Giorgio Valentini

This paper introduces a counterfactual explanation algorithm for graph neural networks to explain patient outcome predictions in patient similarity networks.

This study benchmarks large language models against a traditional diagnostic tool for genetic disease diagnosis using 5,213 case reports, finding LLMs lag behind in accuracy.

The paper trains AI models to predict RNA m5C sites using a curated dataset, achieving high accuracy and revealing the importance of data quality and biological validation.

The paper replaces non-biomedical synonyms with biomedical terms to improve concept embeddings using Word2Vec, reducing mean intracluster distance by 8%.

miss-SNF integrates multimodal patient data with missing sources by recovering patient similarities and identifying clinically relevant subgroups.

The paper analyzes RNA-KG using homogeneous graph embedding methods to predict node types and edges, achieving high accuracy in classification and edge prediction tasks.

This paper introduces a method to guide dimensionality reduction and data fusion in multi-omics data by analyzing intrinsic dimensionality, showing improved performance in predicting survival outcomes.

AutoMAxO uses LLMs to semi-automate MAxO curation for rare diseases, identifying 538 confirmed disease-treatment annotations from 4,918 medical abstracts.

This paper introduces Finenzyme, a Protein Language Model that improves EC category prediction through fine-tuning, generating functionally similar enzymes with preserved structure and function.

The paper introduces a degree-aware edge sampling method to reduce performance inflation in biomedical graph representation learning caused by imbalanced edge degree distributions.

The paper introduces isopret, an algorithm for predicting isoform-specific functions using RNA-seq data, and shows that expression and splicing regulate distinct biological functions.

The paper uses word embeddings from PubMed to predict which protein kinases are relevant for treating specific cancers using machine learning.

SQUIRLS prioritizes splice variants by calculating interpretable features and achieving higher accuracy and speed than existing methods in simulated exomes.

parSMURF is a parallel machine learning tool for detecting pathogenic variants in genomic data, achieving state-of-the-art results with an 80x speed-up over sequential versions.

The paper reports results from the CAFA3 challenge, showing improved computational protein function prediction and new annotations for over 1000 genes through experimental screens in C. albicans, P. aureginosa, and Drosophila.