Complex-Aware Pathway Features for Predicting Xenograft Drug Response

When machine learning models trained on cancer cell culture data are asked to predict how the same drugs will work in patient-derived tumors grown in mice, they mostly fail. A systematic benchmark of models trained on large in vitro pharmacogenomics datasets and evaluated on the NIBR patient-derived xenograft encyclopedia found that drug response prediction from baseline gene expression achieved an average $R^2$ of only about 0.15, with complete failure for some drug cohorts Assessment of modelling strategies for drug response prediction in cell lines and xenografts. The same benchmark shows the failure is not about model fit in vitro — models identify tissue type with 0.79–0.89 accuracy — but about what the features fail to represent: living tumors rewire which proteins physically assemble into working complexes, and no current predictor captures that.

This proposal builds a predictor whose features are pathway activity scores computed over a graph of physically validated protein complexes rather than over unstructured gene lists. The intuition is simple: individual gene expression levels drift between plastic dishes and tumors, but the coordinated activity of a physical multi-protein machine — a scaffold bound to a transcription factor, an E3 ligase bound to its substrate — is the functional unit a drug actually disrupts. We weight the edges of these pathway graphs with complexes confirmed by co-immunoprecipitation (an antibody-based pull-down that captures which proteins are physically bound together in a lysate), and weight the nodes by oncogene overexpression amplitude. The claim we test is that enrichment scores which respect physical complex topology translate across systems better than either single-gene features or conventional topology-blind gene-set scores.

Core Claim

Standard functional enrichment analysis treats a pathway as an urn of genes: every member counts equally, and wiring — who physically binds whom — is ignored. This is a documented limitation of the method class From random to predictive: a context-specific interaction framework improves selection of drug protein–protein interactions for unknown drug pathways, which showed that global, context-free protein-interaction statistics predict drug side effects no better than chance (AUROC 0.51) while a context-specific view reaches AUROC 0.77. We take that principle into the oncology translation problem that paper never touched, and we ground it in physical evidence the field already produces but discards.

Across the cancer literature, co-immunoprecipitation is almost always used as a binary yes/no check for one pre-selected protein pair: CKAP2 binds TFDP1 in colorectal cancer CKAP2 Regulated by TFDP1 Promotes Metastasis and Proliferation of Colorectal Cancer through Affecting the Tumor Microenvironment, IQGAP3 binds GLI1 in lung cancer IQGAP3 activates Hedgehog signaling to confer stemness and metastasis via up-regulating GLI1 in lung cancer, SKP2 binds p27 and curcumol disrupts that binding Curcumol Induces G1 Phase Arrest in SK-Hep-1 Cells by Targeting SKP2-Mediated p27 Degradation, and BRCA1 binds senataxin at transcription pause sites BRCA1 Recruitment to Transcriptional Pause Sites Is Required for R-Loop-Driven DNA Damage Repair. Each study confirms a physical edge and then moves on; the accumulated set of validated physical interactions inside overexpression-driven cancer programs is never assembled into a predictive resource. Our core claim is that this discarded evidence, combined into a complex-topology layer inside functional enrichment, is specifically the missing ingredient for crossing the in vitro-to-in vivo gap — because the benchmark paper's transcriptomic models fail exactly where microenvironment-dependent complex remodeling matters most (they could not predict gemcitabine or paclitaxel xenograft responses at all, while reaching correlation of about 0.5 for erlotinib) Assessment of modelling strategies....

Research Program

The program has three connected components.

First, we curate a literature-scale map of co-immunoprecipitation-validated physical interactions reported in overexpression-driven cancer studies, seeded from the axes above (CKAP2–TFDP1, IQGAP3–GLI1, SKP2–p27, BRCA1–SETX) and extended by systematic screening of the same literature class: non-mutated drivers such as FUT8 in colorectal cancer Identification of hub gene associated with colorectal cancer: Integrating Mendelian randomization, transcriptome analysis and experimental verification, DHDH in hepatocellular carcinoma D-xylose suppresses hepatocellular carcinoma progression by regulating dihydrodiol dehydrogenase and remodeling the immune microenvironment, and m6A-modified circRAPGEF1 bound by IGF2BP3 m 6 A‐Modified circRAPGEF1 Interaction with IGF2BP3 Promotes Hepatocellular Carcinoma Progression via Reprogramming Aspartate Metabolism. Each physical edge is annotated with tissue context and the overexpression amplitude of its members. Edges that have only database (non-Co-IP) support are kept but flagged, so their contribution can be ablated separately. This is the resource the field lacks: existing PPI databases are static and unweighted, and the CSI paper demonstrated that using them without context specificity collapses predictive value to chance level on its own benchmark From random to predictive....

Second, we redefine the enrichment feature. Conventional overrepresentation analysis compares the fraction of a differentially expressed gene list falling in a pathway against a genome-wide background fraction, treating all members identically; we instead compute pathway scores over a graph in which Co-IP-validated physical complexes define tightly coupled subunits, and the expression amplitude of overexpressed oncogenes upweights the subunits they physically occupy. The hypothesis encoded in the weighting is that an overexpressed non-mutated driver acts through the complexes it binds, not through its mRNA level alone — the mechanistic picture established pair-by-pair in the CKAP2, IQGAP3, and SKP2 studies cited above. We implement this using the unified enrichment tooling of funkea: Functional Enrichment Analysis in Python, which integrates five enrichment frameworks and scales to large pharmacogenomic datasets. The output is a per-cell-line, per-xenograft feature vector of complex-aware pathway scores that replaces the raw gene-expression vector in standard drug response models.

Third, we test the mechanism directly in the laboratory, because the computational claim rests on an unverified premise: that protein complex composition actually differs between culture and tumor. We use the CKAP2 colorectal cancer system — cell lines HT29, HCT116, SW480 with established subcutaneous xenografts in nude mice CKAP2 Regulated by TFDP1... — and run unbiased co-immunoprecipitation coupled to mass spectrometry (the pull-down is digested and every captured protein is identified by mass spectrometry, rather than probing one hypothesized partner) on CKAP2 complexes from cultured cells and from excised tumors. The published work validated only the single CKAP2–TFDP1 edge by targeted Co-IP and never profiled the broader interactome; we measure how much of it remodels in vivo. Because Co-IP requires cell lysis, which can create false encounters and destroy weak complexes, we cross-validate any culture-versus-tumor difference with proximity ligation assays in intact cells and tissue sections, the standard in situ control for this lysis artifact. We run the same comparison on SKP2–p27 stoichiometry with and without curcumol, since that complex has a known drug disruptor and a known in vivo xenograft readout Curcumol Induces G1 Phase Arrest....

Proposed Experiments

The computational evaluation follows the benchmark design of the assessment study exactly, so results are directly comparable. Models are trained on the in vitro datasets it used — CCLE, CTRP, GDSC, and gCSI — and evaluated on in vivo response in the NIBR patient-derived xenograft encyclopedia, across its drug cohorts including erlotinib, gemcitabine, and paclitaxel. The benchmark paper reports that cell-line-trained models reached correlation of roughly 0.5 for the erlotinib-treated xenograft cohort but were unable to correctly predict responses for gemcitabine or paclitaxel, with average $R^2$ around 0.15 across seven drugs; it also found AUC-based sensitivity metrics significantly more predictable than $IC_{50}$ ($p < 0.001$) and that gene expression carried nearly all predictive power relative to mutations and copy number Assessment of modelling strategies.... We adopt AUC as the primary in vitro sensitivity label for that reason. Success is defined per drug cohort as xenograft response correlation clearly exceeding the baseline model on the same cohort, with the decisive test being the gemcitabine and paclitaxel cohorts where the baseline fails outright; if complex-aware features move those cohorts from unpredictable to correlation comparable to the erlotinib result, the mechanism claim is supported, and if they do not, the interactome-rewiring hypothesis for the gap is weakened and we will say so.

A second computational experiment isolates the weighting. The same response model is trained three ways on identical data and tuning budget: with raw gene-expression features (the benchmark configuration), with conventional topology-blind gene-set enrichment scores, and with the complex-aware scores. The second variant is the critical ablation, because it separates "aggregating genes into pathways helps" from "physical complex topology specifically helps" — a distinction the enrichment literature's own documented topology blindness predicts matters From random to predictive.... A fourth variant replaces Co-IP-validated edges with generic database PPI edges to measure what physical validation adds over curation.

The laboratory experiment closes the loop. If Co-IP-MS shows substantial interactome remodeling between culture and xenograft for the CKAP2 system, we test whether a predictor trained with culture-derived complex features misassigns drug sensitivity for compounds whose targets sit inside the remodeled complexes, and whether replacing those features with tumor-derived ones repairs the prediction on held-out xenografts. This is a small, decisive test of the causal story rather than a broad screening effort. For comparison context on the drug side, curcumol's in vivo efficacy against SKP2-driven xenografts is already established Curcumol Induces G1 Phase Arrest..., and VE-822 inhibits FUT8-overexpressing SW480 cells with an $IC_{50}$ of 1.245 µM in vitro with no in vivo testing at all Identification of hub gene... — exactly the situation, in vitro efficacy with unknown in vivo translation, that our predictor is meant to address.

Baselines

The naive reference point is a per-drug mean predictor: assign every xenograft the average response of the training cell lines for that drug, with no model at all. No reported result found on Lacuna for this reference point on the xenograft benchmark, so we will compute it ourselves as a floor; it establishes how much of any apparent accuracy is just the cohort's response distribution.

The strongest existing competitor is the benchmark paper's own modeling strategy — support vector machines, random forests, and gradient-boosted classifiers trained on baseline gene expression from CCLE, CTRP, GDSC, and gCSI and evaluated on the NIBR xenograft encyclopedia Assessment of modelling strategies.... It is the strongest competitor because it is the only published method evaluated on this exact cross-system question, and its reported numbers define the bar: average $R^2$ of approximately 0.15 across seven drugs (from 0.06 for sorafenib to 0.27 for PLX4720), correlation of roughly 0.5 for the erlotinib xenograft cohort, and failure to predict the gemcitabine and paclitaxel cohorts. We will re-run its best-performing configuration ourselves rather than quoting its numbers, using the same training data, the same drug cohorts, the same AUC labels, and the same model family and hyperparameter search budget granted to our method, so the only difference is the feature representation. Beating it means materially higher average $R^2$ across the drug panels and, most tellingly, non-trivial correlation on the gemcitabine and paclitaxel cohorts where it achieves none.

Two ablation baselines isolate the proposal's key ingredient within our own pipeline. One swaps the complex-aware pathway scores for conventional gene-set enrichment scores from the same pathways and the same tooling funkea, holding model, data, and tuning fixed; if this matches the complex-aware variant, aggregation rather than topology is doing the work and the central claim fails. The other swaps Co-IP-validated edges for unweighted database PPI edges; this measures whether physical validation in cancer-relevant contexts adds anything over generic interactome curation, the distinction the CSI work found decisive on its own adverse-event benchmark From random to predictive....

Why not just…?

Why not just use a bigger in vitro model, like GraphDRP? GraphDRP represents drugs as molecular graphs processed by graph neural networks and achieves striking in vitro numbers on GDSC — mixed-test RMSE of 0.0243 with a GIN encoder and Pearson correlation of 0.9308 with a combined GCN-GAT encoder Graph convolutional networks for drug response prediction. But it was evaluated only on in vitro $IC_{50}$, on binary mutation and copy-number features that the benchmark study showed contribute almost nothing beyond gene expression, and never on xenografts. The benchmark paper's core finding is that in vitro fit is not the bottleneck — its models already identify tissue type with 0.89 accuracy in xenografts while predicting drug response at $R^2 \approx 0.15$ — so more capacity on the same static features should not close a cross-system gap. That said, this is a genuine threat: capacity plus graph structure might learn complex-like features implicitly. Our re-run of the benchmark configuration under an equal tuning budget, and the explicit capacity-matched comparison, is what would settle it.

Why not just apply the CSI framework as-is? The context-specific interaction framework is the closest existing method: it showed that evaluating downstream interaction partners on a per-phenotype basis lifts AUROC from 0.51 to 0.77 on predicting severe drug side effects across 34 designated medical events From random to predictive.... We adopt its central principle. But it was built for post-marketing clinical adverse events on DrugBank targets and static, unweighted interactome databases; it has never been evaluated on cancer cell lines, oncogene overexpression, or in vivo tumor efficacy, and it contains no empirically validated physical complex layer. Because it is a real threat, we run a CSI-style global-PPI variant as the database-edge ablation described above: if context-specific weighting over a generic interactome matches Co-IP-weighted topology on the xenograft benchmark, then the physical-validation layer adds nothing and we will report that the contribution reduces to transplanting CSI to a new domain.

Why not just run more patient-derived xenograft experiments and skip the modeling? More in vivo data is always informative, but it is exactly what the benchmark study exists to avoid: the NIBR encyclopedia screening was done once, at scale, and the open problem it leaves is predicting the next compound's in vivo behavior without repeating it. The literature is full of the pattern that modeling should shortcut — enoxacin downregulates cell cycle and angiogenesis pathways in angiosarcoma cells in vitro with GO enrichment confirming the pathway-level effect, yet the authors note no comprehensive in vivo therapeutic validation exists Canonical microRNA loss drives tumor development implicating therapeutic efficacy of enoxacin in angiosarcoma, and VE-822 against FUT8 stops at an in vitro $IC_{50}$ Identification of hub gene.... Our wet-lab component is deliberately small and mechanistic — two systems, Co-IP-MS and proximity ligation — sized to test whether interactome remodeling is real, not to brute-force efficacy measurement.

Decisive Checks

The proposal stands or falls on two results. Computationally: do complex-aware enrichment features predict xenograft response on the gemcitabine and paclitaxel cohorts where the published benchmark achieves nothing, under matched data and tuning? Mechanistically: does unbiased Co-IP-MS reveal measurable interactome remodeling between culture and tumor in the CKAP2 system, confirmed in situ, and does incorporating the remodeled complexes repair mispredictions for drugs targeting them? If the first succeeds without the second, we have a useful feature engineering result but no mechanism, and we would report it as such. If the second succeeds without the first, interactome remodeling is real but not yet predictive at benchmark scale — a finding that would still redirect how the field uses its discarded Co-IP evidence.