Scoring Drug-Induced Reversal of a Signed DNA-Repair Licensing Module

When a cancer cell becomes resistant to radiation or platinum chemotherapy, it often does so by turning up homologous recombination (HR) — the high-fidelity pathway that repairs DNA double-strand breaks and stalled replication forks. Computational drug-repositioning methods try to find compounds that undo such resistance by a "molecular mirror" trick: measure which genes go up and down in the resistant state, then search perturbation databases for a compound that pushes exactly those genes in the opposite direction. The problem, documented across the field, is that this mirror is fogged by noise. A whole-genome reversal score rewards compounds that invert thousands of passenger genes and generic stress responses, and it is structurally blind to whether the compound actually touches the causal machinery: the ordered cascade of phosphorylation switches — a kinase tags a protein, a phosphatase untags another, an adaptor dimerizes — that licenses the repair complex onto chromatin.

This proposal makes one specific claim: if we reweight transcriptomic reversal scoring onto a small, hand-curated, signed and ordered map of the HR licensing cascade, the resulting score will predict which compounds functionally suppress HR far better than whole-signature reversal does — and this prediction is falsifiable end-to-end, because suppression of HR can be measured with a repair reporter assay and the physical engagement of the licensing machinery can be measured with co-immunoprecipitation (Co-IP — an antibody-based pulldown that recovers a bait protein together with whatever stably binds it). The drug-discovery literature has reversal scoring, and the DNA-repair literature has licensing cascades and Co-IP, but nobody has connected them as a prediction-and-validation loop.

Core Claim

Whole-signature reversal scoring conflates passenger genes with drivers. Pividori et al. showed that aggregating expression into modules already helps: their PhenoPLIER framework, which matches drugs to diseases via unsupervised latent variables, improved drug–disease prediction from an AUC of 0.566 for raw gene-level matching to 0.632. But their modules are unsupervised, unsigned, and blind to post-translational signaling. Meanwhile the HR signaling literature has spent a decade assembling exactly the causal structure those modules lack: ATR phosphorylates FANCD2 to license Fanconi-anemia-complex loading onto stalled forks (Kupculak et al.), PP2A must first strip an inhibitory CK2 phosphorylation from FANCD2 (Yang et al.), ATM phosphorylation of Abraxas at one residue dimerizes the BRCA1-Abraxas platform (Wu et al.), and SIK2 phosphorylates VCP to clear ubiquitinated proteins off damaged chromatin so RPA and RAD51 can load — the mechanism behind radioresistance in colorectal cancer (Meng et al.). The BTRR anti-crossover complex and its partner RAD54L2 add a resolution-stage arm to this map (Ho et al.).

Our claim is that a reversal score computed over this cascade — with gene signs encoding the direction of pathway activation in the resistant state, and weights encoding position in the causal ordering (upstream kinases and phosphatases weighted as drivers, terminal effectors as readouts) — will rank HR-suppressing compounds better than three competitors: raw anti-correlation over all genes, the summarized Reverse Gene Expression Score (sRGES) of Chen et al., and unsigned overrepresentation analysis of the same HR genes in the style of Yu et al.. The claim is then tested physically: top-ranked compounds should suppress HR in a reporter assay and disrupt Co-IP-detectable licensing complexes; compounds ranked high by whole-signature scores but low by the module score should do neither. That divergence set is the whole experiment.

Research Program

The program has three stages. First, we construct the signed licensing module $\mathcal{M}$: a curated set of roughly 40–60 genes spanning the licensing cascade (ATR/CHK1, ATM, SIK2, VCP, PP2A subunits including PPP2R3A, CK2, FANCD2/FANCI, Abraxas/BRCA1, RPA, RAD51, the BTRR complex, RAD54L2, Speedy A/SUN1 for the meiotic arm), each gene carrying a sign equal to the direction of its change in the HR-high resistant state. The resistant-state signature itself comes from SIK2-high colorectal cancer, the model in which Meng et al. established SIK2-VCP-driven radioresistance, stratified by SIK2 expression in TCGA colorectal tumors where they showed high SIK2 expression tracks with poor survival and poor radiotherapy response. Weights $w_g$ are assigned by cascade position, with upstream switches heaviest, and we will preregister both the gene list and the weights before any compound ranking is computed.

Second, we define the module reversal score for a compound $c$ with perturbation profile $z^{(c)}_g$ (the compound-induced expression change of gene $g$):

$$S_{\mathrm{LM}}(c) \;=\; -\sum_{g \in \mathcal{M}} w_g \,\mathrm{sign}(z^{\mathrm{res}}_g)\, z^{(c)}_g$$

where $z^{\mathrm{res}}_g$ is the resistant-state differential expression of gene $g$. The score is large and positive when the compound pushes module genes opposite to the resistant-state direction, in proportion to their causal weight. We compute $S_{\mathrm{LM}}$ for every compound profiled against colorectal or epithelial lines in LINCS L1000 (12,442 compounds in the Chen et al. library) and in the Tahoe-100M atlas, which after quality control holds 95.6 million single-cell profiles across 17,813 cell line–drug conditions with replicate Pearson correlation of 0.975 (Tahoe-100M), giving us single-cell resolution of whether a compound's reversal of the module holds cell-type by cell-type. A known limitation we will quantify rather than ignore: L1000 directly measures only 978 landmark transcripts, so we will report what fraction of $\mathcal{M}$ is landmark-covered versus imputed, and check whether restricting to landmark-covered genes changes the ranking.

Third, we run the prospective loop in the SIK2-high colorectal cancer cells. We select a panel of roughly 30 compounds: the top-ranked by $S_{\mathrm{LM}}$, a matched set ranked high by sRGES but low by $S_{\mathrm{LM}}$ (the divergence set), and ARN-3236, the SIK2 inhibitor Meng et al. already showed radiosensitizes these cells, as a positive control. Each compound is assayed for functional HR competence with the DR-GFP gene-conversion reporter (the assay in which SIK2 knockdown reduced HR with minimal effect on NHEJ in the same system), for RAD51 and RPA foci formation after irradiation, for nuclear pCHK1 (the ATR-dependence biomarker of Sundararajan et al.), and — the physical endpoint — for Co-IP of the SIK2-VCP complex and the FANCD2/FANCI heterodimer, with Phos-tag gels to read VCP hyperphosphorylation, under compound treatment. Co-IP is the right tool here precisely because these enzyme–substrate licensing interactions are transient and context-dependent, the known weak point of both lysis-based pulldowns and database PPI graphs; we will cross-check any hit with proximity ligation in intact cells, as the Co-IP literature itself recommends (Foster et al.).

Baselines

The naive reference point is raw whole-signature anti-correlation in the original Connectivity Map style (Butte et al.): Spearman correlation between the resistant-state differential expression vector and each compound's perturbation vector, over all measured genes, no filtering, no structure. It is the "do nothing" of reversal scoring, and it anchors how much any structured weighting actually buys.

The strongest existing competitor is sRGES, the dose- and time-normalized reversal score of Chen et al., which is the validated standard in this space: it correlated with drug efficacy at a Spearman's rho of 0.61 ($P = 1.34 \times 10^{-6}$) in liver hepatocellular carcinoma and its top prediction, pyrvinium pamoate, worked in xenografts. That 0.61 is the number to keep in mind as the field's benchmark for reversal-score-to-efficacy correlation — though it was measured on general drug efficacy in liver cancer, not on HR suppression; no reported result found on Lacuna for sRGES performance in any DNA-repair context. We will run sRGES ourselves on the identical L1000 and Tahoe profiles, with the same reference-condition normalization, the same cell-line weighting, and the same compute, so the only difference from $S_{\mathrm{LM}}$ is the scoring function itself. Beating it means a higher Spearman correlation between score and measured DR-GFP HR suppression across the compound panel, and a higher AUC for classifying compounds as HR-suppressing versus inert.

The strongest module-level competitor is PhenoPLIER (Pividori et al.), which reported an AUC of 0.632 versus 0.566 for gene-based matching, and is the published evidence that module aggregation helps. We will compute its latent-variable reversal score for the same compounds under the same tuning budget. If unsupervised latent modules match our curated signed cascade on this task, the hand-curation is not earning its keep, and we would say so.

The key ablation isolates exactly our new ingredient: the same gene set $\mathcal{M}$, scored unsigned and unweighted — that is, standard overrepresentation analysis of the HR pathway in the style of Yu et al., who found HR enrichment among resveratrol's targets by ignoring expression polarity and cascade order entirely. If $S_{\mathrm{LM}}$ beats this ablation, the win comes from signs and ordering, not merely from restricting to HR genes. Finally, ARN-3236 serves not as a competitor but as calibration: a compound with a known mechanism in this exact system (Meng et al.); a scoring scheme that fails to rank it near the top is broken.

Why not just…?

Why not just screen kinase inhibitors against purified SIK2? For this one cell model, that threat is real: ARN-3236 already exists and already radiosensitizes these cells, so if the goal were only "find an SIK2 inhibitor," the proposal would be pointless. The goal is different — to test whether structured reversal ranking finds HR suppressors without knowing the kinase in advance, including non-kinase mechanisms. The precedent is Yu et al.: unbiased reversal of chemoresistance signatures from 567 I-SPY 2 patients flagged fulvestrant, an estrogen-receptor degrader, as a sensitizer in receptor-negative HCC-1937 cells, cutting paclitaxel EC50 from 30.9 nM to 5.17 nM — a prediction no kinase-pocket screen of that line would have produced. Our experiment settles it honestly: if every top $S_{\mathrm{LM}}$ hit turns out to act through a single known kinase pocket, the added machinery is redundant for this setting, and we will report that.

Why not just use clinical HRD scar scores? Because scars measure history, not current capability. Ren et al. profiled 9,262 patients across 17 tumor types and found the discordance directly: 21.3% carried pathogenic alterations in homologous recombination repair genes, but only 13.7% scored HRD-positive at the standard threshold (score ≥ 42), and the authors note scars cannot detect dynamic restoration of repair via reversion or compensatory licensing — exactly the SIK2-style rewiring Meng et al. characterized. Our DR-GFP and Co-IP readouts measure live pathway competence, which is the quantity that determines whether a platinum or PARP-inhibitor strategy will work.

Why not just rank all differentially expressed genes? The published evidence says raw gene lists lose to structured aggregation: PhenoPLIER's 0.632 AUC versus 0.566 for gene-based matching (Pividori et al.), and the underlying problem — thousands of passenger genes and generic stress responses dominating whole-signature scores — is a documented bottleneck in the reversal literature itself. The genuinely open question is whether a curated, signed, ordered module beats unsupervised modules; that is precisely what the PhenoPLIER baseline and the unsigned ablation are designed to answer, and if the curated module loses, the proposal's central claim is wrong.

Why not just use STRING or BioGRID instead of doing Co-IP? Because the databases demonstrably miss this machinery. Ho et al. compared their BioID proximity map of the BLM complex directly against BioGRID and GeneMANIA and found the curated databases missed the majority of functional proximal interactors — including the RAD54L2 partnership, whose direct physical interaction with BLM still awaits formal Co-IP confirmation. Gonzalez et al. explicitly list noisy, context-free PPI databases as a core limitation of their own PDGrapher model. In this proposal Co-IP is not a decorative validation step; it is the ground-truth measurement of whether a predicted compound physically engages or disrupts a licensing complex, which no database query can substitute for.

Decisive Checks

Two results decide the proposal. The first is the divergence set: compounds ranked in the top decile by sRGES but the bottom half by $S_{\mathrm{LM}}$ should show no HR suppression in DR-GFP, no loss of RAD51 foci, and no change in Co-IP-detectable SIK2-VCP or FANCD2/FANCI complexes — while compounds ranked high by $S_{\mathrm{LM}}$ (including ones with unremarkable whole-signature scores) should suppress HR and perturb at least one licensing complex at the physical level. If both groups suppress HR equally, the curated structure adds nothing over global reversal and the claim is dead. The second is the correlation test: across the full panel, Spearman's rho between each score and measured HR suppression, compared against the 0.61 benchmark sRGES achieved for general drug efficacy in liver cancer (Chen et al.) — noting again that no reported result found on Lacuna for any reversal score's performance against a direct HR functional assay, which is exactly the missing number this proposal exists to produce. As a secondary check with clinical reach, we will ask whether compounds scoring high on $S_{\mathrm{LM}}$ against HR-high signatures are enriched among drugs whose L1000 profiles invert the HR module in the high-HRD-score tumors of the Ren et al. cohort — the beginning of a dynamic, scar-independent companion diagnostic for repair competence.

References

  1. Projecting genetic associations through gene expression patterns highlights disease etiology and drug mechanisms
  2. Phosphorylation by ATR triggers FANCD2 chromatin loading and activates the Fanconi anemia pathway
  3. PP2A licenses the FANCD2/FANCI complex for chromosome loading
  4. Structure of BRCA1-BRCT/Abraxas Complex Reveals Phosphorylation-Dependent BRCT Dimerization at DNA Damage Sites
  5. Salt‐inducible kinase 2 confers radioresistance in colorectal cancer by facilitating homologous recombination repair
  6. The BLM-TOP3A-RMI1-RMI2 proximity map reveals that RAD54L2 suppresses sister chromatid exchanges
  7. Reversal of cancer gene expression correlates with drug efficacy and reveals therapeutic targets
  8. Multi-target modulation of the homologous recombination pathway by resveratrol promotes DNA damage repair in POI: integrated network pharmacology, molecular dynamics simulation, and experimental validation
  9. Tahoe-100M : A Giga-Scale Single-Cell Perturbation Atlas for Context-Dependent Gene Function and Cellular Modeling
  10. Nuclear <scp>pCHK1</scp> as a potential biomarker of increased sensitivity to <scp>ATR</scp> inhibition
  11. HPF-MPI: A Programming System Supporting Task and Data Parallelism
  12. Session Introduction
  13. Computational drug repositioning for the identification of new agents to sensitize drug-resistant breast tumors across treatments and receptor subtypes
  14. Pan-cancer analysis of homologous recombination deficiency and homologous recombination repair–associated gene alterations in solid tumors from a large Asian cohort
  15. Combinatorial prediction of therapeutic perturbations using causally inspired neural networks