Many cancer treatments — ionizing radiation, platinum cross-linkers, the alkylator melphalan, and PARP inhibitors, drugs that block a single-strand-break repair enzyme — work by injuring DNA faster than a cell can fix it. Resistant tumors fight back actively: they speed up endogenous DNA repair to erase lethal lesions before cell-death programs commit. But two features of this resistance are systematically missed. First, repair competence is not spread evenly through a tumor; it concentrates in minority subclones — genetically distinct descendant populations of the founding tumor cell — that bulk sequencing, which averages millions of cells together, washes out. Second, DNA damage is itself a signal: in DAB2IP-deficient bladder cancer cells, radiation-induced breaks activate the ATM kinase, which not only coordinates repair but also drives NF-κB (a transcription factor complex that switches on inflammatory and survival genes) into the nucleus to keep the damaged cell alive; knocking down ATM leaves over 40% of double-strand breaks unrepaired at 24 hours and cuts the surviving fraction after a 2 Gy dose from 0.69 to 0.37 (ATM mediates DAB2IP-deficient bladder cancer cell resistance to ionizing radiation through the p38MAPK and NF-κB signaling pathway). That damage-to-survival relay, and the clonal heterogeneity it acts on, are absent from 2D plastic monoculture, invisible to bulk assays, and ignored by every drug-response model we are aware of.
We propose to build and test a model that predicts, for each inferred subclone in a tumor, survival under genotoxic treatment with and without NF-κB-pathway inhibition. The core move is to treat therapy response not as a property of the average cell but as the outcome of selection among clones whose fitness depends on two coupled, clone-specific variables: repair capacity and damage-induced NF-κB activation. The model is trained on large perturbation compendia, validated against clinical cohorts where the competing models have reported numbers, and subjected to one decisive wet-lab test in which mixed clones of known composition are followed from culture into xenografts.
Core Claim
Clinical failure of genotoxic therapy is governed in part by a subclone-specific survival circuit in which damage-induced ATM→NF-κB/p38 signaling complements DNA repair capacity. A model that resolves this circuit per clone will do two things existing models cannot: it will predict patient response to genotoxic agents better than subclone-aware models that ignore inflammatory signaling and better than bulk transfer models that ignore clonal structure; and it will predict that NF-κB-pathway inhibition selectively eliminates repair-high subclones in vivo while appearing far less effective in 2D culture — a specific, falsifiable mechanism for one slice of the in vitro–in vivo translation gap.
The gap we fill is real and documented. SubNetDL infers subclones from bulk mutation data and predicts therapy response, but its features are static mutation calls — it contains no DNA repair kinetics and no inflammatory axis. DeepSADR transfers drug-response prediction from 966 cell lines to 555 TCGA patients and even includes a DNA-repair pathway cluster, but it collapses each tumor into one bulk expression vector. CellVS-Net predicts how drugs rewire gene networks, but only from bulk perturbation profiles. And the mechanistic paper that established the ATM→NF-κB survival axis (He et al.) relied entirely on 2D monoculture, bulk Western blots, and population-level foci counting. No work resolves the circuit at single-clone resolution, and no work carries it across the culture-to-patient boundary.
Research Program
The first workstream defines the circuit as a measurable module. From the mechanistic literature we assemble a gene set covering the damage sensors and repair effectors (ATM, ERCC1, the DNA-PKcs catalytic subunit whose inhibitor-bound structures are known, RAD51, POLQ), the splicing machinery that controls which repair isoforms are actually functional (CLK1 and SRSF5, since resistance to PARP inhibitors is carried specifically by the ERCC1-202 splice isoform rather than total ERCC1, and commercial antibodies cannot distinguish it (Genome-wide CRISPR-Cas9 screening identifies CLK1 inhibition as a strategy to restore PARP inhibitor sensitivity via ERCC1 isoform switching)), and the inflammatory axis (TLR4, MyD88, p38/MAPK14, RELA, NFKBIA, NLRP3, IL1B). We include backup-pathway genes because repair redundancy is the rule, not the exception: melphalan directly induces POLQ-mediated alternative end joining, and blocking one repair route pushes cells into that error-prone backup (Polymerase theta inhibition impairs tumor growth and amplifies melphalan-induced DNA damage in multiple myeloma). We also treat NF-κB as a dynamic state rather than an expression level, because its activation is oscillatory and shaped by cytoskeletal feedback (Actin networks modulate heterogenous NF-κB dynamics in response to TNFα), which is one reason static pathway summaries mislead.
The second workstream builds the model. We extend the perturbation-network approach of CellVS-Net, which predicts post-treatment gene-network rewiring from LINCS L1000 profiles, with two ingredients it lacks: subclone conditioning and a mechanistic coupling term. For each subclone $k$ inferred from bulk data, the model takes the clone's mutational profile and module state, predicts its post-perturbation network rewiring, and converts rewiring into a survival weight,
$$w_k = \exp(-\gamma\, D / R_k)\,\bigl(1 + \beta\, A_k\bigr), \qquad p_k' \propto p_k\, w_k,$$
where $D$ is the administered damage, $R_k$ is clone $k$'s repair capacity (including backup-pathway redundancy), $A_k$ is its damage-induced NF-κB activation, $\gamma$ and $\beta$ are context-fitted coefficients, and $p_k'$ is the predicted post-treatment clonal fraction from the pre-treatment fraction $p_k$. NF-κB inhibition sets $\beta$ to zero, so the model makes clone-specific, per-treatment predictions rather than one score per patient. An important honesty note: LINCS L1000 is gene-level, so isoform identity cannot be read from it directly; in the computational arm we proxy isoform switching through the CLK1/SRSF5 splicing module, and we measure ERCC1-202 directly by targeted assays in the wet-lab arm.
The third workstream validates the model on the benchmarks where competitors have reported numbers. On a genotoxic-restricted split of the TCGA cancer–drug pairs used by SubNetDL (their 507-patient, 10-pair setup), we compare subclone-resolved predictions against SubNetDL itself, against a DeepSADR-style bulk model augmented with NF-κB genes, and against tumor mutational burden. On the multiple myeloma cohorts GSE24080 and MMRF-CoMMpass, where bulk POLQ expression rises across clinical stages and tracks shorter survival (Qun Li et al.; no reported result found on Lacuna for the exact POLQ–survival hazard ratio, so we first quantify bulk POLQ ourselves as a reference), we test whether subclone-resolved POLQ-plus-NF-κB state stratifies progression-free survival better. For the masking question — how large must a repair-high clone be before bulk assays see it — we simulate bulk profiles from single-cell references using the Kassandra database's 2,286 simulated mixtures and the Salcher lung adenocarcinoma atlas of roughly 161,000 cells (Charting cell-type-specific positive genetic interaction at single-cell resolution for lung adenocarcinoma), whose epithelial classifier reached a mean AUROC of 0.974 in bulk validation — our reference point for how well a single-cell signal can survive translation to bulk tissue.
The fourth workstream is the decisive experiment. We construct two isogenic systems: 5637 DAB2IP-deficient bladder cancer cells, where the circuit was mechanistically established, and a multiple myeloma line with a POLQ-high, repair-dependent phenotype. In each, we mix a known resistant clone with a sensitive clone at roughly 10:90 and treat with the relevant genotoxic agent (radiation; melphalan) with and without NF-κB-axis inhibition (ATM knockdown or an IKK/p65 inhibitor), in 2D culture and in subcutaneous xenografts. We read out clonal composition by single-cell sequencing and per-clone γ-H2AX foci and nuclear RELA. The prediction is sharp: in 2D, NF-κB inhibition should show the modest, near-uniform sensitization already reported (surviving fraction 0.69 falling to 0.37 with ATM knockdown in He et al.); in vivo, it should preferentially eliminate the repair-high clone and shift composition — and bulk Western blots of the same tumors should fail to show why. That culture-versus-xenograft divergence, predicted in advance at the clonal level, is the falsifiable heart of the proposal.
Baselines
The naive reference point is standard clinical biomarkers: tumor mutational burden and single-gene repair markers such as bulk POLQ or total ERCC1 mRNA. These are what a clinic would use today without any of this machinery. We will run TMB exactly as benchmarked in SubNetDL, and we will quantify bulk POLQ on GSE24080 and MMRF-CoMMpass ourselves, since no reported result found on Lacuna gives its prognostic number. Beating these is necessary but not sufficient — they are the floor, not the competitor.
The strongest existing competitor is SubNetDL (A network-based deep learning model integrating subclonal architecture for therapy response prediction in cancer), the only published model that both resolves subclones and predicts clinical therapy response; it reports a median AUROC of 0.74 across 10 TCGA cancer–drug pairs, 0.81 on CCLE/GDSC preclinical pairs, and 0.77 on the MSK immunotherapy cohorts, significantly beating TMB there (P = 0.02 and P = 6.0×10⁻³). We will run it on the same genotoxic-restricted TCGA subset, same cross-validation splits, same feature budget and tuning compute as our model, with its SciClone deconvolution untouched. Beating it means a higher median AUROC on the genotoxic pairs with bootstrap intervals excluding 0.74, and — more tellingly — correctly calling relapse in patients whose tumors contain a small repair-high, NF-κB-active subclone that SubNetDL's static mutation features cannot represent. The strongest bulk competitor is DeepSADR (DeepSADR), which reports an AUC of 0.957 on Sorafenib (versus 0.811 for the next best model) and an average clinical AUC of 0.856 with its adaptive readout, versus 0.662 without it; we run it with NF-κB genes added to its pathway set, so the comparison isolates clonal resolution rather than feature availability.
The key ablation is our own model with the coupling term removed: set $\beta$ to zero for all clones, yielding a purely cell-intrinsic repair-capacity selection model. This isolates the inflammatory circuit as the ingredient the proposal is actually about. The precedent for expecting a large effect is DeepSADR's own readout ablation, where removing one module dropped average AUC from 0.856 to 0.662; if removing the NF-κB coupling costs our model little, the core claim is wrong and we will say so. A second ablation collapses subclones into one bulk profile before prediction, isolating the heterogeneity ingredient; Count Bridges (Count Bridges enable Modeling and Deconvolving Transcriptomic Data) and POPPY, which outperformed CIBERSORTx and MuSiC on static deconvolution, serve as the deconvolution baselines for the mixture-recovery experiment.
Why not just…?
Why not just use TMB or a bulk DDR gene signature? The file's own evidence says this falls short. SubNetDL significantly outperformed TMB on both MSK cohorts and, critically, showed that shuffling subclonal structure destroys predictive signal — aggregate mutation and expression counts genuinely wash out the information that matters. Single-gene repair markers fail for a second reason: total ERCC1 misleads because non-functional splice isoforms are expressed and antibodies cross-react (CLK1/ERCC1 isoform switching), and knocking down DNA-PKcs in DAB2IP-deficient cells failed to restore radiosensitivity while ATM knockdown did (He et al.) — resistance depends on which kinase-driven survival crosstalk is active, not on repair-enzyme abundance. Bulk signatures are the floor we must beat, and prior results say the floor is low.
Why not just add NF-κB genes to DeepSADR and skip the clonal machinery? This is a real threat, and we treat it as a first-class baseline rather than arguing around it. If the mechanism acts through average pathway activity, a bulk NF-κB-augmented DeepSADR could capture much of it at a fraction of the complexity. Two things make us doubt it: DeepSADR's own ablation shows that how information is aggregated matters enormously (0.856 versus 0.662), and SubNetDL shows subclonal structure carries signal that bulk aggregation loses; a 10% resistant clone with high NF-κB activation shifts a bulk vector only marginally. But the experiment settles it, not the argument — the augmented bulk model runs on identical data and tuning budget, and if it matches our model on the genotoxic subset, we will conclude the clonal machinery was unnecessary and report that.
Why not just run the combination experiment (NF-κB inhibitor plus genotoxic agent in a xenograft)? Also a genuine alternative, and simpler. But it answers "does sensitization occur in one model," not "which patients and which subclones" — and NF-κB inhibition is exactly the setting where that distinction decides feasibility, because NF-κB governs basal immunity and tissue homeostasis, so systemic suppression is broadly toxic, and its oscillatory dynamics make static dosing hard to calibrate. A sensitization-only experiment cannot tell you when the toxicity is worth paying. Our wet-lab arm is deliberately small and designed as a test of the model rather than a standalone efficacy study: the known 10:90 clonal mixture is what makes the model's prediction of post-treatment composition falsifiable, which a conventional single-line xenograft cannot do. If the model adds nothing beyond what the experiment shows, that is a finding — but only the mixture design can reveal it.
Decisive Checks
Three results would make the case. On the genotoxic-restricted retrieval task built from LINCS L1000 in the style of the CellVS-Net benchmarks, the model should exceed the reported Hits@1 of 0.1250 for CellVS-Net (molecular fingerprints manage only 0.0357, random 0.0179), specifically by retrieving NF-κB-axis inhibitors as perturbations that reverse genotoxic-induced rewiring — a retrieval pattern no cell-intrinsic model should produce. On the TCGA genotoxic subset, it should beat SubNetDL's reported median AUROC of 0.74 with the coupling ablation showing a DeepSADR-sized drop when $\beta$ is zeroed. And in the mixture experiment, the model's predicted post-treatment clonal composition should match observed single-cell composition in xenografts while diverging in 2D culture — with bulk Western blots of the same tumors unable to explain the difference. That last contrast, between what single-clone resolution sees and what bulk measurement cannot, is the proposal in one experiment.
References
- ATM mediates DAB2IP-deficient bladder cancer cell resistance to ionizing radiation through the p38MAPK and NF-κB signaling pathway
- A network-based deep learning model integrating subclonal architecture for therapy response prediction in cancer
- DeepSADR: Deep Transfer Learning with Subsequence Interaction and Adaptive Readout for Cancer Drug Response Prediction
- Cell-Level Virtual Screening
- Genome-wide CRISPR-Cas9 screening identifies CLK1 inhibition as a strategy to restore PARP inhibitor sensitivity via ERCC1 isoform switching
- Polymerase theta inhibition impairs tumor growth and amplifies melphalan-induced DNA damage in multiple myeloma
- Actin networks modulate heterogenous NF-κB dynamics in response to TNFα
- Charting cell-type-specific positive genetic interaction at single-cell resolution for lung adenocarcinoma
- Count Bridges enable Modeling and Deconvolving Transcriptomic Data
- Cross-modal transfer learning for mapping bulk transcriptomes at cellular level