When four drugs for the same cancer produce statistically indistinguishable response rates, progression-free survival, and one-year survival, clinicians are told to choose among them by toxicity alone. That is the situation in EGFR-mutant advanced non-small-cell lung cancer (NSCLC), where a multiple-treatments meta-analysis of 12 phase III trials ($N = 1{,}821$) found erlotinib, gefitinib, afatinib, and icotinib therapeutically equivalent on efficacy endpoints (1-year overall survival odds ratio 1.04, 95% CI 0.79–1.36) and ranked them only by adverse events — afatinib had a 100% Bayesian probability of causing the worst grade 3–4 diarrhea (Liang et al.). I propose that this "efficacy plateau" is partly an artifact of treating drugs as unlabeled arms $A$ and $B$ and patients as static baseline covariates. Two literatures say so but have never been joined: pharmacogenomic work shows that resistance to a single EGFR inhibitor is driven by highly heterogeneous, sample-specific bypass networks (Counterfactual Erlotinib Study), and longitudinal profiling shows tumors and hosts change state under therapy in ways that baseline snapshots cannot see — one patient's olaparib-resistant prostate cancer escaped not through the drug's target pathway but through wholesale neuroendocrine transdifferentiation ($MYCN$ up 4.8-fold, an $RB1$–$COL28A1$ fusion) (Horak et al.).
The core claim: comparative efficacy within a drug class can be predicted and explained per patient by (i) representing each drug not by its name but by its predicted binding targets propagated across the protein-interaction network and signed by the direction in which the drug perturbs those targets (the ingredient network pharmacology currently lacks), and (ii) conditioning the prediction on early on-treatment molecular kinetics rather than pre-treatment snapshots, with the resulting per-patient ranking explained by counterfactual pathway attribution instead of standard overrepresentation tests. The deliverable is a method that, where a population-level meta-analysis says "these drugs are equal," says which one this patient's tumor is most sensitive to and which pathway modules are responsible.
Core Claim
Formally, each drug $d$ in the class gets a vector $\mathbf{w}_d$ over pathway modules: predicted binding affinities from target-prediction repositories are propagated over the protein-protein interaction network (as in the morusin aortic-dissection study, which docked seven hub targets at $-9.0$ to $-7.0$ kcal/mol), but each propagated edge is weighted by the sign and magnitude of the drug-induced expression change of the downstream gene, taken from perturbation signatures. The predicted preference for drug $A$ over drug $B$ in patient $i$ is then
$$\hat\Delta_{AB}(i) = \big(\mathbf{m}_i(t_k)\big)^\top \big(\mathbf{w}_A - \mathbf{w}_B\big),$$
where $\mathbf{m}_i(t_k)$ is the patient's pathway-module state measured at an early on-treatment time point $t_k$ (from serial liquid biopsy or early repeat transcriptomics), not at baseline $t_0$. The novelty is entirely in the right-hand side's two ingredients: no existing comparative-efficacy method uses signed, propagated target-network representations of drugs, and none uses on-treatment kinetics. The pathway-module state $\mathbf{m}_i$ is computed with direction-aware enrichment — module scores reflect up- versus down-regulation — because standard overrepresentation analysis counts genes without signs and, more damningly, membership in a canonical KEGG pathway yields causal intervention hit rates of only ~40% versus 74–77% for mechanism-weighted panels (Counterfactual Erlotinib Study).
This sits deliberately between four existing works. DeepCRE already predicts cross-drug response for patients with a domain-separation network, but its drugs are unstructured embeddings on static baseline expression — no target prediction, no interactome, no time axis. The counterfactual erlotinib study does causal attribution but for a single agent on static pharmacogenomic profiles. Sequential regression and simulation manufactures head-to-head comparisons across trials without a common control arm, but only from demographic covariates — it was blind-validated against the prospective SEAVUE trial in Crohn's disease (simulated week-8 remission 44.9% ustekinumab / 45.9% adalimumab versus real 50.3% / 47.7%, $p = 0.9$), proving the counterfactual-simulation premise while leaving its covariate set entirely non-molecular. And Dr.VAE showed perturbation signatures improve sensitivity prediction, but from a single 6-hour snapshot, one drug at a time, with no comparative ranking. I take the causal-comparison premise from the first, the target-network machinery from network pharmacology, the perturbation-signing idea from the second, and move the observation window onto the treatment course.
Research Program
Three components, in increasing order of empirical risk. First, build the signed target-network drug representations for the EGFR-TKI family. Pull predicted targets for erlotinib, gefitinib, afatinib, and icotinib from the public repositories used by the morusin study (TargetNet, PharmMapper, SuperPred, SwissTargetPrediction); no reported predictive performance number for these repositories exists on Lacuna, so their output will be treated as a hypothesis to be validated downstream, not trusted input. Propagate affinities over the STRING interactome with signed edge weights derived from perturbation signatures (LINCS L1000 in cell lines; on-treatment biopsies in patients). The immediate testable prediction: afatinib (irreversible pan-ERBB binder) and gefitinib (reversible EGFR-selective) should differ on $\mathbf{w}_d$ precisely in modules where cell lines differ in cross-drug sensitivity — if the four vectors are near-identical, the network layer adds nothing and the proposal's first leg fails early. A second structural risk, documented directly in the file: basal expression of direct physical targets and their immediate interactors correlates poorly with drug sensitivity, and compact downstream panels beat whole-genome prediction ($R = 0.45$ versus $0.31$ across 961 cell lines and 265 drugs, Informative Genes Study). This is why the module state $\mathbf{m}_i$ uses propagated downstream modules, not the docking hubs themselves.
Second, the comparative prediction layer. Train the ranking model $\hat\Delta_{AB}(i)$ on GDSC/CCLE cross-drug response data (the erlotinib study draws on 571 cell lines with matched expression and mutation data; icotinib's availability in these panels is uncertain and must be checked — if absent, the cell-line phase proceeds with the three globally available TKIs and icotinib enters only the clinical phase). The output is a per-sample ranking of the class, evaluated by how often the model identifies the TKI with the lowest measured $IC_{50}$ for that specific line and by Kendall's $\tau$ between predicted and observed pairwise contrasts. Learn from DeepCRE's domain-separation design for the cell-line-to-patient transfer, since that paper already demonstrated the cell-line-to-TCGA bridge works (+17.7% over single-drug benchmarks; wet-lab confirmation in 5 of 8 patient-derived organoids).
Third, the longitudinal and clinical leg — the riskiest and the part no existing dataset fully supports, which I state plainly. No cohort on file offers within-patient, multi-time-point, multi-drug molecular sampling; the longitudinal evidence is single-arm or single-patient. So this component has two tiers. Tier one uses what exists: serial ctDNA cohorts like Fu et al.'s limited-stage SCLC study — whose Bayesian kinetic algorithm reached time-dependent AUC 0.796 in training and 0.745 in independent validation, and which showed consolidation-immunotherapy benefit confined to a high-risk kinetic tier (risk score > 0.457) (Fu et al.) — as the methodological template: early on-treatment ctDNA kinetics replace the static baseline $\mathbf{m}_i(t_0)$. Tier two is a prospective sub-study in EGFR-mutant NSCLC of sequential TKI use, exploiting the clinical reality that patients who progress on one TKI are routinely switched to another: each switch is a natural within-patient cross-drug comparison, with serial liquid biopsies before and 2–8 weeks after each switch. The target is roughly 40–60 patients yielding 25–40 within-patient switch comparisons. The endpoint is whether the kinetic-conditioned ranking predicts which switch drug yields deeper ctDNA clearance and longer progression-free survival. This is admittedly a data-generation exercise as much as a modeling one; the file's longitudinal cohorts (Horak et al.; Sailani et al., whose iPOP study tracked 105 people across 902 visits and found 1,133 seasonal analytes) prove serial multi-omic sampling is feasible at this scale, but none of them compares drugs.
The explanation layer runs throughout: for each predicted ranking, counterfactual pathway attribution — in the spirit of the BIRD-20 causal-normalizing-flow approach, which raised 3-gene intervention hit rates from ~40% to 74–77% over plain KEGG membership (Counterfactual Erlotinib Study) — identifies which module shifts would flip the ranking, and a subset of these is tested by CRISPR interventions in cell lines and by an In-SMARchip-style organoid assay as DeepCRE did.
Baselines
The "do nothing" reference point is the current standard of care implied by the meta-analysis: treat the four TKIs as interchangeable on efficacy and choose by toxicity hierarchy (Liang et al.). Concretely, this baseline assigns every patient the gefitinib/icotinib-family preference and expects zero per-patient molecular discrimination; any model whose per-line drug ranking beats chance is doing something the plateau says is impossible, so even modest ranking accuracy against this reference is informative. A second naive reference is the SUCRA-style population ranking from a Bayesian network meta-analysis applied uniformly to all patients, which by construction cannot vary by patient.
The strongest existing competitor is DeepCRE's Domain Separation Network (DeepCRE), the only published system that predicts patient-level cross-drug response; it outperformed single-drug-learning benchmarks (DrugCell, PaccMann) by an average of 17.7% and achieved leading results in 12 of 13 TCGA tumor types. It will be run on exactly the same GDSC/CCLE and TCGA inputs, the same train/validation splits, and the same tuning budget as the proposed model, so the only systematic difference is the drug representation (signed propagated target networks versus unstructured drug embeddings) and the observation window (kinetics versus baseline). Beating DeepCRE means a higher per-sample correct-best-drug rate and higher pairwise contrast $\tau$ under identical evaluation — and the 17.7% margin it holds over single-drug models is the scale of improvement that counts as meaningful here. Dr.VAE (Dr.VAE) runs as a secondary baseline for the perturbation ingredient; it beat Ridge logistic regression, random forest, and SVM-RBF on at least 14 of 26 drugs (53.8%), with AUROC gains up to 4.4%, but it predicts one drug at a time, so it enters the comparison by ranking drugs through separately trained models.
Two ablations isolate the proposal's key ingredients. The network ablation replaces $\mathbf{w}_d$ with a one-hot drug identity vector, keeping the architecture and data identical; if performance does not drop, the network-pharmacology ingredient is decoration. The kinetics ablation sets $\mathbf{m}_i(t_k)$ to $\mathbf{m}_i(t_0)$, collapsing to a static baseline model; the difference measures what the longitudinal window buys. A third, smaller ablation strips the sign from propagated edges (unsigned topology, as in standard network pharmacology) to test whether causal directionality — the field's acknowledged blind spot — actually matters for ranking.
Why not just…?
Why not just run a Bayesian network meta-analysis? It is the accepted tool, and the population-level plateau result is real. But network meta-analysis assumes transitivity across trials, degrades to "Very Low" certainty under GRADE when trials are heterogeneous — as in the 40-trial, 3,088-patient TCM–metformin network where safety reporting existed in only 20% of trials (Che et al.) — and, most importantly for this proposal, it averages away exactly the sample-specific bypass-resistance structure that the counterfactual erlotinib study demonstrated (Counterfactual Erlotinib Study). The plateau and per-patient heterogeneity are not contradictory; they are two views of the same data. This proposal uses the NMA as its population anchor and tests whether molecular features explain the residual between-trial heterogeneity the NMA cannot — a complementary, not competing, role.
Why not just use DeepCRE or another static baseline model? This is the real threat, since DeepCRE already does cross-drug prediction well and my cell-line and TCGA phases will lose to it if static baselines suffice. The honest answer is that the decisive evidence does not yet exist and the switch-study is designed to produce it: Horak's patient (Horak et al.) acquired resistance through a lineage switch invisible to any pre-treatment profile, and Fu et al.'s dynamic ctDNA algorithm stratified immunotherapy benefit that static assessment could not (Fu et al.). If, on the switch cohort, the static-baseline version of my own model (the kinetics ablation) matches the kinetic version, the longitudinal claim is dead and I will report that — the ablation exists precisely to let this alternative win cheaply if it is right.
Why not just measure the direct target? All four TKIs bind EGFR; if direct-target logic worked, the plateau would be molecularly confirmed and this proposal pointless. But the informative-genes study showed basal expression of known physical targets and their immediate interactors correlates poorly with sensitivity across 961 cell lines and 265 drugs ($R = 0.31$ for whole-genome prediction, lifted to $0.45$ only by compact downstream panels, Informative Genes Study), and Horak's patient escaped through a bypass with no reversion in the target gene. The unsigned-topology ablation settles this empirically within the proposal: if direct-target and hub-expression features match the propagated signed modules, the network layer was unnecessary and I will have documented that.
Why not just do standard KEGG/GO overrepresentation on the on-treatment biopsies? It is cheap and interpretable. But it is direction-blind, inflated by co-regulated genes, and — measured directly in the file — pathway membership alone yields ~40% causal intervention hit rates versus 74–77% for mechanism-weighted attribution (Counterfactual Erlotinib Study). The proposal uses ORA as a comparator in the attribution experiments: the decisive check is whether counterfactual module attribution predicts which interventions flip drug rankings at a rate ORA-derived panels do not.
Proposed Experiments and Decisive Checks
Experiment one (retrospective, cell lines): on GDSC/CCLE EGFR-TKI response data, evaluate per-line correct-best-drug rate and pairwise contrast $\tau$ for the full model against DeepCRE, Dr.VAE-as-ranking, single-drug learners, one-hot and unsigned-topology ablations, and the plateau reference. Decisive check: the full model must beat DeepCRE under identical data and tuning, and the one-hot ablation must underperform the signed-network representation — otherwise the network-pharmacology claim fails. Experiment two (transfer): apply the trained model to TCGA EGFR-mutant tumors with recorded EGFR-TKI response and progression-free survival, mirroring DeepCRE's evaluation regime; note the validation-collapse risk documented in TCGA predictive modeling, where a gemcitabine model held 82.3% cross-validation and 85.7% validation accuracy while the same pipeline for 5-fluorouracil collapsed from 84.1% to 52.9% (Leveraging TCGA Data) — so external validation on a held-out tumor source, not cross-validation, is the standard. Experiment three (mechanistic): counterfactual attribution identifies predicted ranking-flipping modules; CRISPR interventions in selected cell lines measure the intervention hit rate, with the 74–77% BIRD-20 figure as the reference scale and ~40% (plain pathway membership) as the floor to beat. Experiment four (longitudinal, prospective): the sequential-TKI switch sub-study, with early post-switch ctDNA kinetics feeding $\mathbf{m}_i(t_k)$; the endpoint is whether the kinetic-conditioned ranking predicts ctDNA clearance depth and time to progression on the switched drug, benchmarked for scale against Fu et al.'s reported time-dependent AUC of 0.796 (training) and 0.745 (validation). The single most decisive outcome across all four: a patient subset, invisible to the population meta-analysis, in which the model correctly calls the best TKI and the attribution layer names a module whose perturbation experimentally flips the call. If instead the kinetic and signed-network ablations match the full model everywhere, the plateau is real at every resolution this toolkit offers — itself a publishable and consequential negative result, since it would say the missing structure lies somewhere none of these four literatures currently looks.
References
- Network Meta-Analysis of Erlotinib, Gefitinib, Afatinib and Icotinib in Patients with Advanced Non-Small-Cell Lung Cancer Harboring EGFR Mutations
- Identification of Heterogeneous Erlotinib Response Gene Sets Using Sample-Specific Counterfactual Causal Attribution
- Response to olaparib in a PALB2 germline mutated prostate cancer and genetic events associated with resistance
- Integrated Network Pharmacology, Single‐Cell Transcriptomics Unveil the Mechanistic Role of Morusin in Aortic Dissection
- DeepCRE: Transforming Drug R&D via AI-Driven Cross-drug Response Evaluation
- Sequential regression and simulation: a method for estimating causal effects from heterogeneous clinical trials without a common control group
- Dr.VAE: improving drug response prediction via modeling of drug perturbation effects
- Modeling cancer drug response through drug-specific informative genes
- Integrated circulating tumor DNA-based prognostic algorithm for limited stage small-cell lung cancer under definitive chemoradiotherapy and utility in consolidation immunotherapy benefit prediction
- Deep longitudinal multiomics profiling reveals two biological seasonal patterns in California
- Efficacy and safety of traditional Chinese classic prescriptions combined with metformin in the treatment of type 2 diabetes mellitus: a Bayesian network meta-analysis
- Leveraging TCGA gene expression data to build predictive models for cancer drug response