A Three-Axis Framework for Therapy Prioritization in MBD4-Deficient Tumors: De-confounded Synthetic Lethality Modeling, Replication-Stress Pharmacology, and PARP1 Trapping Decoupling
Background: Loss-of-function mutations in MBD4 (methyl-CpG-binding domain protein 4) cause a specialized base excision repair (BER) deficiency characterized by hypermutated CpG>TpG signatures and sensitivity to immune checkpoint blockade. However, post-immunotherapy treatment prioritization remains empirically unmapped. In clinical oncology, DNA damage response (DDR) defects are frequently extrapolated to poly(ADP-ribose) polymerase inhibitor (PARPi) sensitivity by analogy to homologous recombination deficiency (HRD), without testing whether upstream glycosylase loss generates the nicked single-strand break (SSB) intermediates required for toxic PARP1-DNA trapping. Methods: We developed a three-axis framework integrating mechanistic curation, large-scale pharmacogenomics (DepMap 24Q2 CRISPR/RNA-seq, n=1,517; GDSC2 dose-response profiling, n=928; PRISM Repurposing), and a 5-Layer De-confounded Synthetic Lethality Machine Learning (SL-ML) Engine (SL-XGBRanker). To prevent functional dilution from variants of uncertain significance, MBD4 loss-of-function (MBD4-LOF) was strictly restricted to truncating and canonical splice-site alleles (21 of 36 mutated lines), independently validated by nonsense-mediated mRNA decay (NMD; two-sided p2s = 0.0367), while excluding 15 lines harboring 19 non-truncating variants (18 missense, 1 inframe deletion). All primary pharmacogenomic comparisons were evaluated using two-sided Mann-Whitney U tests (p2s), bootstrap 95% confidence intervals (CIs), Benjamini-Hochberg false discovery rate (BH-FDR) multiplicity correction across six pre-specified candidate axes, and Two One-Sided Tests (TOST) for expression equivalence. Results: Our analysis delineates three distinct therapeutic axes in MBD4-LOF tumors: (1) Cytidine Analog Vulnerability (Axis 1): Mechanistic structural evidence demonstrates that MBD4 loss impairs excision of 5-substituted pyrimidine analogs and deaminated bases, creating a direct biochemical bottleneck to gemcitabine and 5-fluorouracil incorporation. (2) ATR-Check1 Replication-Stress Vulnerability (Axis 2): In GDSC2 profiling, MBD4-LOF cell lines exhibit significant sensitivity to the ATR inhibitor ceralasertib (AZD6738). In the clinically relevant TP53-mutant background-which disables the G1/S checkpoint and forces reliance on ATR-mediated S/G2 arrest-MBD4-LOF lines (n=11 vs n=619 WT) demonstrate pronounced ceralasertib hypersensitivity (Delta LN_IC50 = -1.07, 95% CI [-1.62, -0.41], Cohen's d = -0.74, p2s = 0.0060, BH-FDR q = 0.036; Delta AUC = -0.087, d = -0.89, p2s = 0.0017), with 11/11 leave-one-out (LOO) jackknife iterations retaining two-sided significance (p2s <= 0.014). This genotype-associated dependency remains robust in the unstratified pan-cancer cohort (n=14 vs n=914 WT; Delta LN_IC50 = -0.73, 95% CI [-1.31, -0.11], d = -0.50, p2s = 0.0430; Delta AUC = -0.056, d = -0.56, p2s = 0.0259) and after purging microsatellite instability-high (MSI-H) lines (n=10 vs n=906 WT MSS; Delta LN_IC50 = -0.91, 95% CI [-1.50, -0.22], d = -0.62, p2s = 0.0307). By contrast, Wee1 inhibition (adavosertib; n=15 vs n=920 WT; Delta LN_IC50 = -0.35, p2s = 0.1667) does not reach statistical significance. (3) Absence of Single-Agent PARPi Sensitivity (Axis 3): While PARP1 transcript abundance strongly governs pan-cancer PARPi cytotoxicity across n=481 cell lines (rho = -0.416, p < 10^-21 across five clinical PARPi), MBD4-LOF tumors (n=19 vs n=1,498 WT) show zero compensatory PARP1 upregulation (median 6.77 vs 6.66 log2[TPM+1], Delta median = +0.11, 95% CI [-0.39, +0.71], p2s = 0.605; TOST equivalence p_TOST = 0.0199 within +/- 0.50 log2[TPM+1]). Direct GDSC2 evaluation of MBD4-LOF (n=8) versus WT (n=478) confirms complete absence of single-agent PARPi sensitization across talazoparib (p2s = 0.973), olaparib (p2s = 0.522), niraparib (p2s = 0.877), rucaparib (p2s = 0.804), veliparib (p2s = 0.538), and the composite 5-drug mean PARPi Z-score (Delta median Z = -0.024, 95% CI [-0.84, +0.60], p2s = 0.841). Under Leave-One-Anchor-Out cross-validation across 12 canonical DDR genes (370 gene-drug pairs), our 5-Layer De-confounded SL-ML Engine achieves an out-of-fold AUROC of 0.9716, AUPRC of 0.8007, and Top-1 accuracy of 82.86%, zero-shot ranking gemcitabine (#1) and ceralasertib (#2) as the top therapeutic dependencies for held-out MBD4 while rejecting all PARP inhibitors (ranks #13, #27, #28, #31). Conclusion: MBD4-LOF tumors exhibit a selective, TP53-potentiated pharmacological dependency on ATR kinase inhibition alongside mechanistic vulnerability to cytidine analogs, but lack both the nicked BER repair intermediates and the PARP1 target expression required for single-agent PARP inhibitor efficacy.