---

Introduction

Cancer rarely stops at a single lesion. A pancreatic tumor, for instance, typically activates KRAS and then inactivates CDKN2A, TP53, and SMAD4 in turn, so most such tumors have lost two or three suppressors at once. If losing one tumor suppressor can create a dependency, losing two together might create one that neither produces alone: a higher-order, or three-way, synthetic lethality, in which a third gene becomes essential only against the doubly-mutant background. Such a target would be exquisitely selective, present as a vulnerability only in the cancers that carry both losses.

The cancer dependency map, which measures how much every cell line needs every gene, lets us look for these third genes directly. But the same property that makes cancer genomes informative, that they delete DNA in blocks rather than one gene at a time, sets a trap. A gene that appears essential in a two-loss background may owe its vulnerability to just one of the losses, with the other an innocent neighbor deleted alongside it. Reading the third hit therefore requires the same driver-versus-passenger discipline the genome itself demands. We built the screen as a single pipeline (Figure 1) and applied that discipline throughout.

Results

A genome-wide screen for the third gene

For each pair of tumor suppressors, we identified the cell lines that had lost both, by damaging mutation or by deletion, and asked, across the whole genome, which gene was more essential in that double-loss background than elsewhere, correcting for tissue of origin so that a vulnerability of one cancer type would not masquerade as a consequence of the two losses. Run on a background whose answer is known, the screen reconstructs it in full. Cells that have lost CDKN2A together with its chromosome-9 neighbor MTAP depend on the entire PRMT5 methylation module, with the partner protein WDR77 at the top, PRMT5 itself just behind, and PELO, PRPF6, and METTL16 alongside (Figure 2, left). As a discovery, cells that have lost both TP53 and RB1, the two master brakes on the cell cycle, become dependent on the cell-cycle engine itself, led by the transcription factor E2F3 and including SKP2, CDK2, the E-type cyclins, and the replication-licensing factors CDT1 and ORC6 (Figure 2, middle). A method that rebuilds a drug-target pathway from a genetic background alone is worth trusting with an unknown one.

The passenger trap

The higher-order setting invites a specific error, and the screen falls into it unless checked. It is tempting to read "PRMT5 is essential in CDKN2A and MTAP double-loss cells" as a genuine three-way interaction, a vulnerability that needs both losses. It is not. When the cells are separated by single versus double loss, PRMT5 is just as essential in cells that have lost MTAP alone, and cells that have lost only CDKN2A are no more dependent on PRMT5 than normal cells (Figure 2, right). The dependency is driven entirely by MTAP; CDKN2A is a passenger, deleted because it sits beside MTAP on the chromosome. This is the common case, not the exception. Across the backgrounds we tested, most third-gene dependencies were carried by whichever single loss was the stronger driver, with the second contributing little. Genuine three-way interactions, in which a gene needs both losses and neither alone, do exist but are quantitatively modest, and they are found only by asking, for every hit, whether either loss on its own already explains it.

A third gene in real genetics: pancreatic cancer

Applied to the genetics that actually define pancreatic ductal adenocarcinoma, the screen points at a clinically pursued target, and the passenger logic explains why. In cells that have lost both SMAD4 and TP53, the strongest third-gene dependency is VPS4A (Figure 3). VPS4A has a paralog, VPS4B, that lies beside SMAD4 on chromosome 18 and is co-deleted with it in about two-thirds of pancreatic cancers. VPS4A becomes essential precisely when VPS4B is lost, a relationship that holds cleanly across the panel. Here the two named losses are again marker and mechanism: TP53 marks the aggressive genotype, SMAD4 marks the 18q deletion, and the actual vulnerability is created by the co-deleted passenger VPS4B, leaving its paralog VPS4A to be targeted. The third gene, read correctly, is a paralog left holding an essential job alone.

A model of third-gene targets and their cancers

We turned the screen into a systematic model, scoring every gene as a candidate third hit against every pair of common tumor suppressors and classifying each result, by the single-loss test, as a genuine higher-order interaction or a passenger-driven one. The model returns a ranked catalogue of third-gene targets (Table 1): from the cell-cycle engine exposed by combined RB1 and TP53 loss, through the PRMT5 methylation axis exposed by MTAP loss and the proteasome subunit PSMD6 in BAP1 and PBRM1 co-mutant tumors, to VPS4A in SMAD4-deleted cancers. Several are already validated or in the clinic; others, such as ERBB2 emerging against combined BAP1 and KEAP1 loss, are hypotheses the same framework generates.

Because each target is defined by a specific pair of losses, its relevance to a cancer is simply how often that cancer carries both (Figure 4). The map assigns each cancer a leading third-gene target: VPS4A in pancreatic (52% of tumors), head-and-neck, and biliary cancers; the E2F and cell-cycle axis in lung, breast, and bladder; PSMD6 in kidney cancer; and the PRMT5 axis in myeloid malignancies. As with any target that rides a deletion, the two-loss background is both the hypothesis and its companion diagnostic.

Several of these assignments are supported by independent work. The E2F and cell-cycle dependency of RB1 and TP53 co-mutant tumors is the biology of small-cell lung cancer, where RB loss renders cells dependent on E2F3 and on mitotic kinases (7, 8). KLF5, the third gene the model returns against combined ARID1A and SMAD4 loss, is independently reported as synthetic lethal with ARID1A, and KLF5 protein accumulates and is leaned upon when SMAD4 is lost (9). VPS4A in 18q-deleted cancers and the PRMT5 axis in MTAP-deleted cancers are established, clinically pursued vulnerabilities (1-3). Others, such as PSMD6 in BAP1 and PBRM1 co-mutant renal tumors, a genetically defined subtype (10), and ERBB2 against combined BAP1 and KEAP1 loss, arise in real cancer contexts but remain, for now, hypotheses the model generates rather than validated targets.

A named case: 1p/19q co-deleted oligodendroglioma

The idea has a birthplace, and it is worth returning to. Oligodendroglioma is defined by an IDH mutation and the co-deletion of chromosome arms 1p and 19q. On 1p36 sits ENO1, one of two genes for the glycolytic enzyme enolase. When ENO1 is deleted, cells survive on its paralog ENO2, but they can no longer afford to lose it: silencing ENO2, or inhibiting enolase pharmacologically, selectively kills ENO1-deleted cells. This was the first demonstration of collateral lethality (11) and is now the basis of a brain-penetrant enolase inhibitor (12). The dependency map agrees, with a sharp condition (Figure 5): ENO2 becomes essential only in cells whose ENO1 is homozygously deleted, not in those that have merely lost one copy.

This is the same lesson the screen taught throughout, in its most literal form. The vulnerability is created by a passenger, ENO1 is not the driver of oligodendroglioma, and it is exposed only when the loss is complete. In the 1p/19q co-deleted tumor the arm-level event removes one ENO1 copy and primes the genotype; the collateral vulnerability is realized fully where a focal event removes the second. Marker, mechanism, and the depth of the deletion are three separate things, and a target is earned only by keeping them apart.

A dosage-sensitive target for the heterozygous genotype, and its safety

Because 1p/19q co-deletion is heterozygous, it does not silence a gene so much as halve its dose, so the collateral strategy that fits it is not the homozygous one of ENO1 but a dosage-sensitive one. Scanning genes on 1p and 19q for paralogs that become essential as the partner's expression falls, rather than only when it is deleted outright, returns candidates that respond to a single-copy loss. The clearest is ATP1A1, the housekeeping subunit of the sodium-potassium pump: as its neuronal paralog ATP1A3, encoded on 19q, is reduced, cells come to depend on ATP1A1 (Figure 6), and the dependency holds in brain-derived lines.

This target carries an unusually explicit on-target liability, and it shows why modality matters. ATP1A3 is among the most important neuronal genes in the genome; its loss causes severe movement and developmental disorders (13). A small molecule against the sodium-potassium pump, such as a cardiac glycoside, cannot tell the alpha-1 and alpha-3 subunits apart, because they share a conserved drug-binding pocket, so it would inhibit ATP1A3 in normal neurons and reproduce that neurological toxicity, on top of the cardiac toxicity these drugs are known for (14). An antisense oligonucleotide changes the calculus. Selectivity for an oligonucleotide is a matter of sequence, and the ATP1A1 and ATP1A3 messages differ enough that an ASO can lower ATP1A1 alone while leaving ATP1A3 untouched; delivered into the cerebrospinal fluid, as CNS antisense drugs already are in the clinic (15), and too large to cross the blood-brain barrier, it stays within the central nervous system, isolating ATP1A1 inhibition to the brain and sparing the heart and kidney that constrain a systemic small molecule. Under that modality the window is real: the tumor, having lost ATP1A3, depends on ATP1A1, while normal neurons keep ATP1A3 and normal glia keep the alpha-2 subunit to buffer the loss. The same target that is treacherous for a small molecule becomes tractable for a sequence-selective one, provided the knockdown stays below the tumor's reduced buffer without exhausting the normal cell's.

Discussion

Higher-order synthetic lethality is real. Two-gene-loss backgrounds do create third-gene dependencies, the screen finds them, and some, like PRMT5 in 9p21-deleted cancers and VPS4A in 18q-deleted pancreatic cancer, are already being pursued as drugs. But the two-loss framing is seductive in a way that must be resisted. Because cancers delete chromosome segments, the two genes a clinician names are often not two independent hits but a driver and its passengers, and the true vulnerability may be created by a third deleted gene that no one named. The claim that a target requires two specific losses is therefore a strong one, and it is earned only by decomposition: showing that neither single loss, on its own, already produces the dependency.

The practical guidance is the same in the genome and in the analysis. Do not credit a correlated bystander until an independent test has ruled it out. For a two-loss target, that test is the single-loss comparison; a genuine higher-order interaction survives it, a driver-plus-passenger does not. Applied with that discipline, the screen delivers what it promises: a short list of third-gene targets, each tied to a specific, biomarker-defined cancer subtype, in which the gene that matters, and the loss that truly enables it, have been correctly identified.

Code availability

The screen described here is released as an open-source tool, A View to a Kill, under the MIT license at github.com/bajabio/a-view-to-a-kill. In one line it takes two genes lost in a cancer as the background and returns a rank-ordered list of third-gene targets that become selectively essential when both are lost:

Each target comes with a tissue-corrected selectivity score and, importantly, the single-loss decomposition that labels it a genuine higher-order interaction or a co-deleted passenger. Given a biomarker-defined background the tool reproduces the cases discussed above without being told the answer: CDKN2A + MTAP returns the PRMT5/WDR77 axis, and RB1 + TP53 returns the E2F/cell-cycle machinery. It is meant for hypothesis generation; every nomination still needs a modality and an experimental test. The repository holds the code, the ranked example outputs, and pointers to the public DepMap, Ensembl, and GTEx data it runs on.

References