DepMap 26Q3 synthetic-lethality scan: 22 novel, ranked candidate pairs (exploratory hypotheses, none confirmed)

Summary

The kit scanned 684 anchors against 17,786 genes across 1,206 cancer cell lines (12.2M tests). It recovered all 4 required controls (WRN/MSI, STAG1/STAG2, SMARCA2/SMARCA4, ARID1B/ARID1A) and 6 of 7 known pairs. Of 7,681 hits, 1,902 survived a lineage and confounder audit. Of 49 reviewed pairs, 37 showed no clear prior literature, and 22 novel pairs were ranked after the safety review into 2 tier A, 10 tier B and 10 tier C. Two further caveats apply. The platform's novelty check labels the whole scan "similar" to published DepMap synthetic-lethality work, so it is not "possibly_new", and its single review was inconclusive. All pairs are lab hypotheses, not confirmed results. No preregistered confirmatory tests were run.

What was analyzed

• Inputs: DepMap 26Q3 CRISPR gene effect (1,361 models, 1,206 lines in the scan), damaging mutations, copy number (WGS), and expression. External sources were Europe PMC, OpenAlex, STRING, GTEx v8, gnomAD v4.1, Open Targets, DEMETER2 RNAi, PRISM/GDSC2, and cBioPortal TCGA.
• Controls: 6 anchors × 17,786 genes = 106,711 tests, 20 permutations, FDR 0.1, effect ≤ −0.2.
• Anchor definition: 6,350 anchors defined, with deep deletion at CN < 0.3, a cap of 30% of lines, and sex-chromosome genes removed (2,449).
• Main scan: 12,165,001 tests; 2,369 pairs and 7,681 hits passed.
• Confounder audit: all 7,681 hits adjusted for lineage, TP53, KRAS, MSI, target copy number and 5 PCs, with 200 permutations; 1,902 survived.
• Novelty review: 49 pairs checked against literature.
• Safety and druggability: 1,902 pairs, 792 genes.
• Replication: a within-DepMap lineage hold-out and PRISM/GDSC drug replication, plus an RNAi check.
• Stability and context: a 1,000-resample bootstrap for the 22 ranked pairs, and TCGA anchor frequencies for 12 final candidates.
• Exploratory scan: expression-loss anchors, 55.4M tests.
• Compute: 88 runs (75 succeeded), 3.9 CPU hours, 12 work groups done and 2 dropped.

Key results

• Controls (exploratory):

MTAP–MAT2A was missed.

• Verdict classes from the scan: 11 target_candidate, 38 needs_review, 37 known_pair. Novelty split of the 49 reviewed: 37 novel, 9 uncertain, 3 known.
• Tier A:

Both are tier A, with no safety flags and a hold-out pass.

• Other ranked pairs: SUV39H1→TIMM17A (14 lost lines, beta −0.57, CI [−0.85, −0.29]) is the strongest effect among them, but it is graded by one run as tier B and by another as tier C.
• Other pairs reaching the final-candidate stage: CDH11→LGR4 (coef −0.47), CT45A1→FEM1B (−0.39), ARL11→KPNA4 (−0.592; 9 lost lines, 2 lineages), IDH mut→RAD54L2, and TIMP1→TIMM17A. The full set of 12 final candidates is in the hold-out tables.
• Replication:
  • Drugs: 1 pair replicated in both PRISM and GDSC, and it is not novel.
  • RNAi: only 2 of 22 pairs were testable, and 1 replicated (SMARCB1_LoF→NABP2).
• TCGA anchor loss (context only): IDH mut is frequent in diffuse glioma (0.870). HACD4 loss is 0.174 in diffuse glioma, 0.093 in pleural mesothelioma and 0.084 in pancreatic cancer. ARL11 is 0.098 in prostate cancer. SUV39H1 and CDH11 are rare (≤0.033).
• Safety: 15 small-molecule-tractable genes among the 22 novel pairs in one run, and none tractable in Open Targets by the director's account.
• Expression-loss scan: 46 paralog pairs outside the kit table, with permutation FDR 0.28.

drug_replication_final.csv (group #10, first 25 rows)

Alternatives explored

These directions went beyond the request.

• Expression-loss anchors and a paralog scan: a weak, ranked screen only (FDR 0.28).
• PRISM/GDSC drug replication: no novel pair replicated.
• Confounder audit of all hits: removed 75% of hits (1,902 of 7,681 survived).
• Declined: a second scan of the remaining anchors was dropped because it duplicated the first scan. A repeat RNAi replication and a repeat TCGA frequency check were declined because both had already been done.

What did not work

• MTAP–MAT2A was not recovered.
• Of the 7,681 hits, 5,092 had weak adjusted effects and 3,402 were concentrated in a single lineage.
• The RNAi replication was possible for only 2 of the 22 pairs, because the target was absent from the RNAi data or too few lines had lost the anchor.
• Drug replication found no novel pair replicated in both datasets.

Confidence and limitations

• Confidence is moderate to low (director: 0.5). The claim has level "candidate", novelty "similar", and one inconclusive review. There are no preregistered tests.
• The hold-out is inside DepMap, so it is not independent.
• The bootstrap resamples lines rather than lineages, so the CIs are too narrow, especially for pairs with 10–14 lost lines.
• GTEx is bulk tissue.
• TCGA frequencies show where the anchor is lost, not where the lethality works.
• Tier assignments are unstable between runs for some pairs.
• Lab validation is required: CRISPR knockout in isogenic anchor-lost versus intact lines, with normal-cell controls.

Next steps

• Validate HACD4→HBS1L and SMARCB1_LoF→NABP2 in isogenic models and in glioma, myeloid, and kidney rhabdoid lines.
• Test SUV39H1→TIMM17A and ARL11→KPNA4 in more lines, since each rests on 14 or fewer lost lines.
• Assess tractability of the top targets in ChEMBL, and test them in an independent dataset such as other CRISPR or RNAi screens.
• Check normal-tissue toxicity in non-malignant cell models.