Drug discovery roadmap: screening in mice → maps of dependency → designing in silico

Overview

1. 1950s-1990shistoric

   Screening nature and hoping

   The NCI screened tens of thousands of compounds in mouse leukaemias and, from 1990, in a panel of sixty human cell lines. It found the periwinkle alkaloid vincristine, the yew-bark taxane paclitaxel and the antibiotic dactinomycin, and it established the pipeline everyone still uses: cells, then mice, then people. What it could not do was say which people.

   Hide era· 5 records

   National Cancer Institute (NIH)⟳Vincristine⟳Paclitaxel / nab-paclitaxel⟳Dactinomycin (actinomycin D)Patient-derived xenografts
2. 1990s-2015historic

   Targets, structures and high-throughput screens

   Cloning the oncogenes gave discovery a target; crystal structures gave it a shape to fit; robotic screening of millions of compounds and later DNA-encoded libraries gave it throughput. Imatinib was the proof. The cost was a generation of drugs that hit their target and did nothing for patients, because the cell-line and xenograft models that selected them did not represent human tumours. Nine in ten oncology drugs entering trials still fail.

   Show era· 5 records
3. 2015-2024current

   Maps of dependency and models closer to the patient

   Genome-wide CRISPR knockout screens across a thousand cell lines (DepMap) list which genes each cancer cannot live without, and expose synthetic-lethal pairs such as PRMT5 in MTAP-deleted tumours and WRN in mismatch-repair-deficient ones. Patient-derived organoids keep a tumour's architecture and drug response in a dish; xenograft banks keep it in a mouse; TCGA, GENIE and CPTAC supply the genomes and proteomes to interpret them. The first drugs found this way are now in trials.

   Show era· 12 records
4. 2020-2026current

   Structure prediction and generative design

   AlphaFold turned protein structure into a lookup, and AlphaFold 3 (2024), Boltz and Chai extended it to drug and antibody complexes; RFdiffusion and ESM3 design proteins from scratch. Insilico's generative chemistry produced the first AI-discovered drug to reach phase 2; Isomorphic's first oncology candidate entered trials; Recursion merged with Exscientia to pair image-based biology with design; Xaira launched with over a billion dollars to build discovery around these models. The honest scorecard: faster hit-to-candidate, no approved cancer drug yet.

   Show era· 16 records
5. 2015-2026current

   New modalities as platforms

   Discovery is no longer only about small molecules. Degrader and molecular glue platforms remove proteins that cannot be inhibited; ADC linker chemistry (Araris) and bicyclic peptide conjugates (Bicycle) turn a payload into a targeted drug; oligonucleotides silence genes; chemoproteomics finds covalent handles on KRAS. Each platform generates candidates faster than trials can test them, which moves the constraint downstream.

   Show era· 10 records
6. 2026-2030emerging

   Functional precision medicine

   Instead of inferring drug response from genotype, test the drug on the patient's own cells: organoid pharmacotyping in pancreatic cancer, the PARIS organoid screen, tumour fragments kept alive with their vessels, BH3 profiling in leukaemia. The proposal that would make it a field is to grow each trial patient's tumour as organoids and let the results decide which platform arm opens next, with shared reference organoid and xenograft panels so every laboratory tests against the same models.

   Show era· 8 records
7. 2026-2030emerging

   Perturbation data at the scale the problem needs

   Tahoe-100M measured a hundred million single cells across 1,100 drugs and fifty cancer lines; Arc's Virtual Cell Atlas and State model, Geneformer and scGPT are the attempts to learn from that scale how a cell will respond to a perturbation it has never seen. Rigorous benchmarks show current models barely beat simple baselines on unseen contexts, which is the right kind of bad news: the problem is now measurable. The virtual cell roadmap follows this in detail.

   Show era· 8 records
8. 2030+speculative

   In silico first

   The long-range bet is that a candidate is designed, its dose chosen and its toxicity screened in silico and on linked human organ chips before the first mouse, and that digital twins reduce the size of the trials that follow. That requires models that generalise, which requires data that reproduce. A funded replication in every cancer biology PhD and a registry for preclinical experiments that did not work are the unglamorous prerequisites.

   Show era· 7 records
9. What sets the pacecurrent

   Models, reproducibility and the valley of death

   Preclinical models still do not predict people, fewer than half of landmark findings reproduce, and most academic discoveries die before anyone tests them in humans because no one funds the step between. Companies hold compound libraries and negative results that would save others years. A shared compound pool for rare cancer researchers and a guaranteed purchase prize for the first drug against a named hard target are two proposals that attack the incentive problem directly.

   Hide era· 8 records

   Lab models that fail to predict what happens in patientsPreclinical results do not reproduceThe valley of death between lab and productSecrecy and intellectual property block collaborationFailures are hiddenThe undruggable driversA shared compound library that rare cancer researchers can actually useA guaranteed purchase prize for the first drug against a named hard target

Every era's records, trial outcomes and papers, and every watch item, as JSON.