Research

My work follows one axis: how much biological semantics to formalise in a representation, and what that formalisation buys a learning system. The lines below are the rungs of that ladder, from a reusable semantic substrate to the models that learn on it.

Research lines

A semantic substrate for multi-omics

A property-graph infrastructure that couples an ontology-grounded backbone derived from PheKnowLator and OWL-NETS with per-cohort quantitative multi-omic collections, inside a single multimodal database. Entities and measurements stay jointly queryable: a vector-plus-index data model stores dense per-sample arrays separately from a position-to-feature mapping, so the document count stays linear in samples and cohorts and entity-to-measurement joins resolve inside one query.

The resulting graph holds over 780,000 ontologically grounded entities and 11 million typed relations, coupled to the multi-omic profiles of the TCGA and TARGET cohorts across five layers at native precision, with at least 93% feature coverage on four of the five layers. Built once, rather than rebuilt per study.

Technologies: multimodal NoSQL property graphs (AQL), OWL-NETS, Gene Ontology, Reactome, CHEBI, RO. Status: under review.

Cross-species drug repurposing — PathogenKG

Antimicrobial resistance needs candidate targets across many species, and existing knowledge-graph repurposing systems are built around human disease biology. PathogenKG integrates protein interaction networks from 31 bacterial pathogens with orthology relationships, functional annotation and validated drug–target associations into one graph of over 3 million triples, so that evidence propagates between phylogenetically related organisms.

The hard part is imbalance: 302 known drug–target edges among millions of triples. The training pipeline combines focal loss, type-constrained negative sampling and targeted oversampling, validated on an external biomedical benchmark. Ranked predictions for nine case-study compounds were assessed by domain experts, who rated 22.2% of the top-20 biologically plausible.

Technologies: heterogeneous GNN encoders, CompGCN, knowledge-graph embedding, link prediction. Status: in press, ECML PKDD 2026 (Applied Data Science Track).

Molecular-scale subtype classification — MOGNN-TF

Most prior-knowledge GNNs restrict the prior to a single network type, usually protein interactions, and are evaluated on binary or few-class settings. MOGNN-TF integrates three complementary knowledge sources, protein interactions, post-transcriptional regulation and transcriptional regulation, into a single multi-relational graph where genes, microRNAs and transcription factors are first-class node types, and is evaluated on the full 27-class TCGA pan-cancer taxonomy.

It reaches macro-F1 0.79, outperforming two graph-based baselines with significance confirmed by Friedman and corrected Wilcoxon tests across eleven seeds. A four-axis ablation shows the transcriptional layer can substitute the microRNA layer at no measurable cost, which makes subtype prediction possible from transcriptomic data alone.

Status: under review, IEEE Journal of Biomedical and Health Informatics. Code: Zenodo.

Multi-scale subtype classification — MOKG-HGNN

The same biomedical knowledge graph that underlies the substrate, this time connecting the molecular layer upward to pathways, Gene Ontology terms and diseases as explicit node types, learned with typed message passing and a per-scale readout over three interchangeable heterogeneous backbones.

It does not beat the molecular baseline on aggregate accuracy, and that is the interesting part: what the knowledge graph buys is measured directly. The model degrades less than half as fast under feature scarcity, the two accuracy curves crossing at 50–100 genes, below which the semantic scaffold wins outright, in precisely the clinically realistic regime. It is three times more stable across splits, and it recovers a distinct pathway signature for nearly every subtype, matching canonical drivers such as the androgen-receptor programme in prostate cancer and glutamate-receptor signalling in glioma.

Status: under review, IEEE J-BHI Special Issue. Code: Zenodo.

Deliberation as representation — OmicsCouncil

Latent embeddings of multi-omic data score well on accuracy and poorly on calibration, auditability and resilience to missing inputs, the three properties that high-risk medical AI regulation treats as prerequisites. OmicsCouncil represents a patient by the trace of a multi-agent deliberation rather than by a dense latent vector: modality-specific agents emit typed evidence claims, cross-examine one another, and submit the survivors to a graph-mediated arbiter that grades each claim against a biomedical knowledge prior.

On TCGA-BRCA the representation is interpretable by construction, attains the lowest expected calibration error of the four models tested, and gives up about one accuracy point for those properties. The paper also reports an honest negative result: swapping the statistical agent for an LLM agent degrades classification without improving the trustworthy metrics.

Status: in press, REHMED workshop @ ECML PKDD 2026.

LLMs, retrieval and knowledge graphs

A parallel line on making biomedical knowledge reachable: graph-based information retrieval, BERTopic-driven topic modelling for genomics, retrieval-augmented generation and knowledge-graph-enhanced RAG in the nutrigenetic domain, and REDAC, a web application that runs RNA-seq differential expression analysis from natural-language queries and grounds its pathway interpretation in PubMed through a RAG module over two open LLMs.

Status: published across iiWAS 2024, IRonGraphs @ ECIR 2024, KES 2024, AIBio @ ECAI 2025 and Bioinformatics Advances.

Nutrigenetics

Where I started. A genome-scale curated dataset of nutrition-associated genetic polymorphisms, built by MeSH-driven literature mining and cross-referenced with GWAS data, with a co-occurrence analysis of gene–diet interactions, followed by a computational study of missense variants in nutrition-related genes.

Status: published in Journal of Biomedical Informatics and International Journal of Molecular Sciences. Data: Zenodo.

Methods and tools

Area What I use
Graph learning PyTorch Geometric, DGL, GCN, GAT, GraphSAGE, R-GCN, CompGCN, heterogeneous message passing
Knowledge representation PheKnowLator, OWL-NETS, Gene Ontology, Reactome, STRING, TFlink, RO, CHEBI
Data TCGA and TARGET pan-cancer cohorts, GWAS Catalog, dbSNP, MeSH and PubMed, BRACS
Databases Property and graph databases, multimodal NoSQL, SQL
Language models Transformers, retrieval-augmented generation, GraphRAG, LLM agents, BERTopic
Evaluation Repeated stratified cross-validation, Friedman and corrected Wilcoxon tests, ablation design, calibration and interpretability analysis
Engineering Python, R and Bioconductor, Docker, Git, reproducible releases with archived DOIs

Experience abroad

Visiting researcher, Computer Laboratory, University of Cambridge, March to June 2025, in the research unit of Prof. Pietro Liò. Biomedical knowledge-graph embeddings and graph-based retrieval; graph neural networks benchmarked on Cora, Cornell and Chameleon; prior-knowledge integration in multi-omics GNNs (MOGONET, MOGAT, MoGCN, MPK-GNN); cross-species knowledge-graph construction and link prediction for drug repurposing, which became PathogenKG; and a GraphRAG framework for LLM-driven knowledge-graph construction.

Collaborations

Service

Program committee member for the REHMED workshop at ECML PKDD 2026 and for the 1st AIBio workshop at ECAI 2025; reviewer for a Soft Computing special issue on AI and knowledge graphs in smart industry and health applications; volunteer organising staff at ECML PKDD 2026.

Supervision and teaching

Invited lecturer and teaching assistant for Information Retrieval and Natural Language Processing (MSc in Computer Engineering, Federico II), where I gave seminars on LLM-based topic modelling, designed the assignment Biomedical Knowledge Graphs for Information Retrieval in Precision Medicine and graded it for the exam. Co-supervisor of one MSc and one BSc thesis on knowledge graphs for biomedical data.

If you are a student looking for a thesis or an internship on these topics, get in touch.