unseen cells, organoids or animals
No evidence yet for efficacy, toxicity or clinical value
The useful output is a shorter experiment list, not a medicine
GRAFT could help researchers prioritise which drug–cell combinations to measure when comprehensive perturbation experiments are unaffordable. The authors released code, models, processed matrices and fixed analysis inputs, making the computational claims unusually inspectable. They also compared multiple correlation, direction and magnitude metrics instead of presenting one favourable score. Funding came from Chinese national and Shenzhen programmes; the authors declared no competing interests.
But the work is an unreviewed computational replay. Transcriptional similarity does not establish target engagement, dose response, toxicity, selectivity or benefit in an organism. Cancer cell lines supplied the source contexts, while some metrics were selected after the internal comparison. No new compound was synthesized, no cell was prospectively treated to test a prediction, and no patient outcome was measured. The assessment would change with preregistered prospective assays in unseen primary or organoid systems, independent drug-response measurements from new donors, and evidence that model-guided selection finds useful hits more efficiently than experimental or simpler statistical baselines.[1]
Watch the background, then check the evidence
Illustrative photo · not a video frame
The Royal Society · 44 min 28 sec · Background explainer
Is this the AI future we want? | The Royal Society
YouTube loads only when you choose; playback starts when you press play.
A Royal Society discussion with Madhumita Murgia and biomedical engineer Alison Noble on AI’s potential in healthcare, science and culture, recorded at the 30 June 2026 event.
Evidence check: A discussion of opportunities and risks, including future possibilities. It is not evidence that every application described is available or clinically proven.
Published 2026-07-01. Source checked 14 October 2026. Check captions on the publisher’s video page.
Check the healthcare evidencePhoto: Daniel Sone / National Cancer Institute / Unsplash · Licence
Continue the story
Read next
Continue exploring ai for science: compare the finding and its evidence limits.
Can sequence-only AI speed drug hit selection without replacing assays?
An unreviewed model trained on about 82 million protein–ligand records improved early virtual-screening enrichment and ranked compounds in milliseconds. Its lab-in-the-loop result was simulated on existing datasets, not a prospective drug programme.
Read the next reportExplore the full source and evidence record
What this means for people
- A reliable prioritisation tool could reduce the number of expensive experiments needed before a drug programme advances. A misleading one could concentrate laboratory budgets on familiar cell states and compound classes while overlooking unusual biology or toxicity.
Global context
Most perturbation atlases over-represent a small set of laboratory cell lines. The China-led study broadens computational transfer across public datasets, but useful translation will depend on independently measured primary cells and disease-relevant systems from more populations and laboratories.
What the evidence does not yet show
- Unreviewed preprint; all performance tests replayed existing public perturbation data.
- The main benchmark used five LINCS cancer cell lines and transcriptional responses, not patient outcomes or organism-level efficacy.
- External query panels varied from 4,530 drugs to only 86, and HAP1 showed weaker, platform-mismatched transfer.
- The donor analysis reused genetic measurements supplied to the model and did not include independent drug-treatment experiments.
- The study did not measure dose response, toxicity, pharmacokinetics or clinical benefit.
What to watch next
- Prospective assays on model-selected drug–cell pairs
- Independent donor-derived cell or organoid validation
- Performance against simpler experimental-design baselines
- Dose, exposure-time and toxicity measurements
- Pre-registered metrics and external laboratories
- Whether prioritisation reduces experiments without losing useful hits
Living evidence record
Impact record IAI-1MSQPIZ
Evidence stage
Studied
Confidence
Supported
Reporting basis
Source analysis
Independent or research support
Present
Record status
Monitoring
Last checked
1 November 2026
Source trail
1 direct source across 1 source type.
People impact
Documented in this record.
Uncertainty
Limits and next checks are explicit.
Stages describe the evidence available—not whether a technology is good or bad. See the public method.
Single-source reporting disclosure
This record analyses one direct source. It can establish what bioRxiv published or reported, but it is not independent corroboration of every performance claim or predicted outcome. The confidence label will change only when broader evidence is added.
Evidence trail
Sources used for this report
Links checked 1 November 2026
Research question, topics and regions
Cross-context prediction of transcriptional drug responses from genetic perturbation data
Topics: Drug discovery · Genetic screens · Cell models · AI for science · Preprint
Regions: Asia · China
This report is labelled source analysis. We summarise and analyse source material in our own words; company statements remain attributed claims until independently supported. Translated summaries preserve the meaning of the original source and link back to it. Read our editorial standards.
Reader commentary
What does this mean where you are?
Tell us what you have seen at work, in your community or in everyday life. A question is welcome too. No account is needed; we review notes before publishing them.
Explore commentary across the portal →Published reader notes
0No published reader notes yet. You can start the evidence-led discussion above.
Prefer a private correction or response? Contact the newsroom.