When the cell lines are not the patients

DepMap is the most useful resource in target discovery and also a non-random sample of human cancer. Cell lines are the tumours that agreed to grow in plastic, and that is not a lottery. For many purposes the bias is tolerable. For synthetic lethality it is structural, because the whole claim rests on two events occurring together. If the co-occurrence pattern in the panel differs from the pattern in patients, the screen cannot see the interaction that matters clinically. It is not that the screen gets the answer wrong. It is that the question never appears.

The scale of the skew is documented. In a survey of 554 published leukaemia and lymphoma cell lines, Drexler and Quentmeier (2020) found that 54 percent were established from patients at relapse or at a refractory or terminal stage, against 34 percent at diagnosis or initial presentation. Of the ten bona fide Hodgkin lymphoma lines, all ten came from terminal or refractory patients and seven of ten from pleural effusions. The panel is not a cross-section of the disease. It is a cross-section of its end stage.

What is documented about those hits is uncomfortable in its own right. Ku et al. (2020) found that published SL screen hits overlap at the pathway but not the gene level, and that most SL phenotypes are strongly modulated by cellular and genetic context, offering evidence for why most reported synthetic lethals are not reproducible. Lin et al. (2019) knocked out the putative targets of ten clinical-stage cancer drugs and found efficacy unaffected, meaning the stated target was not the mechanism. That is the honest evidence base: not that cell line SL fails in patients, but that it frequently fails to reproduce at all.

Expand

NPM1 prevalence from Thiede et al. (2006) and TCGA-LAML (2013); DepMap line count verified against the current DepMap model list and the CCLE mutation profile, September 2026.

One cell line for the largest AML subgroup

NPM1 mutation appears in roughly 27 to 30 percent of adult AML and in 45 to 50 percent of cytogenetically normal AML. DepMap currently lists 52 AML models. Exactly one of them carries the NPM1c exon-12 frameshift: OCI-AML3, ACH-000336. The only other established NPM1c line, IMS-M2, has no DepMap or Cell Model Passport accession at all. Calling that one model is generous, because OCI-AML3 also carries DNMT3A R882C and NRAS Q61L, so any dependency found in it is confounded by two further driver lesions. Why NPM1-mutant AML resists immortalisation is not fully resolved and we will not invent a mechanism, but the consequence is not in doubt: the experimental literature for the single largest molecularly defined AML subgroup rests on two cell lines, one of which most screens never touch.

MDS is missing its own biology

SKM-1 was derived in 1989 from a 76-year-old whose MDS had already progressed to monoblastic leukaemia. MOLM-13 came from a 20-year-old with AML M5a at relapse, arising from antecedent MDS. Both are real, useful, well-characterised lines. Neither reports the ineffective haematopoiesis that defines primary MDS. A 2026 review of MDS models puts it directly: lines frequently used in MDS-oriented studies were derived after leukaemic transformation and therefore reflect clonal evolution and disease progression rather than the biology that defines the disease. MDS progenitors proliferate poorly in vitro and depend on stromal support, which is why the gap has persisted.

For a disease defined by what happens before transformation, that is the wrong tissue.

Why this is worth solving now

Myeloid care has moved from chemotherapy-only to mutation- and risk-stratified treatment, with hypomethylating agent plus targeted agent combinations at the centre: in VIALE-A, azacitidine plus venetoclax extended median overall survival to 14.7 months against 9.6. Approvals now span FLT3 (midostaurin, gilteritinib, quizartinib), IDH2 (enasidenib), IDH1 (ivosidenib, olutasidenib), BCL-2 (venetoclax) and menin (revumenib in 2024 and 2025, ziftomenib in 2025). Almost all of these are AML labels; ivosidenib is the one carrying an MDS indication. The indication is demonstrably tractable for precision therapy. What is scarce is not chemistry but targets, particularly for resistant populations whose options are exhausted.

Many myeloid subgroups are defined by a gene that is already functionally compromised, whether mutated, epigenetically silenced, or expressed at critically low levels. Those loss-of-function events characterise specific patient populations and frequently drive treatment resistance. That gives a synthetic lethal search an unusual starting advantage: one half of the pair is already a validated biomarker with a diagnosed population attached.

Layer 1: prediction that never saw a cell line

In our preprint (Prosz et al., bioRxiv 2026; not yet peer reviewed), open-weight language models are asked to estimate the likelihood of a synthetic lethal interaction for a gene pair with no fine-tuning and no training set. The models reason from biological knowledge encoded during pretraining, and every prediction carries a written mechanistic chain a scientist can reject on its merits.

Expand

Multiple open-weight models were tested with a uniform prompt against known screens, then used to predict 398,277 clinically relevant gene pairs. From Prosz et al., bioRxiv 2026.

The benchmark is Olivieri et al. (Cell 2020), a map of the DNA damage response built from 31 CRISPR screens against 27 genotoxic agents in RPE1 cells. Two things follow from that, and both belong in the open. RPE1 is a non-cancer, near-diploid line, which makes it a genuinely independent benchmark precisely because it sits outside the cancer dependency panel. It also means this benchmark does not by itself demonstrate performance on myeloid biology.

Expand

Performance saturates around 0.73 AUC across the larger models, while both published comparators sit near chance. From Prosz et al., bioRxiv 2026.

Performance saturates around 0.73 AUC: gpt-oss-20B at 0.768, Qwen2.5-72B at 0.735, gpt-oss-120B at 0.716 and Qwen2.5-32B at 0.715, with Qwen2.5-32B chosen for the production screen on performance-to-cost grounds. The two published comparators, SLant at 0.501 and MAGICAL at 0.549, sit at or near chance on the same data. Validation against a dataset released after the training cutoff argues the result is not leakage. 398,277 pairs have been screened, with predictions public at github.com/Paureel/LLMsynthlet.

The claim here is narrow. These priors come from the literature rather than from a panel of cultured lines, so the predictions are not conditioned on which tumours grow in plastic. That is not automatically better. It is differently biased, and the biases are stated below.

Layer 2: clinical overlay, where the ranking is decided

Raw predictions are filtered and re-ranked against patient sequencing rather than cell lines: TCGA-LAML (200 de novo adult AML), Beat AML (672 specimens from 562 patients), the Papaemmanuil et al. MDS cohort (738 patients, Blood 2013) and Haferlach et al. (944 MDS patients, Leukemia 2014). Three questions decide whether a predicted pair survives.

Worked subgroups would be chosen with the partner. Two obvious candidates: TP53-mutant disease, at 5 to 10 percent of de novo MDS and AML, 30 to 40 percent of therapy-related cases and 70 to 80 percent of complex karyotype (where, per Bernard et al. 2020, only multi-hit status carries the adverse phenotype); and splicing-factor-mutant disease, where spliceosome mutations affect roughly 45 to 50 percent of MDS overall, with SF3B1 at 20 to 30 percent, SRSF2 at 10 to 15 percent and U2AF1 at 8 to 11 percent.

Layer 3: agentic evidence assembly, explicitly not validation

For each surviving candidate, an agent assembles a structured dossier from UniProt, STRING, Reactome, DrugBank and ClinicalTrials.gov: mechanistic rationale, existing tool compounds, active clinical programmes, known safety signals, and an audit trail. The boundary matters and we would rather overstate it: this is annotation, not evidence that a prediction is true. Its purpose is rejection speed, letting a scientist dismiss a candidate in ten minutes rather than ten days. In the preprint, this filtering scored the top 100 predicted pairs for novelty and feasibility and surfaced 11 high-novelty, high-feasibility candidates in the RPE-1 context. The failure modes such a layer must avoid are ones we have written about at length.

The first deliverable is not a target. It is a calibration.

Precision, recall, F1 and AUROC get reported openly, so a partner can decide how much to trust the novel predictions from how well the method does on known biology. There is a tension in that design worth naming rather than burying: benchmarking against known SL pairs measures performance where cell line data exists, which is precisely the regime we argue is unrepresentative. It is the best available check and it is not a complete one.

• Literature-derived priors inherit the literature's biases: well-studied genes are over-represented, and absence of evidence reads as absence of interaction.
• Prediction is triage, never proof. Nothing here substitutes for a screen in a relevant model or a patient sample.
• The benchmark showing 0.715 to 0.768 AUROC was run in a non-cancer line against genotoxic agents, not in myeloid disease.
• Clinical cohorts carry their own ascertainment biases and mostly lack matched functional data.
• Normal-tissue tolerance is the hardest question in the whole design and the least well served by public data.

This generalises past myeloid disease to any indication where the panel does not look like the patients. We have run the analogous prioritisation problem before: our combinatorial therapy case study covers the same arithmetic at a different scale, where two genes give roughly 200 million candidate pairs and three give about 1.3 trillion, and ranking before testing is the only tractable move.

If your group works in myeloid malignancy, or anywhere the cell line panel misrepresents your patients, the interesting conversation is which subgroup to start with and which positive controls would convince you. That is a conversation we would like to have.