Research Article  ·  Computers in Biology and Medicine  ·  Published March 3, 2026

Natural Language Processing of Biomedical Text to Map and Prioritize Protein–Disease Associations in HFpEF

Clodomir Santana  ·  Chitra Mukherjee  ·  Arnib Quazi  ·  Ronaldo Menezes  ·  Vladimir Filkov  ·  Dibakar Sigdel  ·  Howard Choi  ·  Imo Ebong  ·  Padmini Sirish  ·  Nicholas R. Anderson  ·  Xuan Wang  ·  Heng Ji  ·  JiaWei Han  ·  Baback Roshanravan  ·  Leighton T. Izu  ·  Thomas W. Smith  ·  Nipavan Chiamvimonvat  ·  Colleen E. Clancy  ·  Martin Cadeiras  ·  David A. Liem

Computers in Biology and Medicine, 205: 111599  ·  DOI: 10.1016/j.compbiomed.2026.111599

Literature Corpus 38M+ PubMed abstracts screened for CVD relevance
CVD Abstracts 5.5M publications curated across 6 cardiovascular disease groups
HFpEF Proteins 5,124 proteins computationally identified and ranked for HFpEF
Biopsy Validation 2.65× enrichment factor over chance in HFpEF patient tissue data

The Problem

Identifying clinical biomarkers and drug targets for cardiovascular disease has traditionally relied on biochemical and omics experiments. These studies have generated millions of publications — but the knowledge they contain is scattered across a fragmented, ever-growing literature that now exceeds 38 million indexed papers on PubMed alone.

No single researcher, lab, or even institution can synthesise that body of work. Proteins of high diagnostic or therapeutic relevance go unnoticed simply because the evidence for them is buried across thousands of papers that have never been read side-by-side.

This problem is especially acute for heart failure with preserved ejection fraction (HFpEF), a subtype of heart failure where the left ventricle contracts normally but stiffens during filling. HFpEF is entangled with aging, obesity, diabetes, hypertension, and hyperlipidemia, has no approved disease-modifying therapy, and remains poorly understood at the molecular level.

"The validation of promising clinical biomarkers, molecular mechanisms, and novel drug targets in cardiovascular disease is hindered by a vast and fragmented biomedical literature, which now exceeds 38 million publications indexed in PubMed."

Previous text-mining efforts in cardiology have focused on electronic health records or narrow bibliometric analyses. No study had previously applied NLP at this scale to systematically map protein–disease associations across all major CVD groups and then focused specifically on HFpEF.

The Approach

The authors developed VITAL (ValIdated Text-mining using Advanced Language models), a computational framework built around the CaseOLAP algorithm (Context Aware Semantic Online Analytical Processing). VITAL scores each of the 20,428 known human proteins against every disease domain of interest, producing a ranked list of protein–disease associations derived entirely from published text.

Key Findings

Applied to the full PubMed corpus, VITAL mapped protein–disease associations across six CVD groups and produced a rich molecular portrait of HFpEF and its biological neighbourhood.

IHD, CM, and CVA share the densest molecular landscapes

Ischemic heart disease, cardiomyopathy, and cerebrovascular accidents each had over 7,000 associated proteins and substantial pairwise overlap — CM and IHD alone share 6,595 proteins — suggesting convergent molecular pathways across these three conditions.

Known biomarkers confirmed at high scores

Established clinical biomarkers — troponin-I (IHD), BNP (cardiomyopathy), ACE and renin (heart failure), TNF and galectin-3 (inflammation) — all received CaseOLAP scores above the statistically derived cut-off for their respective diseases, validating the scoring framework.

4,879 proteins converge across HFpEF and all comorbidities

Of 5,124 proteins found in HFpEF literature, 4,879 are shared across all four comorbidity domains. Aging showed the strongest computational correlation with HFpEF (r = 0.61), suggesting the deepest shared biological pathways of any comorbidity studied.

Eight top-ranked proteins are consistent across all domains

TDO2, SPARC, TCO1, MYPC3, and ASA were the highest-ranked HFpEF proteins and remained top-ranked across comorbidities and pathological mechanisms alike, implying they are central molecular drivers rather than domain-specific signals.

HFpEF proteins localise outside the heart

Cross-referencing with the Human Protein Atlas revealed that the top computationally ranked proteins are primarily expressed in liver, pancreas, and adipose tissue — not cardiac tissue — providing computational evidence that HFpEF is fundamentally a multisystemic disorder driven by inter-organ signalling.

Computational predictions validated in patient biopsies

Seven of the top 20 NLP-ranked HFpEF proteins appeared in the top 100 differentially expressed proteins in myocardial biopsies from HFpEF patients, with an enrichment factor of 2.65× (p = 0.00265) over random expectation. Vinculin, desmin, and SPARC were among the validated candidates.

Why It Matters

VITAL is not a replacement for wet-lab science — it is a navigation system for it. By ranking proteins against disease domains at the scale of the entire biomedical literature, it surfaces candidates that experiments alone would take decades to reach.

Infographics

The VITAL pipeline: from biomedical text to prioritized proteins
VITAL processes the entire PubMed literature in five automated stages. Each stage builds on the previous: raw text is mined for biomedical entities, co-occurrence relations are extracted, a weighted knowledge graph is constructed, and proteins are finally ranked by their association strength with HFpEF.
From noisy associations to a focused HFpEF protein landscape
The raw knowledge graph contains thousands of protein-disease co-occurrences spread across the literature. VITAL filters and ranks this network to surface the proteins most strongly and specifically linked to HFpEF — transforming an unnavigable graph into a tractable priority list for experimental follow-up.
Raw association graph thousands of proteins · noisy
→
HFpEF priority subgraph top-ranked proteins · actionable
Named entity recognition: detecting proteins and diseases in text
The first NLP stage scans each sentence for biomedical entities. Protein mentions (blue) and disease terms (amber) are tagged independently before relation extraction links co-occurring pairs into candidate protein-disease associations.
Protein entity detection e.g. TITIN, MYH7, TNNT2, TTN
Disease entity detection e.g. HFpEF, fibrosis, hypertrophy

Paper & Authors

Authors

Clodomir Santana (co-first) Division of Cardiovascular Disease
University of California, Davis, USA
Chitra Mukherjee (co-first) Division of Cardiovascular Disease & AI for Health Center
University of California, Davis, USA
Arnib Quazi Division of Cardiovascular Disease & Center for Precision Medicine
University of California, Davis, USA
Ronaldo Menezes Department of Computer Science
University of Exeter, UK
Vladimir Filkov AI for Health Center
University of California, Davis, USA
Dibakar Sigdel Division of Cardiovascular Disease
University of California, Davis, USA
Howard Choi Division of Cardiovascular Disease
University of California, Davis, USA
Imo Ebong Division of Cardiovascular Disease
University of California, Davis, USA
Padmini Sirish Division of Cardiovascular Disease
University of California, Davis, USA
Nicholas R. Anderson Department of Medical Informatics
University of California, Davis, USA
Xuan Wang Department of Computer Science
Virginia Tech, Blacksburg, USA
Heng Ji Department of Computer Science
University of Illinois, Urbana-Champaign, USA
JiaWei Han Department of Computer Science
University of Illinois, Urbana-Champaign, USA
Baback Roshanravan Department of Nephrology
University of California, Davis, USA
Leighton T. Izu Department of Pharmacology
University of California, Davis, USA
Thomas W. Smith Division of Cardiovascular Disease
University of California, Davis, USA
Nipavan Chiamvimonvat University of California, Davis & University of Arizona
College of Medicine, Phoenix, USA
Colleen E. Clancy Center for Precision Medicine & Dept. of Pharmacology
University of California, Davis, USA
Martin Cadeiras Division of Cardiovascular Disease
University of California, Davis, USA
David A. Liem (corresponding) Division of Cardiovascular Disease & AI for Health Center
University of California, Davis, USA

Cite This Work

@article{santana2026vital, title = {Natural language processing of biomedical text to map and prioritize protein--disease associations in {HFpEF}}, author = {Santana, Clodomir and Mukherjee, Chitra and Quazi, Arnib and Menezes, Ronaldo and Filkov, Vladimir and Sigdel, Dibakar and Choi, Howard and Ebong, Imo and Sirish, Padmini and Anderson, Nicholas R. and Wang, Xuan and Ji, Heng and Han, JiaWei and Roshanravan, Baback and Izu, Leighton T. and Smith, Thomas W. and Chiamvimonvat, Nipavan and Clancy, Colleen E. and Cadeiras, Martin and Liem, David A.}, journal = {Computers in Biology and Medicine}, volume = {205}, pages = {111599}, year = {2026}, doi = {10.1016/j.compbiomed.2026.111599} }

Open access · CC BY 4.0