MicroRNAs are short, non-coding RNA molecules that act as master regulators of gene expression, fine-tuning how much protein is produced from thousands of genes at once. Because they circulate stably in blood and other bodily fluids, they have become prime candidates for liquid biopsy diagnostics, a far less invasive and less expensive alternative to brain imaging or lumbar puncture. Altered blood microRNA profiles have been linked to Alzheimer’s disease, Parkinson’s disease, and many cancers. Yet building reliable classifiers from microRNA data is notoriously difficult: datasets typically contain hundreds of microRNAs measured on only a few dozen patients, a recipe for overfitting, and the small cohort sizes mean that any single study captures only a fragment of the true biological picture.

GraphiRNA, developed by Veronica Buttaro, Antonio Pellicani, Gianvito Pio, Domenica D’Elia, Cristina Pizzulli, Michelangelo Ceci and colleagues at the University of Bari and partner institutions, tackles the problem by representing each patient not as a raw vector of expression values but as a point in a learned embedding space that encodes the biology of the microRNAs themselves. The first step is to build a correlation network: for every pair of microRNAs, the method computes the Pearson correlation of their expression values across the training patients, then keeps only the correlations that pass a statistical significance test. Rather than using an arbitrary user-defined threshold on correlation strength, the researchers rely on the standard p-value machinery of hypothesis testing, with an option to apply the Benjamini-Hochberg correction for multiple comparisons.

Each node in this network is then enriched with information about the microRNA’s actual nucleotide sequence, using RNABERT, a pre-trained transformer model for RNA. RNABERT was trained on two tasks: masked language modeling, which teaches it to predict hidden nucleotides from context, and structural alignment learning, which encodes information about RNA secondary structure drawn from the Rfam database. The result is a 120-dimensional feature vector for every microRNA that captures both sequence patterns and structural characteristics, properties that determine binding affinity and target specificity and therefore the molecule’s regulatory role.

The most inventive element of the method is how it handles negative correlations. Most existing approaches either delete links with negative correlation values or replace them with their absolute values, both of which destroy information. A strong negative correlation, after all, is a meaningful biological relationship: when one microRNA rises, the other reliably falls. GraphiRNA instead introduces synthetic mirror nodes into the network. For every negatively correlated pair, two new nodes are created whose feature vectors are point-reflections of the originals across the center of the embedding space. A strong negative correlation between microRNAs i and j can then be faithfully represented as a strong positive correlation between i and the mirror of j, and between j and the mirror of i. The sign of the relationship is preserved without breaking the algorithms that expect non-negative edge weights.

With the network built, the method applies GraphConv, an inductive graph neural network, to learn embeddings for every node. Trained in a self-supervised link-prediction setting, the network learns to pull connected microRNAs together in the embedding space, with the weighted mean aggregator ensuring that stronger correlations exert more influence. Two convolutional layers capture both direct and two-hop neighborhood relationships. The synthetic mirror nodes are then discarded, but their influence has already been absorbed into the embeddings of the real microRNAs through neighborhood aggregation. Finally, each patient’s normalized expression profile is used as a set of weights over the microRNA embeddings, so that a patient becomes a weighted combination of the biological context of every microRNA they express. Patients with similar disease states land close together in this space, and a Random Forest classifier with 100 trees reads out the diagnosis.

The evaluation drew on three real datasets from the NCBI Gene Expression Omnibus, spanning 300 healthy controls, 115 patients with mild cognitive impairment, and 841 Alzheimer’s patients aged 60 to 85, evaluated with stratified five-fold cross-validation. GraphiRNA outperformed a range of competitors, including multi-view methods based on raw expression data, stacked autoencoders, graph attention networks, and graph transformer networks. The best competitor was the autoencoder-based approach of Venugopalan and colleagues, but a Wilcoxon signed-rank test confirmed that GraphiRNA’s advantage was statistically significant, winning in five out of five cross-validation folds with a p-value of 0.03125.

The clinically striking result concerns mild cognitive impairment, or MCI, the transitional state that often precedes frank Alzheimer’s disease and is the hardest class for algorithms to recognize. Competing deep learning methods achieved high precision and accuracy on the abundant Alzheimer’s and control classes but nearly failed to detect MCI at all. GraphiRNA, by contrast, recognized the MCI class with recall between roughly 54 and 56 percent and, in most configurations, achieved perfect precision, meaning that when it flagged a patient as MCI it was almost always right. Because early intervention at the MCI stage offers the best window for slowing progression, the authors argue this balance is precisely what a screening tool would need. The method also proved unusually robust in a transfer learning test, where models trained on two studies were applied to a third collected on a different experimental platform: while most competitors collapsed into predicting a single class, GraphiRNA maintained balanced performance across all three.

An ablation study quantified the value of the synthetic-node trick. Compared with a variant that simply took the absolute value of negative correlations, the full method improved the F1-score by up to 30.9 percent, confirming that the sign of a correlation carries diagnostic information worth keeping. A biological validation added further credibility: the fifty most central microRNAs in the learned network, analyzed through experimentally validated gene targets and enrichment analysis, converged on pathways central to Alzheimer’s biology, including neurotrophin signaling, the Alzheimer’s disease pathway itself, BDNF signaling for synaptic plasticity, DYRK1A regulation, and neuroinflammatory and glutamatergic processes.

The authors caution that work remains before GraphiRNA reaches the clinic. Recall on the healthy control class was still sub-optimal, and the team plans to address class imbalance, improve explainability to isolate candidate biomarker microRNAs, and validate predictions on freshly recruited patient blood samples. They also intend to test the framework on other neurodegenerative diseases and cancers. But the core message stands: by letting an algorithm see the conversation among microRNAs rather than listening to each molecule in isolation, and by refusing to throw away half of the correlation signal, GraphiRNA shows how graph-based thinking can turn a cheap blood test into a far sharper diagnostic instrument.

Subject of Research: Graph-based patient embedding from microRNA expression data for Alzheimer's disease diagnosis

Article Title: A novel graph-based patient embedding method for the diagnosis of Alzheimer’s disease from microRNA expression data

Article References: A novel graph-based patient embedding method for the diagnosis of Alzheimer’s disease from microRNA expression data. (n.d.). https://doi.org/10.1007/s10618-026-01275-y

Image Credits: AI Generated

DOI: 10.1007/s10618-026-01275-y

Keywords: Alzheimer's disease, microRNA, graph neural networks, patient embedding, liquid biopsy, mild cognitive impairment, RNABERT, correlation network, machine learning, biomarkers, bioinformatics, early diagnosis