Abstract
In this paper, a computational model is proposed for exploring metabolite-disease pairs and has good performance in predicting potential metabolites related to diseases through adequate validation. This result suggests that integrating literature and functional associations can be an effective way to construct disease associated metabolite network for prioritizing candidate diseases-related metabolites.
Introduction
Metabolites serve as the "terminal products" of cellular processes, representing the final output of complex interactions between genes, proteins, and the environment. Because they reflect the actual physiological state of an organism, they are premier candidates for clinical biomarkers. However, a significant "metabolic gap" exists: while thousands of metabolites have been identified, their specific disease associations remain largely unknown.
Traditional experimental methods for mapping these associations are prohibitively slow and expensive. Computationally, the field faces a "data sparsity" challenge, where many metabolites lack sufficient functional data for accurate role prediction. To address this, researchers developed FLDMN (Fused Literature and Disease-associated Metabolite Network), a framework that prioritizes disease-related metabolites by leveraging indirect relationships—the principle that similar diseases likely involve similar metabolites.
Methodology: The FLDMN Pipeline
The FLDMN model employs a multi-step pipeline to integrate diverse data types and rank candidate metabolites.
- Mapping and Similarity Calculation: The process begins by integrating data from the Human Disease Ontology (DO) and the Human Metabolome Database (HMDB). Disease similarity is calculated using FNSemSim, which assesses the overlap in shared gene functional networks between two pathologies.
- Collaborative Filtering (CF): Borrowing techniques from recommendation engines, the model calculates metabolite similarity. If two metabolites are associated with highly similar sets of diseases, they are treated as functionally related. This helps overcome the "cold start" problem for metabolites with few known associations.
- Literature Fusion: To enhance the network's robustness, text-mining scores from the STITCH database are incorporated. These scores quantify the frequency with which a metabolite and a disease or protein are co-mentioned in scientific literature.
- Random Walk with Restart (RWR): The final fused network is analyzed using the RWR algorithm. Starting from a known disease-related metabolite, the algorithm "walks" through the network to identify and rank other nodes based on their "reachability," effectively prioritizing the most likely candidates for the target disease.
Data Sources
The study utilized several key datasets to build and validate the model:
- HMDB (Human Metabolome Database): Two versions (Dec 2017 and April 2018) were used to perform prospective validation, training the model on older data and testing it against newly discovered associations.
- Disease Ontology (DO) & MEDIC: Provided standardized nomenclature and disease mapping.
- STITCH: Supplied literature-based interaction scores between chemicals and proteins.
Results and Discussion
The FLDMN model demonstrated significant predictive power, achieving an average AUC of 76.03% across 19 different diseases.
The Synergy of Data Fusion
A critical insight from the study was the "synergy effect." When tested individually, functional similarity (DMN) achieved an AUC of ~64%, and literature mining (ST_SUBNET) achieved ~62%. However, the fused FLDMN model's performance jumped to 76%, indicating that literature and functional data capture complementary biological truths that are more powerful when combined.
Case Studies: Alzheimer’s and Leukemia
The model's practical utility was validated through specific case studies:
- Alzheimer’s Disease: FLDMN ranked S-Adenosylhomocysteine in the top 3% of candidates. Although not in the training set, independent literature confirms it as a known marker. It also prioritized Substance P (top 10%), a neuropeptide linked to late-onset Alzheimer’s.
- Leukemia: The model correctly identified Putrescine and N1-Acetylspermidine within the top 5% of candidates, both of which are well-documented in clinical leukemia studies.
Conclusion and Future Directions
While FLDMN offers a powerful tool for biomarker discovery, it is not without limitations. Text-mining data can be noisy, potentially introducing false-positive links. Furthermore, the model's performance varies with data density; for diseases with very few known associations, such as Celiac disease (AUC ~59%), the "seed" information remains a bottleneck.
Future work will focus on integrating additional biological layers, such as protein-protein interaction networks and gene expression data, to further refine the prioritization of candidate metabolites and bridge the metabolic gap.
Summary updated