2026-08-18
Cell perspective: 15 challenges for generative AI in biology, where data trails language by 1000x, linear baselines hold, and blind tests should decide.
Generative AI has cleared the molecular layer of biology: protein structure prediction, de novo protein design, pathogenic mutation scoring. The open question is whether the same machinery transfers up a level, to predicting how cells and multicellular systems behave, which is what immunology, cancer, and neurodegeneration actually need. This Cell perspective, from Andrea Califano's group at Columbia with 15 authors across Columbia, Stanford, Northwestern, CMU, Rockefeller, and Yale, argues that it does not yet, and that the gap is structural.
Three numbers carry the argument. Between 2006 and 2022, only 14.6% of clinical trials ended in drug approval, falling to 6.7% in 2014-2023. Language models train on trillions of tokens; the largest biology datasets hold 10^10 to 10^11 measurements, roughly a thousand-fold fewer, and even a billion-cell atlas such as CellxGene does not close the gap. The combinatorics are hostile too: assembling a 60S ribosomal subunit takes up to 47 proteins in concert, and combination spaces of that size reach about 3.48×10^150 possibilities, more than the 10^80 atoms in the universe. Brute-force scaling does not get you there.
Modeled on Hilbert's 23 problems from 1900, the paper proposes 15 grand challenges at four levels: molecular interactions (regulatory and signaling logic, epigenetic logic, cell-cell communication), molecular function (synthetic mechanism design, genome-to-function, drug mechanism of action), cellular and systems function (minimal genome design, cell state reprogramming, Boolean biocircuit design, co-culture design), and translation (biomarker identification, toxicity, efficacy, vaccination response, clinical trial outcome).
Each challenge ships with acceptance criteria. Challenge 3 asks for AUROC of at least 0.85 across 3 or more tissue contexts; challenge 7 wants at least 90% overlap with the JCVI-syn3A minimal genome; challenge 11 sets AUC of 0.90 or higher in blinded prospective cohorts with a panel of at most 20 features, replicated in 2 independent cohorts; challenge 15 requires AUC of 0.80 for responder prediction in prospective phase 2/3 trials.
Benchmarks come in two tiers: tier 1 for retrospective relative ranking, tier 2 for prospective blind tests where predictions are sealed before the experiment runs, following CASP. Architecturally, the authors want biological priors pre-wired into models, with interaction networks and physical constraints shaping graph attention instead of leaving the model to learn everything from data. Their own GREmLN, a transcriptomics foundation model with molecular interaction networks inside the attention mechanism, reportedly does more with less data and fewer parameters, especially on the harder task of predicting gene perturbation effects; the authors have filed a patent on it, which the paper discloses.
A dedicated section answers Sutton's "Bitter Lesson" essay (general compute beats human knowledge in the long run) with three counterarguments. Biology is fundamentally data-scarce. Protein interactions follow thermodynamics and electrostatics, so encoding them restricts hypotheses to physically plausible mechanisms rather than baking in human intuition. And medicine cannot wait: language modeling can sit out the wait for compute, while every year of delay in the clinic is measured in failed trials and untreated patients.
A perspective with no new experiments; the evidence is assembled from prior work.
| Evidence | Number | Reference point |
| Clinical trial success | 14.6% (2006-2022) | 6.7% (2014-2023) |
| Biology data scale | 10^10-10^11 tokens | trillions for language, 1000x gap |
| Single-cell foundation models | no better than linear baselines out of distribution | scGPT, Geneformer; Nature Methods 2025 |
| Sample size where nonlinearity helps | >1M individuals | PRSformer |
| 47-protein combination space | 3.48×10^150 | 10^80 atoms in the universe |
The benchmark critique deserves its own paragraph. Cell type classification, the standard single-cell benchmark of telling CD8+ from CD4+ T cells, is biologically trivial: linear regression does about as well as the best published generative models, and the significant p values that separate them only appear at tens of thousands of cells. AUC itself is mostly irrelevant, because the only region of the precision-recall curve a biologist can use is the part with FDR at or below 5%; the other 95% of the area measures nothing actionable. The 1995 Washington Post story that declared protein folding solved, on six correct predictions out of seven, is the cautionary tale: the real solution took 40 more years and prospective validation.
For AI4Science researchers, this is a citable problem map. All 15 challenges come with data availability assessments and named non-AI baselines, computational and experimental, so the direction, the comparison, and the passing score are all specified up front.
For benchmark builders, the two-tier framework and the FDR argument are a concrete indictment of how single-cell foundation models are currently evaluated, building on the 2025 Nature Methods finding that linear baselines match deep models out of distribution and reframing it as structural rather than incidental.
For funders and data platforms, the paper proposes a public-private consortium targeting 500 trials with molecular-level data within 5 years, starting from already-failed trials where competitive sensitivity is low, plus an AI-DREAM style blind benchmarking effort. Human Immunome Project, DepMap, and Tahoe-100M get an explicit role as data infrastructure.
A perspective with no experimental evaluation: the 15 challenges are an agenda, not validated conclusions. The authors concede that biological problems resist Hilbert-style axiomatic formulation and frame the list as an open-ended program.
Interests require discounting. GREmLN is the authors' own model with a patent filed, so the central claim that graph priors help rests partly on self-report, and several authors hold positions at Biohub, CZI, and GenBio.AI, institutions whose data-production plans align with this agenda.
The acceptance thresholds (AUC 0.90 and similar) are asserted rather than derived, and the argument that clinicians need interpretable models is stated without evidence. The 15 challenges also span wildly different difficulty, from mutation effect prediction, where AlphaMissense already operates, to clinical trial outcome prediction, which may stay open for decades; presenting them as one list risks misreading priorities.
The paper discloses that Claude Sonnet 4.6 was used to check references and grammar during writing; the concepts are the authors'.