Two Structural Regimes, Six Cancer Types, Thirteen Experiments: What Network Topology Reveals About Drug Response

The Question

Every protein in a cancer signaling network has a position. It has more or fewer connections. Its presence -- or absence -- affects how well the physical interaction layer and the co-expression layer align with each other. This alignment is measurable. We call the change in alignment when a node is removed its delta-r (change in inter-layer Pearson correlation).

The question: does that structural position predict drug response? If you build a cancer network, measure every protein's delta-r, and then look at which proteins are the most sensitive drug targets in cancer cell line screens -- do the two rankings agree?

The short answer after 13 pre-registered experiments: sometimes yes, sometimes no, and the cases where the answer is no are as informative as the cases where it is yes.

The Framework

We represent each cancer type as a two-layer multiplex network:

Layer A: Physical protein-protein interactions. Which proteins physically bind each other, measured by co-immunoprecipitation, two-hybrid screens, affinity proteomics. This layer is universal -- it reflects biochemical binding capacity, not cancer state.

Layer B: Cancer co-expression. Which proteins are transcriptionally coupled in actual cancer tumor samples or cancer cell lines. This layer is cancer-state conditional -- it reflects the specific co-activation patterns of the cancer being modeled.

The two layers together define a multiplex network. For each protein, we ask: how much does removing it change the alignment between these two layers? That change (delta-r) is the structural metric. Proteins whose removal increases alignment are structural DAMPERS. Proteins whose removal decreases alignment are structural ANCHORS.

To correct for the fact that highly-connected proteins will trivially show large delta-r (more connections = more disruption regardless of structural role), we compute z_dr: a degree-normalized version of delta-r using null models that preserve degree structure while randomizing topology. z_dr measures topology, not connectivity.

The Critical Requirement: Layer B Must Match the Biology

Before we get to the cancer results, a methodological finding that shaped everything else.

Layer B must be built from data that is conditioned on the same biological state as the drug response you are predicting. If you use pan-tissue co-expression data (measuring gene co-expression across all tissues and conditions), the result is a nearly empty second layer for cancer pathway proteins. Why? Because cancer proteins like KRAS, MEK1, CDK4, and TP53 are co-expressed specifically in cancer -- their joint activation is context-specific. In pan-tissue data, this signal disappears.

When we tested this directly: a 100-node cancer pathway network built with STRING pan-tissue co-expression produced 3 edges in Layer B, compared to 555 edges in Layer A. You cannot measure inter-layer alignment with 3 edges in one layer.

Consequence: the framework works because Layer B is cancer-specific (TCGA tumor data, cancer cell line co-expression). That is not a limitation -- it is the design. The structural predictions are about cancer topology, not generic PPI topology.

Two Structural Regimes

After running the framework across six cancer types -- KRAS-mutant NSCLC, colorectal cancer, pancreatic cancer, GBM, BRAF-V600E melanoma, and EGFR-mutant NSCLC -- a pattern emerged.

Five of the six cancer types fall into what we call the Over-Damped Signaling (ODS) regime:
• Structural DAMPERS (positive z_dr) are the drug targets that show strongest cell killing
• Structural ANCHORS (negative z_dr) are less sensitive to their targeted drugs
• The KRAS driver creates a network state where removing damper nodes is maximally disruptive

One cancer type -- BRAF-V600E melanoma -- falls into a different regime (CLASS 6b):
• The relationship inverts. Structural ANCHORS are the most sensitive drug targets
• The BRAF-V600E driver creates direct pharmacological addiction to the MAPK cascade anchor nodes
• The BRAF mutation creates a dependency that exceeds what the topological position alone would predict

This is not a failure of the framework. Both regimes are detectable from the structural analysis before looking at any drug data. The driver gene's own structural position predicts which regime applies: if the driver is an anchor in the cancer network, the cancer tends to be driver-addicted (CLASS 6b). If the driver is neutral or a damper, the ODS regime applies.

The MEK1 Structural Switch

The clearest illustration of the two regimes is MEK1.

In colorectal cancer: MEK1 is a structural DAMPER. Removing it from the network increases inter-layer alignment. The network does not need MEK1 to stay coherent -- it is using MEK1 as a signal amplifier that creates structural decoupling. Pre-registered prediction: MEK inhibitors kill colorectal cancer cells preferentially. Result: confirmed.

In BRAF-V600E melanoma: MEK1 is a structural ANCHOR. Removing it collapses inter-layer alignment. The network depends on MEK1 to maintain the coupled state the BRAF oncogene creates. Pre-registered prediction: MEK inhibitors kill BRAF melanoma cells more strongly than colorectal. Result: confirmed, and with substantially larger effect sizes when tested on an independent assay platform (GDSC2 multi-point IC50 curves vs the original PRISM single-concentration data).

The same drug. The same protein. Opposite structural positions. Opposite predictions about relative efficacy. Both confirmed.

CDK4: The Scale-Robust Signal

CDK4 (cyclin-dependent kinase 4) appears as a structural DAMPER in every cancer type we have tested: colorectal, pancreatic, BRAF-melanoma. It also appears as a DAMPER when we scale the network from 16 nodes (curated pathway network) to 100 nodes (STRING protein interaction network), even in the methodologically stressed 100-node test where Layer B had only 3 edges.

CDK4 has z_dr above the robust threshold even in the 100-node test. It is the only protein that survives both the biological stress test (multiple cancer types) and the methodological stress test (degraded Layer B at scale). Its structural load-bearing position in cancer networks appears to be a property of the human protein interactome generally, not an artifact of small curated network design.

Note: CDK4/6 inhibitors (palbociclib, ribociclib, abemaciclib) show near-zero cell killing in standard drug response assays -- but this is not a structural failure. CDK4/6 inhibitors are cytostatic: they arrest cells in G1 phase (growth stops) but cells survive. The drug response assay (measuring cell death) does not capture cytostasis. The structural prediction is about sensitivity to disruption; the measurement constraint is about the readout. This is one of three confirmed failure modes in the framework's boundary condition table.

When the Structural Prediction Fails: The Three Failure Modes

We pre-registered a systematic investigation of when structural predictions do NOT transfer to drug efficacy. Three failure modes were confirmed:

Pathway disconnection: A protein has a robust structural signal but its drug targets a pathway that does not connect to the cancer's primary driver. Structural position predicts target importance in the context of the network; if the drug's effect cannot reach the driver dependency, the killing does not happen regardless of structural score.

Indirect pathway: A protein is a structural anchor but the drug reaches it only through a multi-step indirect route. The indirect path creates enough slack that blocking the terminal node does not sufficiently perturb the driver-dependent signaling axis.

Genotype-mechanism mismatch: A protein has a robust structural signal and the drug directly targets it, but the drug's mechanism requires a specific genotype to be active. The most concrete case: MDM2 structural inhibitors (nutlins) require wild-type TP53 to function. In a cancer where TP53 is frequently mutated, the drug cannot engage its mechanism regardless of MDM2's structural score.

All three failure modes were pre-registered before looking at drug data. All three were confirmed. The framework's failure modes are as precisely characterized as its successes.

Cross-Cancer Structural Patterns

When we computed the structural similarity between all six cancer types (cosine similarity between their z_dr vectors), the result contradicted the obvious expectation.

Two cancer types with the same driver mutation are structurally LESS similar to each other than two cancer types with different driver mutations. KRAS-mutant colorectal cancer and KRAS-mutant lung cancer are nearly structurally orthogonal. Meanwhile, pancreatic cancer and BRAF melanoma -- different drivers, different tissues, different signaling regimes -- share the most structural similarity in the atlas.

The explanation: a driver mutation is a single-node perturbation. The entire surrounding tissue environment (co-expression patterns shaped by the tissue of origin, epigenetic state, stromal interactions) shapes the network topology far more than which gene has the mutation. Driver identity is not a reliable predictor of structural type.

The pancreatic / melanoma structural partnership has a specific mechanistic basis: both cancer types have strong p53-pathway structural co-amplification (CDK4, MDM2, TP53 are all structural dampers in both). PDAC has high TP53 mutation frequency; melanoma has CDKN2A deletion releasing CDK4. Different routes to the same structural configuration.

Measurement Geometry Invariance

One concern for any framework built on a single assay platform is whether the results are platform-specific. We tested this directly: took the pre-registered predictions built from PRISM data (single-concentration log-fold-change, Broad Institute) and re-tested them on GDSC2 data (multi-point IC50 dose-response curves, Sanger Institute). Different assay geometry. Different cell line panels. Different continents.

The predictions transferred. The MEK inhibitor regime-switch (BRAF-melanoma more sensitive than colorectal) and the PI3K preferential sensitivity in colorectal cancer both replicated on GDSC2 with larger effect sizes than the original PRISM test. The structural signal is not an artifact of the PRISM measurement geometry.

What Is Being Tested Next

The most important open question in the framework is whether the full structural signal survives at 100-node scale when Layer B is built correctly -- from TCGA cancer-specific co-expression rather than pan-tissue data.

The 100-node test with pan-tissue data showed only CDK4 surviving as a robust signal. The 16-node results with cancer-specific data showed the full framework. Building the 100-node network with TCGA colorectal tumor RNA-Seq as Layer B (D14, pre-registration pending) will determine whether the layer source is the causal variable -- and whether structural pharmacology scales.

If the full signal is recovered at 100 nodes with the correct Layer B, the framework transitions from a curated-network result to a genome-scale structural pharmacology method.

All 13 pre-registration files, with timestamps and immutable hashes, are at github.com/vladi160/preregistrations. The analysis is being prepared for publication.