Many cancer studies now claim that a handful of long non-coding RNAs (lncRNAs), which are RNA transcripts that regulate other genes without making proteins, predict survival "independently of stage." These claims are usually tested by adding the RNA risk score to a regression model that already contains a few coarse table columns: age, TNM stage, and grade. The written pathology report is left out. That report often records the features that separate same-stage patients, such as microvascular invasion, margin status, satellite nodules, and which lesion showed which feature. This proposal asks how much of a published lncRNA signature's prognostic value is left once the clinical baseline also knows what the pathologist wrote.
Answering that requires reading reports at the level of individual lesions, and that is where anaphora resolution comes in. Anaphora resolution is the task of working out what a phrase like "it" or "the larger nodule" refers to. In a multifocal liver tumor report, "the second nodule measures 2.1 cm; it shows microvascular invasion" has to be attached to the right lesion. Otherwise a patient-level feature such as "invasion in the dominant lesion" cannot be computed. The core claim is falsifiable. For signatures derived on The Cancer Genome Atlas (TCGA), a large public collection of tumor sequencing and clinical data, much of the reported improvement over stage is absorbed by lesion-resolved report features, once both are evaluated without in-sample optimism. We also test whether explicit reference resolution matters for building those features, or whether simpler text extraction is enough.
Core Claim and Why It Is Different
The lncRNA literature in the file supplies the targets of this audit. An eight-lncRNA HCC signature built on TCGA liver cancer data separates median survival of 84.4 versus 23.1 months. The same paper reports that its risk score correlates with vascular invasion (p = 0.010) and grade (p = 0.001). That correlation is exactly why a richer pathology baseline could absorb the signal. A 14-lncRNA m6A-related gastric signature reports a 5-year AUC of 0.873 on TCGA stomach cancer data. A 6-lncRNA IL2RG-related kidney signature reports 1-, 3-, and 5-year AUCs above 0.70 and claims superiority over TNM. All three set their covariates as tabular fields, and all three were derived on the same TCGA cohorts they report on. The breast cancer 5-lncRNA nomogram shows the usual pattern: AUC of 0.71–0.72 alone and 0.81 combined with clinical metrics. Its single-institution qRT-PCR data are not available to us, so it motivates the study but is not audited.
On the text side, no prior work links coreference to molecular prognosis. BioCoref benchmarks language models on biomedical literature. It finds that pronouns are nearly solved (F1 up to 0.975) and that long-distance definite noun phrases are the main failure. "The larger nodule" and "the previously described mass" are phrases of exactly that kind. Two-layer phenotype fingerprinting of Korean autism reports and ClinicalMamba mine clinical narratives for heterogeneity or outcome prediction, but neither tracks referents explicitly and neither touches molecular data. The obvious combination would be to feed coreference-resolved notes and lncRNA scores into one multimodal nomogram and report a higher AUC. That is not what we propose, because a higher combined AUC says nothing about whether the RNA assay is worth running. Our object is the incremental value of the molecular signature over a stronger text-derived baseline. Reference resolution is included only as the ingredient that makes lesion-level features possible, and it is tested by ablation.
Research Program
The first component is a lesion-attribution corpus and extractor. We will assemble TCGA's archived surgical pathology reports for the liver (LIHC), stomach (STAD), and kidney (KIRC) cohorts. These are scanned documents and will need OCR cleanup. For about 150 reports, oversampling multifocal cases, two annotators will mark lesion mentions, their coreference chains, and a fixed set of prognostic attributes attached to each chain. The attributes are size, differentiation, vascular or perineural invasion, margin, capsule, satellite lesions, and nodal findings.
The extractor has three steps. The first filters non-referential "it", using distributional pleonastic detection or its deterministic cousins. The second links lesion mentions into chains with an open-weight LLM prompted locally, following BioCoref's finding that local-only prompting of an 8B model outperformed larger models and broader context on CRAFT. We add a string-similarity feature for nominal matches, the device Strube et al. showed roughly doubled definite noun phrase F-measure. The third step attaches attributes to chains. From the resulting lesion table we compute patient-level features, such as the worst attribute across lesions, the attributes of the dominant lesion, and lesion count.
The second component is the audit itself. The published signatures were selected on TCGA, so taking their coefficients and scoring TCGA patients is in-sample. We therefore re-run each paper's selection recipe, univariate screening followed by LASSO Cox, inside nested cross-validation with the same candidate lncRNA pools. Every comparison is made on held-out folds. Following Varoquaux et al., who found linear models matched nonlinear ones on clinical items, the primary models are Cox regressions. A gradient-boosted survival model serves as a check that we are not under-fitting the text features.
The third component addresses heterogeneity directly. Within patients sharing the same TNM stage, we ask whether the lesion-resolved report profile or the lncRNA score explains more of the survival variance. The field's argument, from Lee et al. and Holst et al., is that coarse labels hide subgroups. Our question is whether molecular data are needed to find those subgroups or whether the pathologist already wrote them down.
Baselines
The naive reference point is the covariate set the lncRNA papers themselves use: age, sex, TNM stage, and grade from TCGA's structured clinical tables, fit by Cox regression on the same folds. Every other model has to beat this. This is the baseline the published signatures claim to beat.
The strongest existing method is each published signature added to that tabular baseline. For the gastric cohort the reported figure is a 5-year AUC of 0.873 (Huang et al.). For liver it is a median survival split of 84.4 versus 23.1 months, with no C-index reported on Lacuna (Li, Zeng et al.). For kidney it is AUCs "often > 0.70" (Hu et al.). We run it two ways, with the frozen published coefficients and with the nested-CV re-derivation. The gap between the two measures optimism, and only the re-derived version is used to test the claim.
The text baseline without reference resolution uses the same LLM to extract the same attributes from the whole report in one pass, producing patient-level fields directly. It gets the same model, prompt budget, and annotated development set as the full extractor. We also include TF-IDF on the report text fed to a penalized Cox model. The key ablation isolates reference resolution: full lesion-chain extraction versus the one-pass extraction, scored on the multifocal subset. There, the one-pass method has to guess which lesion a feature belongs to.
For the coreference step itself, the reference number is BioCoref's LLaMA-8B local-only F1 of 0.791 on CRAFT, which is literature rather than clinical text. No reported result was found on Lacuna for coreference on pathology reports. We will report chain F1 using both MUC and B-CUBED, because the CAMP system paper showed the two metrics can disagree sharply when chains are merged by mistake.
Beating means two things. For the audit, the retained increment of the signature, defined as the ΔC-index of the signature over the text-augmented baseline divided by its ΔC-index over the tabular baseline, falls below 0.5. The likelihood-ratio test for the signature also loses significance in at least two of three cancers. For the reference-resolution ingredient, lesion-chain features must improve held-out C-index or attribute accuracy over one-pass extraction on the multifocal subset, with a bootstrap confidence interval excluding zero.
Why not just…?
Why not just use TCGA's structured clinical fields for vascular invasion and margins? Some such fields exist, notably for liver cancer, and they will be added to the tabular baseline. If they turn out to be complete enough to absorb the signature on their own, the audit conclusion still holds and the text pipeline becomes unnecessary for that cancer. We will report that plainly. We expect missingness and lesion-agnostic coding to limit them, but this is an empirical question and we treat it as one.
Why not just prompt an LLM to extract fields from the whole report, without chains? This is the most serious threat. For single-lesion reports it will almost certainly match our pipeline. BioCoref suggests pronouns are easy for LLMs and the hard cases are distant definite descriptions. For that reason we claim a benefit only on multifocal reports, and the ablation above is where it lives or dies. If one-pass extraction ties there too, we drop reference resolution from the recommended pipeline. The audit result stands either way.
Why not just re-derive the signatures with proper nested cross-validation and skip the text? That alone would measure optimism. It would not tell us whether the surviving signal is molecular or a proxy for histology the pathologist already recorded, which is what the correlation with vascular invasion in Li, Zeng et al. raises. We run it as a component of the audit, not a replacement.
Why not just use pairwise relative-expression lncRNA scores? Song et al. use them to sidestep batch effects. That addresses portability, not incremental value over pathology. We include a pairwise variant as a secondary molecular arm.
Decisive Checks
The primary outcome is the retained increment per cancer type, computed on held-out folds with 1,000 bootstrap resamples. Secondary outcomes are time-dependent AUC at 1, 3, and 5 years, net reclassification within same-stage strata, and the partial correlation between each signature's score and the lesion-resolved features.
A negative control is a random lncRNA set of matched size put through the same selection recipe. It shows how much apparent increment the procedure produces from noise. The kidney signature's drug-sensitivity claims rest on in-silico prediction and are out of scope.
The main limitation is external validation. GEO expression cohorts lack reports, so conclusions are confined to TCGA. We will state that the audit bounds, rather than settles, each signature's clinical value.
References
- Identification of an Eight-lncRNA Signature as the Prognostic lncRNA Markers in Hepatocellular Carcinoma Patients
- Identification of N6-Methylandenosine-Related lncRNAs for Subtype Identification and Risk Stratification in Gastric Adenocarcinoma
- Prognostic and therapeutic relevance of IL2RG-related LncRNAs in clear cell renal cell carcinoma
- Prognostic value of a lncRNA signature in early-stage invasive breast cancer patients
- BioCoref: Benchmarking Biomedical Coreference Resolution with LLMs
- Mode-Aware Phenotype Profiling from Korean Clinical Reports: An LLM-Derived Two-Layer Fingerprint for Autism Characterization
- ClinicalMamba: A Generative Clinical Language Model on Longitudinal Clinical Notes
- Distributional Identification of Non-Referential Pronouns
- The Influence of Minimum Edit Distance on Reference Resolution
- Patterns of schizophrenia symptoms: hidden structure in the PANSS questionnaire
- Peripheral inflammatory biomarkers define biotypes of bipolar depression
- Clustering identifies endotypes of traumatic brain injury in an intensive care cohort: a CENTER-TBI study
- Description of the UPENN CAMP System as Used for Coreference
- Identification of an Immune-Related Long Noncoding RNA Pairs Model to Predict Survival and Immune Features in Gastric Cancer