Do Herbal Target Maps Carry Survival Signal Beyond Hub Bias?

Network pharmacology studies of traditional herbal formulas follow a common recipe. They look up the formula's chemical constituents in a database, collect the human proteins those constituents are predicted to bind, and intersect that list with disease genes. The overlap is then presented as the mechanism. Across very different formulas and diseases, the "core targets" come out nearly the same: AKT1, TP53, STAT3, TNF, and the PI3K-Akt pathway. That pattern should worry us. These are also the most-studied, most-connected proteins in human biology, so they dominate any database-driven lookup whatever herb goes in.

This note proposes a survival-based test of whether a formula's predicted target map says anything specific. The test uses patient survival data and a regularized Cox model. A Cox model is the standard regression for time-to-event outcomes such as death or relapse; it handles patients whose follow-up ends before the event. We ask whether weighting genes by how heavily a formula is predicted to hit them improves survival prediction beyond two things: clinical covariates, and a control built from equally "hubby" but herb-irrelevant target maps. If it does not, the survival plots and hub-gene claims in this literature are reporting database structure rather than pharmacology.

Core Claim

The claim is methodological and falsifiable. Take a formula's herb–compound–target map and turn it into gene-specific penalty weights in a Cox model. Genes hit by many of the formula's compounds, with high confidence, are shrunk less. Protect clinical covariates from penalization. The formula-derived weights should then improve out-of-cohort survival prediction in the disease the formula is used for, and they should do so by more than two controls. The first control is weights from degree-preserving rewirings of the same map. The second is weights from formulas used for unrelated diseases.

We expect a mixed result. Much of the apparent signal should be explained by hubness alone. Whatever survives the rewiring control is the part of the target map that is actually specific to the herb. Isolating that residual is what would be new.

Two caveats matter. The patients in public tumor cohorts did not take these herbs. So the test does not measure whether a formula prolongs survival. It measures whether the formula's predicted targets are prognostically relevant to the disease, which is the claim that network pharmacology papers implicitly make when they show target-gene survival curves. And because the test operates on predicted targets, it inherits the database's blindness to absorption and metabolism. We treat that as a secondary axis to probe, not something the design fixes.

Research Program

Prior work that motivates the design. Wei et al. deconstruct Fufang Banmao Capsule for non-small cell lung cancer (Integrating Network Pharmacology And Mendelian Randomization To Explore Potential Targets Of Fufang Banmao Capsule Against Non-Small Cell Lung Cancer). They map 230 ingredients to 786 targets through BATMAN-TCM 2.0, intersect those with 152 disease genes, and then narrow to seven causal candidates by Mendelian randomization. They never fit a time-to-event model, and they acknowledge ignoring pharmacokinetics.

Two other studies land on the same hub-heavy pathway sets from quite different herbs. Zheng et al. map Danshen compounds to PI3K-Akt for arterial restenosis (Lu-Jing Zheng et al.). Lei et al. find 36 shared heart-failure targets (Feifei Lei et al.). A multi-omics review describes PI3K/Akt as a recurrent "master switch" across formulas (Multi-omics technology for demystifying herbal medicine). We read that recurrence as the symptom to test, not as a finding.

Prior work that supplies the method. DegreeCox regularizes a Cox model with a quadratic penalty weighted by inverse node degree in a gene network, and shows better cross-cohort generalization in three ovarian cancer cohorts (DegreeCox - a network-based regularization method for survival analysis). Our estimator keeps DegreeCox's form but changes the source of the weights. Where DegreeCox asks how connected a gene is inside the human network, we ask how heavily a specific formula is predicted to engage the gene. The penalized partial likelihood is

$$\hat\beta = \arg\min_\beta \; -\ell(\beta, \gamma) + \lambda \sum_j w_j \beta_j^2 + \alpha\lambda \sum_j w_j |\beta_j|,$$

The terms are as follows:

  • $\ell(\beta,\gamma)$ is the Cox log partial likelihood with gene coefficients $\beta$ and unpenalized clinical coefficients $\gamma$.
  • $\lambda$ and $\alpha$ are tuning constants.
  • $w_j = 1/(\epsilon + c_j)$ is the penalty weight for gene $j$.
  • $c_j$ is the formula's coverage of gene $j$: the sum, over the formula's compounds predicted to hit $j$, of the database confidence score.
  • $\epsilon$ is a small constant.

Keeping clinical covariates unpenalized follows directly from the TCGA benchmark of Herrmann et al. (Large-scale benchmark study of survival prediction methods using multi-omics data). Across 18 cancers, that benchmark found flat omics models usually lose to a clinical-only Cox model, while methods that favor clinical variables hold their ground.

Prior work that constrains the design. Iuliano et al. show that network-penalized Cox models are only as good as the network's edges (Cancer Markers Selection Using Network-Based Cox Regression: A Methodological and Computational Practice). HTINet2 reports false positives and naming conflicts in herb–target databases (HTINet2: herb-target prediction via knowledge graph embedding and residual-like graph neural network). Both points push us to treat the target map as a noisy prior, and to build controls that keep its degree structure while breaking its herb specificity.

What we build. The build has four parts:

  1. Coverage vectors $c$, computed for several formulas from BATMAN-TCM 2.0 and from SymMap's filtered set of 38,002 herb–target relations. Using both databases checks that results do not depend on one source.
  2. A null family. It includes degree-preserving rewirings of each herb–compound–target bipartite graph, target maps of formulas prescribed for unrelated conditions, and random gene sets matched on protein-interaction degree.
  3. A secondary "predicted-coverage" variant. Here $c$ comes from heNetRW's random-walk scores (Heterogeneous network propagation for herb target identification) instead of curated links. This tests whether target-prediction models add or remove specificity.
  4. Where the database supplies absorption-related compound screening, a bioavailability-restricted variant of $c$. This is a first, admittedly crude, probe of the absorption gap Wei et al. flag.

Baselines

Clinical-only Cox model. A Cox model with age, sex, stage, and whatever other clinical fields each TCGA cohort provides. This is the naive reference point and, per Herrmann et al., the hardest one to beat in most cancers. The benchmark reports that only Block Forest marginally outperformed it on average. The file gives no per-cancer C-index value, so there is no reported result found on Lacuna for the clinical Cox baseline on TCGA lung adenocarcinoma. We fit it ourselves on identical folds. Beating it means a gain in Uno's C-index and integrated Brier score that is consistent in sign across repeated outer cross-validation splits and survives transfer to a held-out cohort.

DegreeCox and Net-Cox. DegreeCox with STRING-derived degree weights, and Net-Cox with a Laplacian penalty, both with clinical covariates unpenalized. These are the strongest structured competitors, and they encode exactly the hub prior we suspect the herbal maps reproduce. The file reports no numbers for DegreeCox or Net-Cox on TCGA lung cohorts. Beating them is the central test.

Elastic-net Cox. An elastic-net Cox model with uniform weights, as a sparsity baseline.

Block Forest and DeepSurv. Block Forest, the one method in the TCGA benchmark that edged out clinical Cox, and DeepSurv (DeepSurv: personalized treatment recommender system using a Cox proportional hazards deep neural network) as the flexible nonlinear competitor. For context only, since it is a different dataset, DeepSurv reports a C-index of 0.867 against 0.816 for linear Cox on METABRIC.

Fairness rules. All methods share the same patients, folds, and preprocessing, and the same gene universe: genes present in both the expression data and the interaction network. Each method gets the same hyperparameter search budget, tuned by nested cross-validation.

The key ablation. This is the rewired-coverage model. It is identical to ours except that $c$ comes from a degree-preserving rewiring of the same formula's bipartite graph, averaged over many rewirings. Beating this ablation, and not just the clinical model, is the only result that supports herb specificity.

Why not just…?

Why not just fit the clinical Cox model and stop? This is a genuine threat. Herrmann et al. show molecular features rarely help. But our question is not whether omics beat clinical variables. It is whether a formula's map helps more than a matched null does. If every molecular arm ties the clinical model, that too is informative: network pharmacology target lists would then carry no measurable prognostic content in that disease. The lower-grade glioma cohort, where Herrmann et al. found molecular data did add value, is our positive control that the pipeline can detect signal when it exists.

Why not just use DegreeCox with human-network degree? This is the most serious alternative. We expect formula coverage to correlate strongly with human-network degree. The comparison is therefore built in as a hypothesis arm rather than argued away. If formula weights match DegreeCox and also match the rewired null, the conclusion is that herbal maps are a noisy proxy for hubness. We would report that as the main result.

Why not just do enrichment analysis and Kaplan–Meier curves on a hub gene, as most herbal papers do? Kaplan–Meier curves split on one gene cannot adjust for stage or age. They also collapse a formula's many targets into one variable. Just as importantly, a hub gene such as AKT1 would be chosen for almost any formula, which is precisely the non-specificity at issue. We run this procedure as an extra arm with the same null family, to show how often it "finds" prognostic targets for irrelevant formulas.

Why not just use DeepSurv or a random survival forest? Flexible models may predict slightly better. But a flexible model's gain cannot be credited to the formula map, because the map does not enter it. Our question concerns the prior, not the prediction ceiling. We include DeepSurv so readers can see how much accuracy the interpretable model gives up.

Proposed Experiments

Primary setting. Fufang Banmao Capsule, with Wei et al.'s published ingredient list, on TCGA lung adenocarcinoma and squamous cohorts.

Specificity contrasts. As off-indication formulas, we use the heart-failure herb set of Lei et al. and the Danshen restenosis set of Zheng et al. Their maps are rich in PI3K-Akt but were built for cardiovascular disease.

Positive control. Lower-grade glioma, where molecular features are known to help. We also re-test on the ovarian cohorts used by DegreeCox, to check our reimplementation against its reported qualitative behavior.

Metrics. Uno's C-index and integrated Brier score under repeated nested cross-validation. We also measure signature stability, as the overlap of selected genes across folds, which is the property DegreeCox emphasizes. Proportional-hazards diagnostics are reported for the selected genes, so any time-varying effects are visible rather than hidden.

Decisive check. The central number is the gap between the formula-weighted model and the mean of its degree-preserving rewirings, expressed as an empirical p-value over at least 1,000 rewirings. A further check requires the on-indication formula to beat the off-indication formulas in its own disease, but not in theirs.

How to read the outcome. A positive gap that holds in both databases and survives transfer to a held-out cohort would be the first survival-grounded evidence that a herbal target map encodes something beyond database hub bias. A null gap would be an equally concrete, and probably more common, result. It would give the field a cheap audit that any network pharmacology claim can be run through before it is published.