Natural language processing of biomedical text to map and prioritize protein–disease associations in HFpEF

The validation of promising clinical biomarkers, molecular mechanisms, and novel drug targets in cardiovascular disease (CVD) is hindered by a vast and fragmented biomedical literature, which now exceeds 38 million publications indexed in PubMed. To address the central challenge of navigating and synthesizing a huge fragmented biomedical literature base, we applied our validated machine learning–based text-mining algorithm containing natural language processing (NLP) and incorporated this into a V al I dated T ext-mining using A dvanced L anguage model ( VITAL ) as a complementary framework. Using this approach, we analyzed more than 38 million PubMed abstracts and identified over 5.5 million relevant to six major CVD groups. These curated data then enabled a deep-dive case study on heart failure with preserved ejection fraction (HFpEF). Our computational framework systematically queried, quantified, mapped, and prioritized protein–disease associations, confirming established CVD biomarkers, such as BNP, troponin-I, galectin-3, and renin, and revealing novel protein signatures with potential diagnostic and therapeutic relevance. Ischemic heart disease (IHD, heart attacks), cardiomyopathy (CM, leading to heart failure), and cerebrovascular accidents (CVA, strokes and brain hemorrhages) exhibited the highest protein attribution densities and overlap, suggesting shared molecular pathways. Using HFpEF as a focused case study, our framework identified 5124 proteins associated with this condition, 4879 of which were shared across its major comorbidities (aging, type 2 diabetes/obesity, hypertension, and hyperlipidemia). Additionally, 4991 proteins were co-shared across key pathological mechanisms, including inflammation, mitochondrial dysfunction, and fibrosis, implicating convergent biological networks spanning these domains. To further characterize and prioritize these molecular associations, we performed a series of data science-driven analyses involving HFpEF-associated proteins. The top computationally ranked HFpEF protein candidates were the same top ranked proteins in the comorbidity-domains and in the pathology-domains suggesting that these proteins are important drivers with convergent molecular networks underlying HFpEF. Cross-referencing and validating top-ranked computational HFpEF protein candidates with clinical myocardial and extracardiac biopsy data from HFpEF patients and corresponding controls revealed that most of these proteins are predominantly expressed in the liver, pancreas, adipose tissue, and lymph nodes, rather than in cardiac tissue. This finding supports the emerging concept that HFpEF is fundamentally a multisystemic disorder mediated by inter-organ signaling rather than a disease confined to the heart. Our computational study demonstrates the capacity of text mining to annotate, integrate, and prioritize protein-disease relationships from large-scale textual data, thereby providing a complementary framework to traditional omics approaches for biomarker discovery and drug target identification in CVDs. The clinical validation of candidate biomarkers, molecular mechanisms and therapeutic targets in cardiovascular disease is constraint by the scale and fragmentation of biomedical literature with over 38 million publications on PubMed. To address this challenge, we applied a validated machine learning based text mining framework incorporating natural language processing (NLP) to systematically navigate this body of knowledge. We analyzed all human proteins listed on UniProt (20,428) across six major cardiovascular disease categories (cerebrovascular accident, CVA; cardiomyopathies and heart failure, CM; arrhythmias, ARR; ischemic heart disease, IHD; congenital heart disease, CHD; valve disease, VD) followed by a focused case study in heart failure with preserved ejection fraction (HFpEF). Our method enabled the mapping and prioritization of protein-disease associations across key comorbidities (aging, diabetes/obesity, hypertension, hyperlipidemia) and central pathological mechanisms (mitochondrial function, inflammation, fibrosis) relevant to HFpEF. Our computational results were validated and cross-referenced with real-world patient data supporting the emerging view of HFpEF as a multisystem disorder driven by inter-organ signaling rather than a purely cardiac condition. • The validation of cardiovascular disease (CVD) biomarkers, molecular mechanisms, and therapeutic targets is impeded by the vast scale and fragmentation of biomedical literature. • We applied machine learning–based text-mining on 38 million PubMed abstracts with a focus on heart failure with preserved ejection fraction (HFpEF), comorbidities, and pathologies to quantify protein-disease associations. • The top computationally ranked HFpEF protein candidates overlapped substantially with their top ranked proteins in comorbidities and pathologies. • Cross-validation of the top ranked HFpEF protein candidates were preferentially expressed in the liver, pancreas, adipose tissue, and lymph nodes rather than in cardiac tissue. • Our computational findings support the emerging paradigm that HFpEF may represent a multi-organ systemic disorder rather than a solely cardiac disease.