{"entity":{"id":"idea-bio1-virtual-cell-perturbation","kind":"idea","name":"A virtual cancer cell that predicts what a drug will do before you test it","aka":[],"tldr":"Train a model on millions of experiments where genes and drugs were altered, so it can predict the effect of a new combination without running the experiment.","summary":"Perturbation foundation models trained on Perturb-seq, CRISPR screens and compound-response atlases aim to predict transcriptional and viability responses to unseen perturbations and combinations. The critical missing element is prospective, blinded validation against held-out wet-lab experiments and, eventually, clinical outcomes. Without that, these models risk repeating the overfitting seen in earlier drug-response prediction efforts.","asOf":"2026-09-08","links":[{"label":"Bottleneck evidence (Lab models that fail to predict what happens in patients): Wong, Siah & Lo, Estimation of clinical trial success rates (Biostatistics 2019)","url":"https://doi.org/10.1093/biostatistics/kxx069"}],"tags":[],"related":[],"cancers":[],"sections":[],"technologies":["ai-drug-design","crispr-screens","rna-seq"],"targets":[],"drugs":[],"companies":["recursion","insilico-medicine"],"institutions":[],"pathways":[],"terms":[],"trials":[],"people":[],"bottlenecks":["b-preclinical-models","b-ai-validation","b-combination-space"],"keyPapers":["paper-wong-biostatistics"],"journals":[],"dependsOn":[],"notes":[],"hypothesis":"A perturbation model prospectively predicts the direction and rank order of combination effects in held-out cell contexts substantially better than a strong statistical baseline, and its errors are systematic and characterisable.","rationale":"Combination space is far too large to screen exhaustively, so some form of prediction is unavoidable; the question is whether current models generalise beyond their training distribution, which only blinded prospective tests can answer.","test":"A blinded challenge in which teams predict outcomes of 500 unseen perturbation experiments that are then run in a reference laboratory, with results and baselines published in full.","maturity":"speculative","actor":"data","cost":"medium","horizonYears":4},"route":"/ideas/idea-bio1-virtual-cell-perturbation/","neighbours":{"technology":[{"id":"ai-drug-design","kind":"technology","name":"AI-driven drug & target discovery","route":"/technologies/ai-drug-design/"},{"id":"crispr-screens","kind":"technology","name":"CRISPR functional genomics","route":"/technologies/crispr-screens/"},{"id":"rna-seq","kind":"technology","name":"RNA sequencing & expression profiling","route":"/technologies/rna-seq/"}],"company":[{"id":"insilico-medicine","kind":"company","name":"Insilico Medicine","route":"/companies/insilico-medicine/"},{"id":"recursion","kind":"company","name":"Recursion Pharmaceuticals","route":"/companies/recursion/"}],"bottleneck":[{"id":"b-ai-validation","kind":"bottleneck","name":"AI that is built but not validated or deployed","route":"/bottlenecks/b-ai-validation/"},{"id":"b-preclinical-models","kind":"bottleneck","name":"Lab models that fail to predict what happens in patients","route":"/bottlenecks/b-preclinical-models/"},{"id":"b-combination-space","kind":"bottleneck","name":"Too many combinations to test","route":"/bottlenecks/b-combination-space/"}],"paper":[{"id":"paper-wong-biostatistics","kind":"paper","name":"Estimation of clinical trial success rates and related parameters","route":"/key-papers/paper-wong-biostatistics/"}]}}