When a sentence such as "John loves Mary, and Paul, Virginia" omits its second verb, readers recover the meaning by lining up the second clause against the first, slot by slot: subject with subject, object with object. Natural-language processing gained real accuracy when parsers began scoring this kind of structural parallelism directly, rather than reading each conjunct on its own. We propose to carry that idea, slot-by-slot alignment across parallel branches, into cancer prognosis.
Metabolism has its own coordinate structures. In liver cancer, a single fuel molecule, acetyl-CoA, is rerouted into three downstream branches at once: N-glycan synthesis, the mevalonate (cholesterol) pathway, and sphingolipid synthesis. Current survival models score each gene, or each pathway's average activity, as an independent term. Our claim is that the shape of rewiring carries prognostic information beyond how strongly each branch is switched on. The shape question is whether the corresponding control points in each branch rise together, for example committed step with committed step and terminal step with terminal step. We will define a per-patient score that measures this role-aligned symmetry. We will then test, with an ablation that shuffles role assignments, whether alignment by biological role adds survival information beyond branch activity levels and their interactions.
Core Claim
The precise hypothesis is that tumors which rewire parallel branches at matched functional positions have worse survival than tumors with equal overall branch activity but mismatched internal profiles. The matched-position pattern would point to a coordinated upstream program. A mismatched profile would suggest piecemeal, branch-local changes. The new element is the comparison of corresponding roles across branches within one patient. Existing scores cover two other things. Pathway averages capture level. Co-expression modules capture correlation across patients. Neither compares within a patient whether the same functional step is elevated in each branch.
The source of this structure is concrete. Barace et al. built gene co-expression graphs on liver cancer data from The Cancer Genome Atlas (TCGA-LIHC) and from the Japanese cohort of the International Cancer Genome Consortium (ICGC-LIRI-JP). Using community detection, they found the "acetyl-CoA pivot" and confirmed through CRISPR knockout screens in the Cancer Dependency Map (DepMap) that HMGCR, ALG14 and the transporter SLC33A1 are essential to cancer cells. Their work motivates ours: it establishes that the three branches exist and are biologically real. It stops at modules, though, and never asks whether patients differ in how the branches line up, or whether that difference predicts outcome.
Where the Linguistic Machinery Comes In
Three parsing ideas supply the method, each in a narrow role.
Kato and Matsubara resolved gapping, sentences like the example above where a verb is dropped, by tagging constituents with grammatical roles. They then aligned the leftover phrases in the incomplete clause ("remnants") to their counterparts in the full clause ("correlates") under order-preserving role constraints, raising gapping F-measure from 12.9 to 70.8. We borrow the principle that alignment should run over role labels rather than raw positions. Each branch's genes are assigned pathway-position roles: entry or substrate supply, committed or rate-limiting step, intermediate steps, terminal output, and transcriptional regulator. These assignments come from pathway annotation and are fixed before any survival data is examined.
Hogan improved coordinate noun-phrase parsing by generating the second conjunct conditioned on the structure of the first. The same conditioning gives us a residual variant of the score: how far a branch's role profile deviates from what the other branches predict.
Ficler and Goldberg used a Siamese network, two copies of one encoder whose outputs are compared, to minimize distance between the label paths of symmetric conjuncts. They reached 72.7 F1 on Penn Treebank coordination against 68.9 for the Berkeley parser. Their result supports learned symmetry scoring. However, a few hundred patients per cohort is far too little data to train such a network. So the primary score is fixed, and a learned version is only a secondary experiment.
Kehler showed that only some discourse relations force element-by-element alignment of parallel clauses, while others need only overall meaning to cohere. We turn this into a testable expectation rather than a metaphor. Some metabolic branch points should be "alignment-demanding," meaning role-level symmetry predicts survival. Others should be "level-only," meaning branch averages suffice. A pan-cancer screen across branch points can tell the two kinds apart.
The Score
For patient $i$, let $z_{ibr}$ be the standardized tumor expression of the gene, or the averaged genes, filling role $r$ in branch $b$, with $B$ branches and $R$ roles. Let $\bar z_{ib}$ be that branch's mean over roles. Subtracting $\bar z_{ib}$ removes branch level, which pathway-average baselines already capture, and leaves only the internal profile. The role-aligned parallelism score is
$$P_i = -\frac{2}{B(B-1)} \sum_{b<b'} \sum_{r=1}^{R} \Big[(z_{ibr}-\bar z_{ib}) - (z_{ib'r}-\bar z_{ib'})\Big]^2 ,$$
so $P_i$ is highest when every branch shows the same pattern of which steps are up and which are down. $P_i$ then enters a Cox proportional hazards model, the standard regression for time-to-death data, alongside clinical stage and the branch means. The Hogan-style variant replaces the pairwise distance with the out-of-sample residual of a ridge regression predicting each branch's centered profile from the others, fit only on training patients.
Research Program
The work proceeds in four stages.
Fix the structure. Using TCGA-LIHC tumors only, we fix the branch gene lists and role assignments before touching outcomes. We restrict to tumor-co-expressed genes, following Barace et al.'s warning that normal-liver metabolic genes are broadly downregulated in tumors and swamp the signal.
Compute and validate. We compute $P_i$ and its residual variant, fit Cox models in TCGA-LIHC, and evaluate on ICGC-LIRI-JP with no refitting of roles or gene lists. ICGC-LIRI-JP has tumor expression and survival follow-up, but the file reports no survival C-index for it, so there is no prior number to compare against there.
Screen for alignment-demanding branch points. We build analogous branch-point structures for other one-source, many-branch nodes and run the same comparison across TCGA's 33 cohorts. Examples are the glycolysis outflow points tied to PFKFB3, which the PFKFB3 pan-cancer study linked to survival and immune checkpoints, and the cholesterol arm studied in Zhao et al.'s myeloma work. Branch points where role-aligned symmetry adds information get labeled alignment-demanding.
Check mechanism in cell lines. We test in DepMap whether cell lines with high $P_i$ depend more on the shared upstream genes, SLC33A1 and HMGCR, than low-$P_i$ lines with equal branch means. A coordinated program should concentrate dependence at the hub.
Baselines
The naive reference point is a Cox model on clinical stage and age alone. Every other model must clear it on the external cohort.
The strongest established competitor is a gene-level penalized Cox signature of the kind Wu et al. built for colon cancer. They used LASSO, a regression that zeroes out most coefficients, plus multivariate Cox regression to reach a 9-gene signature, reporting time-dependent survival AUC of 0.740, 0.723 and 0.763 at 1, 3 and 5 years in their TCGA-COAD training cohort. We found no reported result on Lacuna for an acetyl-CoA-branch signature in TCGA-LIHC or ICGC-LIRI-JP. We will therefore fit the Wu-style pipeline ourselves on exactly the union of branch genes, with the same cross-validation folds and tuning budget.
The pathway-level competitor follows the TCGA pathway-versus-gene comparison, which found pathway scores from Gene Set Variation Analysis (GSVA) comparable to gene-level models (mean C-index around 0.5 to 0.8 across cohorts) and more stable. We compute per-branch GSVA scores on the Barace modules and fit penalized Cox on them. This is the "branch level only" model that $P_i$ must improve on.
The interaction baseline adds all pairwise products and the minimum of the branch scores. It is the most direct statistical threat, because it captures co-elevation of branches without any role alignment.
The key ablation recomputes $P_i$ with role labels randomly permuted within each branch, 1,000 times. This preserves the gene sets and the centering but destroys the correspondence between roles.
Beating the baselines means a gain in Harrell's C-index on ICGC-LIRI-JP over the GSVA-plus-interaction model, with a bootstrap 95% interval excluding zero. Harrell's C-index is the probability that the model ranks a random pair of patients correctly by survival time. In addition, the true-role score must exceed the 95th percentile of the shuffled-role distribution. Matching the interaction baseline is reported as a negative result for the alignment hypothesis, not as a success.
Why not just…?
Why not just add interaction terms between branch scores? This is the real threat, because "all branches up together" is exactly what products and minima of branch means encode. Our score is built to be orthogonal to it: centering removes level, so $P_i$ measures only internal profile alignment. If the interaction model matches the full model on the external cohort, and the role-shuffle ablation shows no gap, we conclude that level co-elevation is all there is and the parsing-derived structure adds nothing. The experiment is designed to make that conclusion easy to reach.
Why not just use co-expression modules or group penalties, as Barace et al. did? Modules and grouped LASSO capture which genes move together across patients. They answer "what are the branches," not "does this patient's rewiring line up step-for-step." A group-penalized Cox model on the Barace modules is included inside the GSVA and gene-level baselines. If within-patient alignment were redundant with across-patient co-expression, $P_i$ would add no C-index beyond those models.
Why not just let a neural survival network learn the interactions? In the TCGA Meta-Dataset benchmark, an MLP averaged 71.19% accuracy across 174 clinical tasks against 68.78% for logistic regression and 66.18% for always predicting the majority class. The standard deviations were around 15%, so the gap over trivial baselines is small at roughly 150 samples per task. Ficler and Goldberg's parsing results point the same direction: generic models did not pick up symmetry without a dedicated component. We will still train a small multilayer survival network on the branch genes under the same budget. If it matches $P_i$, the structured score's remaining advantage is interpretability only, and we will say so.
Decisive Checks
The proposal stands or falls on two numbers in ICGC-LIRI-JP. The first is the C-index increment of $P_i$ over the branch-GSVA-plus-interaction model. The second is the position of the true-role score in the role-shuffled null distribution. A positive pan-cancer pattern would strengthen the claim. That pattern would show alignment-demanding branch points, in Kehler's sense, coinciding with branch points that share a single upstream transporter or regulator, together with the DepMap hub-dependence result. A null result at the acetyl-CoA pivot, where the biology is best established, would be strong evidence that branch level, not branch symmetry, is what prognosis tracks.
References
- Application of Graph Models to the Identification of Transcriptomic Oncometabolic Pathways in Human Hepatocellular Carcinoma
- Parsing Gapping Constructions Based on Grammatical and Semantic Roles
- Coordinate Noun Phrase Disambiguation in a Generative Parsing Model
- A Neural Network for Coordination Boundary Prediction
- Common Topics and Coherent Situations: Interpreting Ellipsis in the Context of Discourse Inference
- Glycolytic regulatory enzyme PFKFB3 as a prognostic and tumor microenvironment biomarker in human cancers
- Identification of a cholesterol metabolism-related prognostic signature for multiple myeloma
- Comprehensive analysis of mitochondrial-related gene signature for prognosis, tumor immune microenvironment evaluation, and candidate drug development in colon cancer
- Comparison of pathway and gene-level models for cancer prognosis prediction
- The TCGA Meta-Dataset Clinical Benchmark