Literature maps are now a standard way to summarize a biomedical field. Tools such as VOSviewer and CiteSpace draw a field as a network: terms are nodes, and an edge links two terms that often appear in the same abstracts. Thick edges are then read as "established mechanisms" and future directions. Histopathology, the microscopic study of stained tissue, adds a specific weakness to these maps. Tissue can only be examined after the animal is sacrificed or the biopsy is taken, so a typical mechanism paper measures a molecular marker and a tissue score in the same animal at a single endpoint, correlates them, and reports a mechanism. That is a cross-sectional design. It cannot tell whether the marker drives tissue damage, the damage drives the marker, or the treatment group drives both. A co-occurrence map records all of these papers as the same thick edge.
We propose to rebuild the map of one histopathology-heavy field so that each marker-to-tissue edge carries a record of how it was supported. Was the marker itself manipulated? Did the marker change before the tissue across animals sacrificed at different times? Or were the two only correlated at a terminal endpoint? We then run a forward-in-time test. We freeze the literature at a cutoff year, take mechanistic edges that at that point had no direct experiment manipulating the marker, and ask which edge-scoring method best predicts whether the first later knockout, antibody-blockade, or overexpression study confirms the edge. If evidence-design features predict that outcome better than co-occurrence strength, citation bursts, or a language model's prior, then we have a cheap and checkable way to tell real mechanisms from artifacts of the cross-sectional design.
Core Claim
In tissue-remodeling research such as renal and hepatic fibrosis or steatosis, a large share of the high-weight edges in standard literature maps rest only on terminal-endpoint correlations. Often the correlation is pooled across treatment groups. The negative correlation between Angiopoietin-1 and the fibrosis index reported in the ureteral-obstruction rat study of Li et al. is a representative case: the drug was randomized, but the marker–histology link was not. Our claim is narrow and falsifiable. Group-level temporal ordering, available from serial-sacrifice time-course experiments, plus whether a correlation was computed within or across treatment arms, carries information about whether a mechanism will survive direct manipulation. Popularity measures do not carry that information. The idea borrows the lead-lag logic of longitudinal cohorts, which Clark et al. use to escape snapshot designs. It applies that logic to animals that can each be observed only once, much as SpaTIE orders cross-sectional biopsy patches along a pseudo-temporal axis. The difference is that our ordering comes from the experimental timeline, not from image embeddings.
Relation to Prior Work
Three groups of prior work bear on this proposal, and each plays a different role.
The first group motivates it. Bibliometric studies such as the pulmonary-fibrosis map of Wu et al., the ferroptosis decade review of Ji et al., and the Food-Energy-Water map of Opejin et al. read clusters and burst keywords as research fronts and gaps. None of them distinguishes an edge that has been tested by intervention from one that has merely been mentioned together. The research file names this limitation explicitly: co-occurrence lacks semantic polarity and causal direction. We keep their pipeline as a baseline and add evidence-design labels to it.
The second group supplies methods and sets constraints. The breadth-first language-model approach of Efficient Causal Graph Discovery Using Large Language Models orients causal graphs from textual knowledge. Its own limitations section notes that it may reproduce publication bias and that it assumes no hidden common causes. Our test is designed to measure exactly that failure. CDHS handles hidden common causes in tabular data but needs thousands of samples. Individual papers in this field report tens of animals, so we do not attempt numerical causal discovery on pooled raw data.
The third group shows that the problem is general. CaMIL and CATTMIL remove staining confounders inside image classifiers, and SpaTIE's "virtual perturbation" yields mutation–morphology couplings that its authors concede are associational. All three operate inside a single model. None asks which claims across a whole literature are backed by intervention, and that is our question.
Research Program
Corpus. We assemble a PubMed and Web of Science corpus on renal interstitial fibrosis, liver fibrosis, and diet-induced steatosis. These fields are dominated by Masson's trichrome, Oil Red O, and H&E endpoints in rodent models, as in Chou et al. and the methionine-diet vascular study art_dd15. We draw on full text where licenses allow.
Claim extraction. From each paper we extract claims of the form "marker M is linked to histological outcome H with sign s". A language model performs the extraction, and it is checked against a gold set of roughly 600 papers double-annotated by a pathologist and an epidemiologist.
Evidence coding. Each claim is assigned one evidence design:
- direct manipulation of M with histology read out;
- a serial-sacrifice time course from which we record whether M's change precedes, coincides with, or follows H's change at the group level;
- a terminal-only correlation, further coded as within-arm or pooled across arms;
- a human cross-sectional biopsy correlation.
We also record the animal models, species, and labs involved.
Edge features and map. Edges aggregate these labels into features: the fraction of within-arm support, a lead-lag consistency score, independent replication across distinct injury models, and sign agreement. We render the result as an overlay on a standard VOSviewer-style layout, with edges colored by their strongest evidence type. This makes it visible which "hotspots" are correlation-only. The scoring model itself is deliberately simple, a regularized logistic regression on interpretable features, so that any gain comes from the evidence coding rather than from model capacity.
Baselines
Raw co-occurrence link strength. This is the naive reference point: the link strength VOSviewer would compute for the edge from the frozen pre-cutoff corpus, used directly as a confirmation score. It is what current bibliometric practice implicitly treats as mechanism strength.
Citation-based attention. We combine Kleinberg burst strength from CiteSpace with normalized citation counts of the supporting papers. This tests whether field attention already tracks validity. It sees the same pre-cutoff corpus.
Language-model causal prior. This is the strongest existing method. We run the breadth-first procedure of art_3d8cf572 (Efficient Causal Graph Discovery Using Large Language Models) over the marker and outcome vocabulary and score edges by model confidence. To keep the comparison fair, we use a model whose training data predate the cutoff, give it the same vocabulary, and allow the same number of prompt-tuning trials we spend on our own features. The cited paper reports an F-score of 0.93 on the 8-node Asia benchmark and 0.351 on the 221-node Neuropathic Pain benchmark. No reported result was found on Lacuna for predicting future experimental confirmation of literature edges. Beating this baseline means a higher AUROC with a bootstrap 95% interval that excludes the difference of zero.
Evidence-count ablation. This ablation isolates our key ingredient. It uses the same regression with the count of supporting papers and models but drops all design labels: lead-lag, within- versus pooled-arm, and manipulation type. If the full model does not beat this ablation, the gain comes from volume, not design.
Why not just…?
Why not just ask a language model whether each edge is causal? It is the most serious competitor, and it may perform well on popular edges. However, it learns from the same abstracts that inflate correlation-only edges, and its own authors flag publication bias and the assumption of no hidden common causes as weaknesses. The head-to-head comparison settles the question. We also report performance separately on the edges where the language model is most confident and our design features say "pooled-arm correlation only". That stratum is where the two methods should disagree.
Why not just count supporting papers or citations? If popularity alone predicted confirmation, our design coding would be wasted effort. The evidence-count ablation and the citation baseline test exactly this. We would accept a null result as a finding: it would mean that, in this field, the volume of evidence already carries the design information.
Why not just apply standard systematic-review evidence grading? Grading tools rate risk of bias for the treatment effect in a randomized experiment. They do not ask whether the marker–tissue link inside that experiment is identified, and this is the specific gap that terminal sacrifice creates. Our within-arm and lead-lag codes go below the level these tools assess. Where the two overlap, we include a coarse grade as an extra baseline feature.
Proposed Experiments
Main forward-in-time test. We freeze the corpus at 2018. We take every marker-to-histology edge that has no direct-manipulation evidence before the cutoff and at least one such study published between 2019 and 2025. The label is whether the first such study reports a histological change in the predicted direction; null and opposite-direction results count as not confirmed. We report AUROC, Brier score, and calibration for every baseline. The decisive check is the comparison of the full model against the evidence-count ablation.
Secondary analyses. First, we measure the share of the top-weighted co-occurrence edges and burst clusters that rest only on pooled-arm terminal correlations. This tells us how much of a standard map is directionally unidentified. Second, we test whether lead-lag ordering in serial-sacrifice studies predicts the direction of effect better than chance. This is a group-level analogue of the reverse-causality problem that the NHANES snapshot of Sun et al. cannot resolve.
Robustness checks. We vary the cutoff year between 2016 and 2020. We repeat the main test in hepatic steatosis alone to check transfer across fields. We also bound the effect of publication bias: null manipulation results are underreported, which would inflate confirmation rates for all methods. Because the bias affects every method, relative rankings should be robust, and we test this by reweighting toward registered and preprinted studies where they exist.
Failure condition. If, after a few hundred edges, design features add no discrimination over evidence counts, we will report that terminal-correlation evidence is no worse than its volume suggests. That would weaken the critique of cross-sectional mechanism claims in this literature.
References
- All-trans retinoic acid regulating angiopoietins-1 and alleviating extracellular matrix accumulation in interstitial fibrosis rats
- Adolescent Executive Dysfunction in Daily Life: Relationships to Risks, Brain Structure and Substance Use
- Learning latent progression states from spatial heterogeneity in uterine histopathology
- Pulmonary fibrosis complicated by lung cancer: bibliometric analysis from 2004 to 2024 - research status, trends and future directions
- Trends and hotspots of publications on ferroptosis: A 10 Year overview
- A Bibliometric Analysis of Food-Energy-Water Nexus Literature
- Efficient Causal Graph Discovery Using Large Language Models
- Recovery of Causal Graph Involving Latent Variables via Homologous Surrogates
- CaMIL: Causal Multiple Instance Learning for Whole Slide Image Classification
- Causal ATTention Multiple Instance Learning for Whole Slide Image Classification
- Preventive Effects of <i>Ophiocordyceps </i><i>s</i><i>inensis </i>Mycelium on High-Fat Diet Induced Lipid Dysregulation and Hepatic Inflammation of Mice
- Vascular Remodeling, Oxidative Stress, and Disrupted PPARγExpression in Rats of Long-Term Hyperhomocysteinemia with Metabolic Disturbance
- Exploring the relationship between dietary index for gut microbiota and cognitive function