Biology subjects

Casiraghi, E.

Publications and source records attributed to Casiraghi, E..

12 recordsLinked to original sources

bioRxiv2026

Counterfactual Explanations for Graph Neural Networks in Patient Outcome Prediction

Counterfactual Explanation (CE) algorithms have been successfully applied to uncover the main factors driving computational diagnostic and prognostic predictions on tabular medical data. Recently, a new Network Medicine paradigm has been introduced for patient diagnosis and prognosis using Patient Similarity Networks (PSNs), i.e. graphs where patients are represented as nodes and their clinical and biomolecular similarities as edges. In this context, graph-based algorithms, including Graph Neural Networks (GNNs), can provide predictions using not only individual patient features but also their relations within a network of clinically and biomolecularly similar individuals. In this work, we propose the first CE algorithm tailored to explain diagnostic and prognostic predictions within PSNs. Alongside a contrastive GNN backbone, we introduce a versatile, model-agnostic counterfactual search method compatible with any underlying classifier. Preliminary results on synthetic data and on a cohort of patients affected by the Alzheimers disease show that our algorithm is competitive both with seminal tabular based CE algorithms and GNNExplainer, a well-established method for explaining graph-based classification tasks.

bioinformatics↗

bioRxiv2026

On the state of protein function prediction: a report on the fourth CAFA challenge

BackgroundThe Critical Assessment of Functional Annotation (CAFA) is a community effort held to understand the field of computational protein function prediction. Every three years, since 2010, the organizers initiate an experiment to collect function predictions on a large set of proteins and then evaluate the performance of predicting methods on a subset of proteins that have accumulated experimental annotations between the submission deadline and the evaluation time. CAFA provides an independent and rigorous assessment of the current state of the art, thus leveling the playing field, highlighting successes, revealing bottlenecks, and offering a forum for the exchange of ideas in protein science. Here, we report the results of the fourth CAFA experiment (CAFA4). ResultsCAFA4 featured the participation of 148 methods from 70 research groups on a total of 46,205 unique proteins over a 5-year annotation accumulation phase, the longest in any CAFA. In a comparison across CAFA2-CAFA4 methods, the prediction of Gene Ontology (GO) terms has clearly improved across all three GO aspects and traditional evaluation settings. While not achieving the first rank, several CAFA2 and CAFA3 methods featured in the top ten methods in many evaluations, suggesting that earlier methods still hold relevance. The performance is weaker in the newly introduced "partial knowledge" evaluation category (proteins with experimental annotations before submission deadline that gained additional annotations in the same GO aspect during the annotation accumulation phase), highlighting the need for a new class of methods. The rankings of the methods were stable over the years in traditional evaluation settings, but less so in the new partial knowledge evaluation. Overall, the field continues to progress with some influx of new participants. Sustained efforts will be necessary to substantially advance it.

bioinformatics↗

bioRxiv2025

Structured Chemical Reaction Modeling with Multitask Graph Neural Networks

Modeling chemical reactions requires connecting fine-grained atom-bond edits with broader semantic categories. Yet, most machine learning approaches model these aspects in isolation: atom mapping, reaction center identification, and reaction classification are treated as separate problems. This separation limits accuracy, interpretability, and generalization. In this work, we argue that multitask learning provides a natural and powerful framework for reaction modeling. By jointly predicting mappings, centers, and classes within a single graph neural network, models can leverage structural dependencies between tasks. We proposed MARCC (Mapping-Assisted Reaction Center and Classification), a multitask architecture that achieves state-of-the-art performance on the USPTO-50K benchmark. MARCC demonstrates that multitask supervision not only improves accuracy across all tasks but also provides a structured representation of reaction mechanisms.

bioinformatics↗

bioRxiv2025

AI methods and biologically informed data curation enable accurate RNA m5C prediction

5-methylcytosine (m5C) RNA modifications influence nearly every aspect of RNA metabolism, but their transcriptome wide detection is limited by costly, error-prone assays. To bridge this experimental gap, a wave of AI tools now predicts putative m5C sites in silico. However, most existing approaches prioritize architectural complexity while neglecting data quality, so their reported gains mainly reflect the artifacts inherited from noisy datasets. We inverted this paradigm by constructing a high-confidence, methyltransferase-specific catalog of m5C sites, removing artifacts that confound existing resources. Using this curated corpus, we trained (for the first time in a multiclass setting) three different models (Bi-GRU, CNN, Transformer) to distinguish writer-specific m5C sites from unmethylated cytosines. All AI models converged to similar, nearly optimal, performance (AUPRC > 0.97), and a biologically informed analysis revealed that most errors clustered in unmethylated sites mimicking true positives. By augmenting the training set with these hard-to-predict negatives, mined from millions of unmodified cytosines, the models were forced to exploit more nuanced features such as RNA secondary structure and subtle sequence cues, which sharply reduced transcriptome-wide false positive predictions, and predicted methylated transcripts exhibited strong concordance with known methyltransferase biology. Explainable AI techniques also showed that our AI models effectively capture how sequence mutations disrupt m5C sites, underscoring their potential to prioritize disease-relevant variants. The main findings of our study underscore that AI models can be decisive levers for reliable m5C identification only if fed with curated data and validated through biologically informed computational analysis.

bioinformatics↗

bioRxiv2025

miss-SNF: a multimodal patient similarity network integration approach to handle completely missing data sources

MotivationPrecision medicine leverages patient-specific multimodal data to improve prevention, diagnosis, prognosis and treatment of diseases. Advancing precision medicine requires the non-trivial integration of complex, heterogeneous and potentially high-dimensional data sources, such as multi-omics and clinical data. In literature several approaches have been proposed to manage missing data, but usually limited to the recovery of subsets of features for a subset of patients. A largely overlooked problem is the integration of multiple sources of data when one or more of them are completely missing for a subset of patients, a relatively common condition in clinical practice. ResultsWe propose miss-Similarity Network Fusion (miss-SNF), a novel general-purpose data integration approach designed to manage completely missing data in the context of patient similarity networks. Miss-SNF integrates incomplete unimodal patient similarity networks by leveraging a non-linear message-passing strategy borrowed from the SNF algorithm. Miss-SNF is able to recover missing patient similarities and is "task agnostic", in the sense that can integrate partial data for both unsupervised and supervised prediction tasks. Experimental analyses on nine cancer datasets from The Cancer Genome Atlas (TCGA) demonstrate that miss-SNF achieves state-of-the-art results in recovering similarities and in identifying patients subgroups enriched in clinically relevant variables and having differential survival. Moreover, amputation experiments show that miss-SNF supervised prediction of the overall survival and progression-free interval events with completely missing data achieves results comparable to those obtained when all the data are available. Availability and Implementationmiss-SNF code, implemented in R, is available at https://github.com/AnacletoLAB/missSNF.

bioinformatics↗

bioRxiv2025

RNA Knowledge-Graph analysis through homogeneous embedding methods

MotivationWe recently introduced RNA-KG, an ontology-based knowledge graph that integrates biological data on RNAs from over 60 public databases. RNA-KG captures functional relationships and interactions between RNA molecules and other biomolecules, chemicals, and biomedical concepts such as diseases and phenotypes, all represented within graph-structured bio-ontologies. We present the first comprehensive computational analysis of RNA-KG, evaluating the potential of graph representation learning and machine learning models to predict node types and edges within the graph. ResultsWe performed node classification experiments to predict up to 81 distinct node types, and performed both generic and specific edge prediction tasks. Generic edge prediction focused on identifying the presence of an edge irrespective of its type, while specific edge prediction targeted specific interactions between ncRNAs, e.g. miRNA-miRNA or siRNA-mRNA, or relationships between ncRNA and biomedical concepts, e.g. miRNA-disease or lncRNA-Gene Ontology term relationships. Using embedding methods for homogeneous graphs, such as LINE and node2vec, in combination with machine learning models like decision trees and random forests, we achieved balanced accuracy exceeding 90% for the 20 most common node types and over 80% for most specific edge prediction tasks. These results show that simple embedding methods for homogeneous graphs can successfully predict nodes and edges of the RNA-KG, paving the way to discover novel ncRNA interactions and laying the foundation for further exploration and utilization of this rich information source to enhance prediction accuracy and support further research into the "RNA world". Code AvailabilityPython code to reproduce the experiments is available at https://github.com/AnacletoLAB/RNA-KG_homogeneous_emb_analysis

bioinformatics↗

bioRxiv2024

Fine-tuning of conditional Transformers for the generation of functionally characterized enzymes

We introduce Finenzyme, a Protein Language Model (PLM) that employs a multifaceted learning strategy based on transfer learning from a decoder-based Transformer, conditional learning using specific functional keywords, and fine-tuning to model specific Enzyme Commission (EC) categories. Using Finenzyme, we investigate the conditions under which fine-tuning enhances the prediction and generation of EC categories, showing a two-fold perplexity improvement in EC-specific categories compared to a generalist model. Our extensive experimentation shows that Finenzyme generated sequences can be very different from natural ones while retaining similar tertiary structures, functions and chemical kinetics of their natural counterparts. Importantly, the embedded representations of the generated enzymes closely resemble those of natural ones, thus making them suitable for downstream tasks. Finally, we illustrate how Finenzyme can be used in practice to generate enzymes characterized by specific functions using in-silico directed evolution, a computationally inexpensive PLM fine-tuning procedure significantly enhancing and assisting targeted enzyme engineering tasks.

bioinformatics↗

bioRxiv2024

Replacing non-biomedical concepts improves embedding of biomedical concepts

ObjectivesConcept embeddings are low-dimensional vector representations of concepts such as MeSH:D009203 (Myocardial Infarction), whose similarity in the embedded vector space reflects their semantic similarity. Here, we test the hypothesis that non-biomedical concept synonym replacement can improve the quality of biomedical concepts embeddings. Materials and methodsWe developed an approach that leverages WordNet to replace sets of synonyms with the most common representative of the synonym set. ResultsWe tested our approach on 1055 concept sets and found that, on average, the mean intracluster distance was reduced by 8% in the vector-space. Assuming that homophily of related concepts in the vector space is desirable, our approach tends to improve the quality of embeddings. Discussion and ConclusionThis pilot study shows that non-biomedical synonym replacement tends to improve the quality of embeddings of biomedical concepts using the Word2Vec algorithm. We have implemented our approach in a freely available Python package available at https://github.com/TheJacksonLaboratory/wn2vec.

bioinformatics↗

bioRxiv2024

Intrinsic-Dimension analysis for guiding dimensionality reduction and data fusion in multi-omics data processing

The advent of high-throughput sequencing technologies has revolutionized the field of multi-omics patient data analysis. While these techniques offer a wealth of information, they often generate datasets with dimensions far surpassing the number of available cases. This discrepancy in size gives rise to the challenging "small-sample-size" problem, significantly compromising the reliability of any subsequent estimate, whether supervised or unsupervised. This calls for effective dimensionality reduction techniques to transform high-dimensional datasets into lower-dimensional spaces, making the data manageable and facilitating subsequent analyses. Unfortunately, the definition of a proper di-mensionality reduction pipeline is not an easy task; besides the problem of identifying the best dimensionality reduction method, the definition of the dimension of the lower-dimensional space into which each dataset should be transformed is a crucial issue that influences all the subsequent analyses and should therefore be carefully considered. Further, the availability of multi-modal data calls for proper data-fusion techniques to produce an integrated patient-view into which redundant information is removed while salient and complementary information across views is leveraged to improve the performance and reliability of both unsupervised and supervised learning techniques. This paper proposes leveraging the intrinsic dimensionality of each view in a multi-modal dataset to define the dimensionality of the lower-dimensional space where the view is transformed by dimensionality reduction algorithms. Further, it presents a thorough experimental study that compares the traditional application of a unique-step of dimensionality reduction with a two-step approach, involving a prior feature selection followed by feature extraction. Through this comparative evaluation, we scrutinize the performance of widely used dimensionality reduction algorithms. Importantly, we also investigate their impact on unsupervised data-fusion techniques, which are pivotal in biomedical research. Our findings shed light on the most effective strategies for handling high-dimensional multi-omics patient data, offering valuable insights for future studies in this domain. Graphical Abstract O_FIG O_LINKSMALLFIG WIDTH=159 HEIGHT=200 SRC="FIGDIR/small/576822v1_ufig1.gif" ALT="Figure 1"> View larger version (56K): org.highwire.dtl.DTLVardef@c39539org.highwire.dtl.DTLVardef@79f681org.highwire.dtl.DTLVardef@1269fcforg.highwire.dtl.DTLVardef@c5e31d_HPS_FORMAT_FIGEXP M_FIG C_FIG HighlightsO_LIWe introduce a flexible pipeline to guide in a principled way feature selection and feature extraction methods to reduce the high dimensions and to contrast the curse of dimensionality that affects multi-omics data. C_LIO_LIWe harness the power of cutting-edge Intrinsic Dimensionality (id) estimation through block-analysis, providing an unbiased estimation of the individual ids for each view within a multi-modal dataset. C_LIO_LIWe use an exhaustive set of diverse multi-omics cancer datasets from the well-known TCGA dataset to show that the automatic analysis of the distribution of the block-ids characterizing each omics-view leverages dimensionality reduction, by (1) evidencing feature noise and redundancy, and (2) providing an unbiased estimate of the id for each view, to be used for setting the dimension of the reduced space. This avoids empirical or heuristic choices and allows tailoring the reduction to each data-view. C_LIO_LIThe crucial information gained by block-analysis allowed proposing a two-step dimensionality-reduction approach combining feature selection and feature extraction. Our comparative evaluation shows the effectiveness of the proposed technique and its synergy with state-of-the-art data-fusion techniques applied in a multi-omics context. C_LIO_LIWe show that the proposed reduction pipeline leverages traditional dimensionality reduction and state-of-the-art data-fusion algorithms. Indeed, it obtains effective performance when predicting overall survival events with simple random forest classifiers, often preferred in the biomedical field due to their robustness, efficiency, and interpretable nature. C_LI

bioinformatics↗

bioRxiv2022

Node-degree aware edge sampling mitigates inflated classification performance in biomedical graph representation learning

Graph representation learning is a family of related approaches that learn low-dimensional vector representations of nodes and other graph elements called embeddings. Embeddings approximate characteristics of the graph and can be used for a variety of machine-learning tasks such as novel edge prediction. For many biomedical applications, partial knowledge exists about positive edges that represent relationships between pairs of entities, but little to no knowledge is available about negative edges that represent the explicit lack of a relationship between two nodes. For this reason, classification procedures are forced to assume that the vast majority of unlabeled edges are negative. Existing approaches to sampling negative edges for training and evaluating classifiers do so by uniformly sampling pairs of nodes. We show here that this sampling strategy typically leads to sets of positive and negative edges with imbalanced edge degree distributions. Using representative homogeneous and heterogeneous biomedical knowledge graphs, we show that this strategy artificially inflates measured classification performance. We present a degree-aware node sampling approach for sampling negative edge examples that mitigates this effect and is simple to implement.

bioinformatics↗

bioRxiv2022

An algorithmic framework for isoform-specific functional analysis

Gene Ontology (GO) overrepresentation analysis characterizes the biological mechanisms common to sets of differentially expressed genes identified by high-throughput experiments. To date, GO overrepresentation analysis has mainly been used to evaluate differentially expressed genes, but short- and long-read RNA-seq technologies now allow increasingly accurate identification of differential alternative splicing. The function of most splice isoforms remain unknown, but if acccurate predictions could be made, overrepresentation analysis could be applied to differentially spliced isoforms to assess the functional implications of alternative splicing in RNA-seq experiments. We present isopret (Isoform Interpretation), a new paradigm for isoform function prediction based on the expectation-maximization framework. isopret leverages the relationships between sequence and functional isoform similarity to infer isoform specific functions in a highly accurate fashion. This enabled us to adapt GO overrepresentation analysis, which to date has been limited to differential gene expression, to be extended to assess overrepresentation of GO annotations in differentially spliced isoforms. An analysis of 100 RNA-seq studies including investigations of development, cancer, and common disease demonstrated that expression and splicing regulate different sets of biological functions. We make isopret predictions freely available in a desktop application that can be used to analyze differential expression and splicing in any bulk RNA-seq dataset.

bioinformatics↗

bioRxiv2021

Supervised learning with word embeddings derived from PubMed captures latent knowledge about protein kinases and cancer

Inhibiting protein kinases (PKs) that cause cancers has been an important topic in cancer therapy for years. So far, almost 8% of more than 530 PKs have been targeted by FDA-approved medications and around 150 protein kinase inhibitors (PKIs) have been tested in clinical trials. We present an approach based on natural language processing and machine learning to the relations between PKs and cancers, predicting PKs whose inhibition would be efficacious to treat a certain cancer. Our approach represents PKs and cancers as semantically meaningful 100-dimensional vectors based on co-occurrence patterns in PubMed abstracts. We use information about phase I-IV trials in ClinicalTrials.gov to construct a training set for random forest classification. In historical data, associations between PKs and specific cancers could be predicted years in advance with good accuracy. Our model may be a tool to predict the relevance of inhibiting PKs with specific cancers.

bioinformatics↗