GTex Visualizer

An open source, cloud-based, web application for the analysis of gene expression data in the GTex database to study changes in gene expression ralated to aging and diseases.

ECG Decision Support System

An open source and offline Decision Support System tool for the analysis, automatic identification of abnormalities and annotation of 12-lead ECG signals. The anomaly detection task is performed using an AutoEncoder model trained to reconstruct 2-seconds lenght healthy ECG windowsof. Trained and tested using both synthetical generated data and real ECG signals from the PTB and PTB-XL datasets. The tool provides a user-friendly interface for clinicians to upload and analyze ECG signals. Once the ECG signal is uploaded, the tool automatically detects and shows them to the clinician for the final annotation. Explanations of the prediction are provided to increase clinician trust so they can perform the final annotation.

SARS CoV 2 Spike protein variants analysis trough Protein Contact Networks (PCNs)

PCN-Miner is our open source and offline tool for the analysis of protein structures based on the concept of PCN. A PCN of a given protein is a graph rappresentation of the protein structure. In our research, we constructed and analyzed the PCNs of the SARS CoV 2 Spike variants. Then, we compared our results (for example: aminoacid/node centralities, network communities) to find differences between variants.

Comparing Kolmogorov-Arnold Network Autoencoders against vanilla ones

In this work we compared the performance of Kolmogorov-Arnold Network Autoencoders (KAE) against vanilla Autoencoders (AE) in the analysis of biomedical data. Both linear and convolutional autoencoders were trained and tested on the same datasets. We used a dataset of stetoscopic heart sounds and compared both models for multiple tasks: reconstruction, denoising and inpainting. We found that convolutional KAEs with PixelShuffle layers outperformed all the other models in all tasks.

Key findings:

• Novel AI architecture: First application of Kolmogorov-Arnold Networks (KANs) to medical signal autoencoders
• Outperformed standard CNNs on cardiac audio analysis across 3 tasks: reconstruction, denoising and inpainting
• Best efficiency: Lowest reconstruction error with smallest parameter count among tested architectures
• Applied to AbnormalHeartbeat dataset with stethoscope audio signals for clinical diagnostics
• Demonstrated clear class separation (normal vs. abnormal) in unsupervised latent space learning

Transthyretin

We proposed a new approach to study the transthyretin protein (TTR), a protein involved in amyloidosis. Given the low availability of experimental structures of TTR in the tetramer conformation, we focused on computational methods to model the protein using AlphaFold 3. Predicted structures were validated against experimental data when available A total of 133 TTR mutations were analyzed using various computational tools.

Methods used:

• Structural analysis of the protein using Protein Contact Networks (PCNs) was used to identify central residues and functional domains.
• We computed TM-scores between each variant and wild-type TTR to assess their structural similarity.
• We used ESM2 embeddings and UMAP to visualize the variants distribution in a 2-dimensional space: benign mutations were found near the wild-type TTR, while pathogenic variants were found in more distant regions.
• We also performed tafamidis ligand optimization and de-novo ligand generation using DiffSBDD: optimization increased binding affinity, generation gives best QED drug-likeness.
• We performed docking and binding affinity prediction of the transthyretin protein with its ligands using AutoDock Vina and DiffDock.
• We performed 100-ns molecular dynamics simulations of the V50M TTR variant to study its stability and the stabilizing effects of each ligand.

Key findings:

• Published in Nature: npj Systems Biology and Applications, September 2025
• Analyzed transthyretin (TTR) mutational landscape for drug design in amyloidosis treatment
• Integrated structural profiling across multiple TTR mutations to optimize ligand binding
• Combined computational modeling with drug optimization for precision medicine approach
• Addresses TTR amyloidosis, a disease affecting heart and nerves with limited treatment options
• Contributes to rational drug design pipeline for rare genetic diseases

E-ABIN

We proposed a new tool to analyze gene expression data from the gene expression omnibus (GEO) database. E-ABIN is an explainable module for anomaly detection in biological networks, incorporates both machine and deep learning techniques to perform anomaly detection in gene expression and dna methylation datasets. The tool was evaluated on several benchmark datasets (celiac disease, parkinson, bladder and colorectal cancer), demonstrating its effectiveness in identifying anomalous patterns and providing insights into the underlying biological processes. An explainability module is available to provide different level of explanations (most influential genes for the prediction of the abnormal class): from patient level to dataset level. Most influential genes found using E-ABIN align with known disease-related genes, demonstrating the tool's potential for identifying novel biomarkers and therapeutic targets.

Key findings:

• First-of-its-kind platform: Only tool combining GUI, ML/DL models, explainability, and DNA methylation analysis in one framework
• Perfect accuracy (100% AUC) on colorectal cancer detection, validated on 146-sample dataset
• Identified 7 cancer-relevant genes in bladder cancer including RUNX2, ITGB3, CARD8 validated by medical literature
• Production-ready GUI: Accessible to non-programmers, eliminating technical barriers for clinical researchers
• Outperformed other tools across multiple benchmarks (0.73 vs 0.59 AUC on celiac disease dataset)
• Published in NAR Genomics and Bioinformatics (Oxford University Press), December 2025

ExDiff

A modular and explainable framework combining network simulation and graph neural networks to study spreading dynamics in complex networks. The framework allows to simulate the spreading of a disease in a network, and then to use graph neural networks to learn the patterns of the spreading. The SIRVD model is used to simulate the spreading of a disease in a network, where nodes represent individuals and edges represent interactions between them. The framework is modular and can be easily extended to include other spreading models (e.g., SIR, SIS).

Real case studies:

• Simulation of west-nile virus spreading in both synthetic and real multispecies populations (human, birds and mosquitos) using simulation parameters estimated from italian bulletins.
• Simulation of COVID-19 spreading on both synthetic and real human populations using early and late stage parameters estimated from italian bulletins.
• Simulation of Monkeypox (Mpox) spreading in multispecies population (human and rodent) using parameters from Nigeria and Congo.
• Simulation of measles spreading in human populations using parameters from historical outbreaks in Italy.

Key findings:

• In Monkeypox, targeted vaccination, togheter with quarantine and rodent culling, is the most effective strategy to control the spreading of the disease. With culling having more effects in the early stages of the epidemic.
• In COVID-19, early vaccination of high-risk individuals (nodes with high centrality) and social distancing are key to control the spreading of the disease. With social distancing having more effects in the early stages of the epidemic.
• In Measles, we used XAI ExDiff predictions to guide vaccination and quarantine. XAI guided vaccination improved the effectiveness of vaccination campaigns, by reducing the infection peak from 33.25% (vaccination guided by node centrality) to 10.50% (XAI guided vaccination).

DCAE-SR

A deep learning model that reconstruct electrocardiogram (ECG) signals at super resolution. The model is trained on a large dataset of ECG signals (PTB-XL dataset) and is able to reconstruct the ECG signal at a resolution of 500Hz, while the original signal is sampled at 50Hz and corrupted by ECG artifacts such as EMG, baseline wander, and noise. The model is based on a denoising convolutional autoencoder (DCAE) architecture and explainability techniques are used to undestand the impact of each sample of the signal for the final super-resoluted signal. Performances are evaluated in terms of MSE, RMSE, PSNR and SSIM. Down-stream tasks such as classification shows that applying super-resolution leads to improved classification performance when compared against the interpolation of the original signal.

Key findings:

• Super-resolution innovation: First denoising autoencoder for ECG super-resolution, enhancing low-quality signals for better diagnosis
• 12 citations since 2024 publication in Artificial Intelligence in Medicine (Elsevier, high-impact journal)
• Dual decoder architecture: simultaneously reconstructs clean signals AND generates denoised high-resolution versions from noisy data
• Validated on PTB-XL dataset with superior performance over existing ECG super-resolution methods
• Clinical applications: Arrhythmia detection, microvolt T-wave alternans analysis, cardiac event identification
• Open-source on GitHub with pretrained models for immediate deployment in medical monitoring systems
• Using Explainability techniques to interpret how model performs super-resolution
• Observerved semantic patterns in the latent space learned by the model