Interactive Program
HomeInteractive Program
Search the programme
Room
Track
Day 1
Wednesday 2 September
08:45 - 09:15
Opening Scientific Introduction
Aula MagnaBIOINF Main Track 1
Aula MagnaChair: Davide Chicco09:30 – 09:45PaperBIOINFAn Enhanced Pipeline for Multi-Omic Integration Based on Topological Data Analysis
Veronica Paparozzi, Anna Plaksienko, Marco Pedicini, Christine Nardini
Click to view abstract
Abstract
The advent of high-throughput sequencing techniques have made it essential to employ ad- vanced tools for data integration and interpretation which represent an active area of research. In this work, we propose to build on two existing efforts to enlarge their scope and usabil- ity in multi-omic integration. We exploit Topological Data Analysis (TDA), capitalizing on its graph-based framework, to refine complex interactions characteristic of biological systems; on its intrinsic explainability to enable interpretation of model outputs in terms of biologically meaningful features; and on its robustness to small input perturbations, to ensure stable results in the presence of noisy data, such as omics. Specifically, we utilize a recent tool in TDA, called Harmonic Persistent Homology (HPH), and propose two relevant advances. First, we adopt an integration and clustering tool (i.e. iNETgrate) as an informed means to reduce data size and integrate multiple omic layers by aggregating methylation beta values from loci to the gene level, to be combined with gene expression values. This enables the application of HPH, otherwise limited by the computational burden. Second, as a consequence, we apply HPH to multi-omic molecular data, and not patients, enlarging HPH scope. Finally, we follow up on recent efforts toward performance standardization, and validate our results against biomarkers from a well-known breast cancer benchmark dataset from TGCA.
09:45 – 10:00PaperBIOINFAnalysis of biological networks using Krylov subspace trajectories
H. Robert Frost
Click to view abstract
Abstract
We describe an approach for analyzing biological networks using rows of the Krylov subspace of the adjacency matrix. Specifically, we explore the scenario where the Krylov sub-space matrix is computed via power iteration using a non-random and potentially non-uniform initial vector that captures a specific biological state or perturbation. In this case, the rows the Krylov subspace matrix (i.e., Krylov trajectories) carry important functional information about the network nodes in the biological context represented by the initial vector. We demonstrate theu tility of this approach for community detection and perturbation analysis using the C. Elegans neural network.
10:00 – 10:15PaperBIOINFEvaluating SMT-Based Formal Inference for GRN Synthesis: From Data Density to Structural Rules
Ofri Caspi, Michal Greenberg, Eitan Tannenbaum, Hillel Kugler
Click to view abstract
Abstract
Synthesizing Gene Regulatory Networks (GRNs) from dynamic observations is a fundamental systems biology challenge, essential for deciphering the regulatory mechanisms governing cellular behavior. This work explores a formal inference strategy based on Satisfiability Modulo Theories (SMT) and the Reasoning Engine for Interaction Networks (RE:IN) to systematically identify all networks consistent with biological constraints and trajectory data. The research investigates two dimensions. First, we evaluate the capability of model reconstruction and, separately, examine the impact of observation volume and resolution on the number of consistent solutions (RQ1). Evaluation across 20 models from the Biodivine Boolean Models (BBM) repository shows a 100% recovery rate with five experimental observations, decreasing to 85% with ten observations. This decrease is due to an expressiveness limitation: RE:IN’s regulation condition templates cannot represent the complex or non-monotonic logic of certain biological update functions. Analysis of a subset of models also shows that increasing experimental volume and timestep resolution effectively reduces the solution space. Second, we characterize the formal identification of structural interactions (RQ2) to determine which are logically necessary to explain observed dynamics. We derived and validated six deterministic rules that characterize, within single-activator motifs, the conditions under which interactions are classified as structurally required or disallowed based on topology and trajectory alone. In conclusion, this work evaluates the reliability and exposes the limitations of SMT-based formal inference for genetic network synthesis. It provides a framework for predicting required or disallowed interactions under single-activator motifs, thereby bridging the gap between complex logical inference and biological understanding.
10:15 – 10:30PaperBIOINFA Pipeline for Predicting Variant Associations with Complex Disorders via Functional Annotations
Francesco Gualdi, Zeno Darani, Daniele Malpetti, Marco Scutari, Francesca Mangili
Click to view abstract
Abstract
The study of complex disorders is limited by the lack of structured data sets that integrate genetic association signals with rich functional annotations. While genome-wide association studies (GWAS) provide large collections of associated variants, these are heterogeneous, studydependent, and not directly suitable for machine learning applications. Although frameworks such as Combined Annotation Dependent Depletion (CADD) offer comprehensive variant-level features, there is no standardised procedure to construct curated, balanced, and reproducible data sets for the systematic study of complex traits. We introduce a scalable and reproducible pipeline for generating CADD-annotated variant data sets associated with complex disorders defined via Experimental Factor Ontology (EFO) terms. The pipeline integrates GWAS summary statistics, genome mapping, chromosomestratified negative sampling, and large-scale annotation, while addressing class imbalance and feature redundancy. The resulting data sets provide curated, balanced, and annotation-rich representations of variant–trait associations, suitable for statistical learning, systematic model evaluation and comparison, and analysis of functional annotation contributions and interactions.
Coffee Break
Terrazza del RettoratoChair: Antonello MaruottiA Dependent Dirichlet Process Approach to Population Size Estimation
Lucia Gallucci
Click to view abstract
A Dependent Dirichlet Process Approach to Population Size Estimation
Lucia Gallucci
Abstract
Estimating the size of partially observed populations is a central problem in capture--recapture studies, where only a subset of individuals is detected in the available data. Classical approaches rely on strong assumptions such as population closure and homogeneous capture probabilities, which are often unrealistic in applications and limit their practical usefulness. We propose a flexible Bayesian nonparametric framework for population size estimation under heterogeneous detection. Building on the model of Guindani et al. (2014), we extend the standard capture-recapture formulation by incorporating individual-level covariates through a Dependent Dirichlet Process (DDP) construction. This allows capture probabilities to vary smoothly with observed characteristics, while retaining the flexibility of nonparametric mixture models. Moreover, our approach explicitly incorporates the covariates probability distribution into the DDP model, allowing both cluster allocation and capture intensity to enable a richer representation of heterogeneity. Moreover, the joint structure of the model allows covariate information to be coherently propagated to unobserved individuals, supporting principled inference on their latent covariate profiles and improving estimation of the total population size (n). We develop an efficient Markov chain Monte Carlo algorithm that combines a collapsed Chinese Restaurant Process representation with reversible-jump Metropolis--Hastings updates for the unknown population size. Despite the increased model flexibility, conjugacy is preserved wherever possible, leading to efficient posterior computation and scalable inference. We illustrate the proposed methodology in the context of estimating the prevalence of Autism Spectrum Disorder (ASD) using hospital discharge records in Italy, an inherently incomplete data source, with the aim of estimating the true prevalence.
Parallel sessions
Aula Gini
Chair: Francesco Branda
SS1: AI and Computational Methods for Medical Informatics - Part I
11:00 - 12:00
11:00 – 11:15Poster session
SS1: AI and Computational Methods for Medical Informatics - Part I
Museo dell'Arte Classica
11:15 – 11:30PaperMEDINFDiagnosing Rejection Collapse via Uncertainty DecompositionDiagnosing Rejection Collapse via Uncertainty Decomposition
Walter Endrizzi, ...Walter Endrizzi, Flavio Ragni, Stefano Bovo, Monica Moroni, Giuseppe Jurman, Venet Osmani
Click to view abstract
Abstract
Standard uncertainty-informed rejection can unexpectedly trigger severe performance collapse, exposing localized vulnerabilities that common machine learning metrics typically do not show. We systematically diagnose this failure dynamic using Levodopa-Induced Dyskinesia prediction in Parkinson’s Disease as a proof-of-concept. By training a heterogeneous ML ensemble, decomposing Aleatoric and Epistemic uncertainty and applying unsupervised subgroup discovery, we isolated the precise drivers of these atypical errors. Stratified error analysis revealed two divergent predictive regimes previously hidden by a global evaluation. While the models successfully extracted a predictive signal for one subgroup, the baseline features of a second subgroup lacked discriminative capacity, resulting in a high rate of confident misclassifications. Operating entirely below rejection thresholds, this single subgroup flatlined predictive metrics, driving the collapse of the global rejection curve. Ultimately, we demonstrate that atypical rejection failures stem from subgroup-specific data ambiguity rather than algorithmic deficiencies, making localized uncertainty-aware evaluation a critical methodological requirement prior to real-world deployment.
11:30 – 11:45PaperMEDINFClinRAG-BiLSTM: An Agentic Cloud AI Framework for Explainable and Traceable Breast Cancer Decision SupportClinRAG-BiLSTM: An Agentic Cloud AI Framework for Explainable and Traceable Breast Cancer Decision Support
Adaleta Gicic, ...Adaleta Gicic, Dženana Đonko
Click to view abstract
Abstract
Breast cancer risk prediction requires both strong discrimination and clinically interpretable communication of model outputs. This paper presents ClinRAG-BiLSTM, a retrieval-augmented and agentic cloud AI decision-support framework built on a previously validated Bayesian-optimized BiLSTM predictor for the Wisconsin Diagnostic Breast Cancer dataset [1], [2]. The prior predictive study established the core scoring engine: tensor reshaping of structured diagnostic profiles, Bayesian hyperparameter optimization, and a comparative evaluation against LSTM, FT-Transformer and SAINT baselines. The present extension keeps those predictive results as the foundation and adds a deployed Azure-oriented architecture for PDF guideline retrieval, LLM-based explanation, model registry traceability and user-facing decision support. Across stratified 5-fold cross-validation, the Bayesian-BiLSTM achieved AUC = 0.9985, accuracy = 0.9824, F1-score = 0.9859 and RMSE = 0.0964, outperforming the evaluated baselines. The upgraded system combines prediction results, top-k retrieved guideline chunks and the user question into a grounded response while explicitly limiting the output to clinical decision support rather than autonomous diagnosis.
11:45 – 12:00PaperMEDINFTowards Federated and Explainable Quantum Machine Learning for Possible Epileptic Seizure DetectionTowards Federated and Explainable Quantum Machine Learning for Possible Epileptic Seizure Detection
Francesco Mercaldo, ...Francesco Mercaldo, Hubert Schölnast, Oliver Eigner, Marta Petyx, Antonella Santone, Mario Cesarelli, Fabio Martinelli, Paul Tavolato
Click to view abstract
Abstract
Epileptic seizure detection from electroencephalography recordings represents a critical challenging task due to the non-stationary nature of brain signals, the need for large annotated datasets, and increasing concerns regarding data privacy. To address these issues, we propose a method for federated and explainable possible epileptic seizure detection by means of quantum machine learning. The proposed model integrates classical convolutional neural networks for spatial feature extraction with a variational quantum circuit designed to enhance nonlinear representational capacity. To provide explainability and clinical trust, we consider a multi-level explainability combining class activation mapping, qubit-state analysis and per-qubit saliency maps. Experiments conducted on a real-world dataset demonstrate that the proposed method is able to reach an accuracy of 0.867 in the 10-client setting, with the capability of providing transparent explanations for both classical and quantum model components.
Aula Magna
Chair: Dario Righelli, Duccio Medini, Giuseppe Maccari
SS11: GPU-accelerated analysis of single-cell and spatial omics
11:00 - 12:00
11:00 – 11:15PaperBIOINFToward Simulation-Informed Lipid Nanoparticle Design with Martini Coarse-Grained ModelsToward Simulation-Informed Lipid Nanoparticle Design with Martini Coarse-Grained Models
Mariana Valério, ...Mariana Valério, Pablo Cardona Perez, Salma Maaquili, Mariia Borbuliak, Paulo C. T. Souza
Click to view abstract
Abstract
Lipid nanoparticles (LNPs) have emerged as a central platform for nucleic-acid therapeutics, yet their optimization remains largely empirical due to the complexity of the physicochemical processes governing encapsulation, stability, intracellular trafficking, and endosomal escape. Recent advances in coarse-grained molecular dynamics (CG-MD), particularly with the Martini 3 force field, now enable simulations of increasingly realistic LNP systems across relevant spatial and temporal scales. Beyond mechanistic insight, these simulations may provide access to physically grounded molecular descriptors useful for computational optimization and machine learning workflows. Here, we discuss recent developments in Martini-based CG-MD simulations and their potential contributions to next-generation computational strategies for LNP design. We highlight advances in lipid model development, automated parametrization protocols, approaches for LNP assembly, and emerging strategies for extracting descriptors associated with lipid organization, hydration, RNA interactions, and membrane remodeling. We further discuss how such observables may support large-scale simulation campaigns, simplified LNP mimetics, and future data-driven optimization workflows. Rather than replacing experimental screening, simulation-derived descriptors may help prioritize candidate formulations, rationalize experimental observations, and guide the exploration of the large combinatorial space associated with LNP formulations. Finally, we outline current challenges and future opportunities associated with integrating multiscale molecular simulations, experimental observables, and machine learning methodologies for mRNA delivery systems.
Chair: Duccio Medini, Giuseppe Maccari
11:15 – 11:30PaperBIOINFGPU-accelerated single-cell and spatial transcriptomics on NVIDIA DGX H100: a systematic benchmark of speed, scalability, and biological concordanceGPU-accelerated single-cell and spatial transcriptomics on NVIDIA DGX H100: a systematic benchmark of speed, scalability, and biological concordance
Luca Vedovelli, ...Luca Vedovelli, Corrado Lanera, Daniele Sabbatini, Dario Gregori
Click to view abstract
Abstract
Single-cell RNA sequencing (scRNA-seq) datasets now routinely exceed one million cells, and spatial transcriptomics platforms such as Visium HD generate hundreds of thousands of measurement bins per tissue section, placing severe computational demands on analysis frameworks that remain predominantly CPU-bound. We present a systematic benchmark comparing Scanpy (CPU, 100 cores) against rapids-singlecell (GPU, 1-8 NVIDIA H100 80 GB) on the 1.3-million mouse brain cell dataset, evaluating speed, multi-GPU scalability, memory efficiency, and biological concordance across five dataset sizes (10k-1.3M cells) with five independent repeats. The GPU pipeline achieved up to 120-fold end-to-end speedup at 1.3M cells (435 s vs 52,056 s on CPU), reducing a 14.5-hour analysis to 7.3 minutes. Biological concordance was high: HVG selection was identical (Jaccard = 1.0), PCA loadings perfectly correlated (Spearman |rho| = 1.0), and Leiden clustering concordance ranged from ARI = 0.908 to 0.963. Multi-GPU scaling was sublinear due to CPU preprocessing dominance. A memory-optimised pipeline processed 11.9 million cells on a single DGX H100 node, with CPU RAM as the binding constraint (GPU VRAM at 7.6% of capacity). A differential expression benchmark at 3.4M cells showed pseudo-bulk aggregation 44x faster than cell-level t-test. We extended the benchmark to spatial transcriptomics on three 10x Visium platforms, achieving up to 51.6x end-to-end speedup (Visium HD 8 um) and 3,272x for co-occurrence analysis, with near-perfect spatial autocorrelation concordance (Spearman rho >= 0.9995). These results identify CPU-side preprocessing as the primary bottleneck and provide practical guidance for deploying GPU-accelerated workflows at atlas scale.
Chair: Dario Righelli
11:30 – 11:45PaperBIOINFCudaMon: An R Package to Monitor NVIDIA GPUs, Showcased by Monitoring a GPU-accelerated Single-cell Analysis Workflow in RCudaMon: An R Package to Monitor NVIDIA GPUs, Showcased by Monitoring a GPU-accelerated Single-cell Analysis Workflow in R
Mohammad Amin Zadenoori, ...Mohammad Amin Zadenoori, Riccardo Ceccaroni, Davide Risso, Gabriele Sales
Click to view abstract
Abstract
NVIDIA GPUs have recently started to be used in computational biology, yet R users lack integrated GPU monitoring tools, forcing reliance on external utilities like nvidia-smi. We introduce CudaMon, an R package providing real-time monitoring of GPU utilization, memory, temperature, and power draw via NVML, with data export and visualization utili- ties. Monitoring a GPU-accelerated single-cell RNA-seq pipeline (1M brain cells, RAPIDS workflow) shows compute-intensive steps (PCA, UMAP, t-SNE) exceed 90% GPU utilization, while data management phases reveal bottlenecks. CudaMon facilates resource optimization, performance debugging, and reproducibility for GPU-accelerated R workflows.
Chair: Dario Righelli
11:45 – 12:00PaperBIOINFDualBioGraph: Patient-Held-Out Tumor--Normal Classification Using Dual Graph Neural NetworksDualBioGraph: Patient-Held-Out Tumor--Normal Classification Using Dual Graph Neural Networks
Al Hamna Asif, ...Al Hamna Asif, Rana Abubakar, Amna Younus, Juan Jose Vegas Olmos, Filippo Cugini
Click to view abstract
Abstract
Graph neural networks (GNNs) are increasingly used for single-cell cancer analysis, but their reported gains are difficult to interpret when patient-level leakage, graph construction choices, and computational feasibility are not evaluated together. We present \textbf{DualBioGraph}, a patient-held-out benchmarking framework and architecture suite for heterogeneous cell-gene graph learning in tumor-normal single-cell RNA-seq classification. DualBioGraph evaluates expression-based baselines, graph ablations, and enhanced GNN variants across lung, breast, and prostate cancer cohorts using the same patient/sample-held-out protocol. The framework constructs cell-cell KNN graphs, gene-gene correlation/pathway graphs, and sparse gene-cell expression bipartite graphs. Sparse top-$k$ gene-cell edge selection is used to reduce the cost of heterogeneous message passing and support execution on a modest GPU. We benchmark Dual Bipartite GraphSAGE, Residual Dual GraphSAGE, Gated Expression-Fusion GraphSAGE, Latent-Graph GraphSAGE with raw/PCA/autoencoder KNN construction, Sparse Bipartite GraphSAGE, and attention-based DualBioGraph variants with expression fusion, contrastive learning, and patient-domain regularization. The final benchmark shows that the best-performing model differs by cancer cohort: PCA-logistic regression achieves the highest F1 score in lung cancer, graph-based variants achieve the strongest results in breast cancer, and an MLP baseline performs best in prostate cancer. These results show that heterogeneous graph learning is competitive and computationally practical under realistic patient-held-out validation, but it should not be reported as universally superior to simpler baselines. DualBioGraph therefore provides a reproducible benchmark for evaluating when graph structure improves single-cell tumor-normal classification.
Chair: Dario Righelli
Aula III
Chair: Annamaria Carissimo, Dario Righelli
SS5: Biostatistical and Artificial Intelligence Methods for Personalized Medicine and Health
11:00 - 12:00
11:00 – 11:15PaperBIOINFA topological object-proposal pipeline for cellular fiber segmentationA topological object-proposal pipeline for cellular fiber segmentation
Riccardo Ceccaroni, ...Riccardo Ceccaroni, Valerio Reffo, Camilla Pezzini, Roberta Sartori, Marco Sandri, Pierpaolo Brutti, Davide Risso
Click to view abstract
Abstract
Segmentation of muscle fibers in H&E histology is less standardized than nuclei or whole-cell segmentation. Boundaries can be weak, staining can vary within the same image, and local intensity or texture may fail exactly where object separation is needed. However, the geometric information is often stable: in cross-sectional views, fibers tend to close into loops. 1-dimensional Persistent Homology (PH) is designed to capture such closed structures by tracking them across intensity thresholds. In this setting, topology becomes an object-discovery principle: a persistent loop provides evidence of a candidate fiber. This work presents an early-stage PH-guided object-proposal pipeline that extracts 1-dimensional persistent classes from H&E images and converts them into candidate instance masks. The resulting candidates are filtered by persistence and area, then post-processed to reduce redundant concentric detections. In the absence of ground-truth annotations, we use Cellpose as a reference point rather than as a competing baseline. The topological pipeline produces a larger set of candidate objects and covers a larger fraction of estimated non-background tissue, but these measurements are descriptive and do not establish segmentation accuracy. These preliminary results suggest that 1-dimensional PH can provide a candidate object-proposal step for fiber structures, which could then be refined by dedicated segmentation methods, including deep-learning approaches such as Cellpose. Because mask extraction is naturally parallelizable, future GPU development could reduce processing time and make the workflow feasible within practical runtimes.
Chair: Dario Righelli
11:15 – 11:30PaperBIOSTATQuantum Machine Learning for Missense Variant Pathogenicity Prediction in the KCNQ Gene FamilyQuantum Machine Learning for Missense Variant Pathogenicity Prediction in the KCNQ Gene Family
Markel Garcia, ...Markel Garcia, Sara Capponi, Aitor Bergara, Aritz Leonardo
Click to view abstract
Abstract
Missense variants in the KCNQ family of voltage-gated potassium channel genes are associated with a broad spectrum of neurological and cardiac disorders for which computational tools capable of differentiating benign and pathogenic variants are essential. Here, we present the first application of quantum machine learning (QML) to the problem of genetic pathogenicity prediction. We employ a hybrid quantum framework combining data re-uploading quantum neural networks with quantum support vector machines and applied it to pathogenicity classification across the full KCNQ gene family. Systematic exploration of circuit architectures revealed that optimal hyper parameter choices enable the proposed quantum model to achieve predictive performance superior or comparable to state-of-the-art classical machine learning approaches, reaching AUC-ROC values of approximately 0.94 across the full dataset. Although no clear quantum advantage is observed at present, these results demonstrate that quantum feature representations can effectively capture biologically relevant patterns in variant data and establish QML as a viable framework for future applications in computational genomics and precision medicine.
Chair: Annamaria Carissimo
11:30 – 11:45PaperBIOSTATZero-shot phenotype prediction and screening of engineered immune receptor variants using contrastive learning-based workflowZero-shot phenotype prediction and screening of engineered immune receptor variants using contrastive learning-based workflow
Jie Shi, ...Jie Shi, Yenho Chen, Kyle Daniels, Shangying Wang, Sara Capponi
Click to view abstract
Abstract
Chimeric antigen receptors (CARs) are engineered T cell receptors that identify precisely cell surface antigens. By integrating new co-stimulatory signaling components into the CAR T intracellular domain, it is possible to engineer CAR T cell functions. Point mutations in co-stimulatory domains might affect CAR T cell phenotype but exploring experimentally CAR T phenotypes caused by different co-stimulatory variants remains often unexplored. Here, we present an ESM2-based framework for predicting CAR T cell phenotypes, specifically cytotoxicity, and for in silico screening co-stimulatory domain variants. The framework combines a consensus-motif-guided mutation augmentation module with contrastive continued pretraining to adapt protein large language model representations to CAR signaling motifs and uses supervised prediction of cytotoxicity from sequence embeddings. The learned representations capture consensus-dependent effects of single-site mutations across the tested motifs, and the supervised model achieves held-out cytotoxicity prediction performance of R^2 = 0.705. The model generalizes to external natural signaling sequences not included in the training dataset, predicting cytotoxicity trends consistent with experimental measurements. This framework enables amino acid level analysis of CAR co-stimulatory motifs and provides a computational strategy for accelerating the search and design of signaling domain variants with desired cytotoxicity.
Chair: Annamaria Carissimo
11:45 – 12:00PaperBIOSTATA Landmarked Penalised Weibull Framework for Time-Updated Survival Modelling of Heterogeneous Clinical Outcomes in Multiple SclerosisA Landmarked Penalised Weibull Framework for Time-Updated Survival Modelling of Heterogeneous Clinical Outcomes in Multiple Sclerosis
Francisco José Aparicio Serrano, ...Francisco José Aparicio Serrano, Ariadna Masot-Llima, Agustín Pappolla, Susana Otero-Romero, René Carvajal, Álvaro Cobo-Calvo, Manel Alberich, María Jesús Arévalo, Georgina Arrambide, Cristina Auger, Joaquín Castillo, Manuel Comabella, Ingrid Galán, Daniel Hernández, Carlos Nos, Jordi Río, Breogán Rodríguez-Acevedo, Jaume Sastre-Garriga, Ángela Vidal-Jordana, Ana Zabalza, Àlex Rovira, Xavier Montalban, Mar Tintoré, Marco Lorenzi, Deborah Pareto, Carmen Tur
Click to view abstract
Abstract
Predicting the longitudinal course of multiple sclerosis (MS) requires modelling frameworks capable of integrating time-varying data and generating updated risk estimates as new clinical information accumulates. We present a time-updated survival modelling framework based on landmarked Weibull accelerated failure time (AFT) regression, applied to the prediction of heterogeneous MS outcomes. At each annual landmark, outcome-specific models are refitted using all data available up to that time point, incorporating both baseline features (age, sex, lesion number, oligoclonal band status) and time-updated covariates updated via last observation carried forward (LOCF): current disability level (i.e., expanded disability status scale [EDSS]), prior relapses and progression events, new MRI lesions, and treatment exposure. We modelled two representative outcome types: reaching EDSS 3.0 (moderate disability) as an absolute endpoint (terminal for that outcome model once reached), and next relapse as a recurrent event (patients remain at risk after each occurrence). The framework was developed on 1334 patients from the Barcelona first-attack cohort (median follow-up 9.7 years) and externally validated on 450 patients from a clinical trial (BENEFIT study; median follow-up 7.5 years). Internal validation yielded C-index values of 0.71–0.78 for EDSS 3.0 and 0.67–0.73 for relapse prediction. External validation in the independent BENEFIT cohort demonstrated comparable discrimination and overlapping bootstrap confidence intervals, supporting generalisability across cohorts. The framework generates time-updated individual risk trajectories with associated confidence intervals, demonstrates generalisability across independent cohorts, and provides a practical approach to dynamic survival modelling in longitudinal biomedical data.
Chair: Annamaria Carissimo
Parallel sessions
Aula III
Chair: Antonello Maruotti, Graziano Pesole
BIOSTAT Main Track 1
12:00 - 13:15
12:00 – 12:15PaperBIOINFCellChat Hotspot: A Focused Lens on Tumor Microenvironment CommunicationCellChat Hotspot: A Focused Lens on Tumor Microenvironment Communication
Dario Monaco, ...Dario Monaco, Mirea Dioguardi, Maria Dipalma, Eliseo Mattioli, Francesco Alfredo Zito, Francesco Giovannelli, Angela Ricco, Oronzo Brunetti, Antonella Argentiero, Simona De Summa
Click to view abstract
Abstract
In recent years, single-cell RNA sequencing (scRNA-seq) has enabled the investigation of intercellular communication through Cell–Cell Communication (CCC) tools such as CellChat. Despite their potential, these approaches remain challenging to interpret, particularly when global analyses generate complex interaction networks that may include biologically irrelevant signals. To address this limitation, we developed a framework termed CellChat Hotspot, designed to improve the biological interpretability of inferred communication networks. This approach integrates gene set enrichment analysis with a context-driven filtering strategy, selecting both genes and cells based on their functional relevance prior to communication inference. Applied to stage II colon cancer data, CellChat Hotspot enabled the identification of condition-specific interaction patterns associated with recurrence. Non-recurrent patients showed enrichment of immune-related processes, whereas recurrent cases were characterized by pathways associated with invasion and metastasis. Notably, a reduction in CDH-mediated signalling between CALB1⁺ and epithelial cell populations was observed in recurrent patients, suggesting a potential disruption of cell–cell adhesion mechanisms. Overall, this framework provides a biologically guided strategy for studying intercellular communication in complex systems such as the tumor microenvironment.
Chair: Graziano Pesole
12:15 – 12:30PaperBIOSTATEvaluating in-context learning with prompting regimes for transformer-based synthetic health tabular data generationEvaluating in-context learning with prompting regimes for transformer-based synthetic health tabular data generation
Amanda Bertgren, ...Amanda Bertgren, Fredrik Öhberg, Paolo Soda, Ulf Näslund, Patrik Wennberg, Christer Grönlund
Click to view abstract
Abstract
Tabular synthetic data generation (SDG) can facilitate research and development in the health domain, as access to health data is restricted. To accommodate for its structural complexity, such as multimodal distributions and fragmentation, transformer models provides an alternative to conventional SDG methods since they are better equipped to infer global spatial coherence due to textual context of tabular records. However, methods for optimising their adaption to tabular SDG are still largely unexplored. We contribute to the understanding of transformer-based SDG by evaluating fidelity and privacy of three prompting regimes, zero shot, one shot and few shot training, for tabular health SDG using small transformer models. We evaluated llama 3.2 (1B and 3B parameters) and Phi (1.3B and 2.7B parameters), on a real clinical dataset. Our results showed that zero shot learning was less prone to hallucinations and performed overall better for fidelity, but at the loss of privacy in comparison to the other regimes. Our results showed no differences in performances across model sizes and families. Future studies should extend the analysis to larger models and more datasets, and explore strategies to reduce hallucinations.
Chair: Antonello Maruotti
12:30 – 12:45PaperBIOSTATMultivariate Model-Based Landmarking Approach in Mixture Cure Models to Dynamic PredictionMultivariate Model-Based Landmarking Approach in Mixture Cure Models to Dynamic Prediction
Bianca Ferraro, ...Bianca Ferraro, Marco Alfò, Marta Cipriani
Click to view abstract
Abstract
In this work, we extend model-based landmarking approaches for mixture cure models by considering multiple longitudinal biomarkers, rather than analyzing each biomarker separately. This approach captures the correlation among biomarkers and it may provide more informative patient-specific summaries for prediction of post-landmark survival. We applied our method to 388 renal transplant patients with serial measurements of glomerular filtration rate, haematocrit, and proteinuria during the post-transplantation follow-up period. The mGLMM revealed strong correlations among random effects, confirming the importance of a multivariate approach. Proteinuria emerged as the primary predictor of post-landmark graft failure across landmark times. The multivariate model demonstrated superior discriminative ability compared to univariate approaches, as evidenced by time-dependent AUC values. These findings demonstrate the value of jointly modeling multiple biomarkers while accounting for their correlation structure in dynamic prediction within mixture cure models.
Chair: Antonello Maruotti
12:45 – 13:00PaperBIOSTATFunctional Fuzzy Clustering of Longitudinal biological marker's trajectoriesFunctional Fuzzy Clustering of Longitudinal biological marker's trajectories
Silvia D'Elia, ...Silvia D'Elia, Valeria Paggi, Marco Alfò, Maria Brigida Ferraro
Click to view abstract
Abstract
Serum thyroglobulin (Tg) is a key biomarker in the follow-up of differentiated thy- roid cancer (DTC) after total thyroidectomy and radioiodine remnant ablation. However, serial Tg measurements are often sparse, irregularly timed, and heterogeneous across patients, making longitudinal interpretation challenging. This study aimed to model post-treatment Tg measurements as functional data, identify latent trajectory profiles through fuzzy clustering, and assess their association with risk stratification and response-to-treatment. Patients treated with total thyroidectomy followed by radioiodine remnant ablation and with at least 5 years of follow-up after initial treatment were included. Log-transformed Tg values were analyzed to reduce the influence of extreme observations. Individual trajectories were reconstructed using B-spline smoothing, and spline coefficients have been used as low-dimensional descriptors of temporal patterns. These features were analyzed through fuzzy clustering, allowing partial membership across clusters. The optimal solution identified three clusters (28, 169, and 43 patients) all characterized by high average membership degrees. Clusters differed mainly in average Tg level rather than slope, corresponding to low, intermediate, and persistently higher Tg profiles. Clusters did not mirror ATA 2015 baseline risk classes, suggesting that initial stratification did not fully capture subsequent biomarker dynamics. In contrast, clear differences emerged in response-to-treatment categories: Cluster 1 was mainly associated with excellent response, whereas Cluster 3 showed the highest frequencies of structural and biochemical incomplete response. Functional Data Analysis (FDA) combined with fuzzy clustering provides an interpretable framework for analyzing irregular Tg trajectories and may support personalized follow-up beyond baseline risk stratification in DTC patients.
Chair: Antonello Maruotti
13:00 – 13:15PaperBIOSTATReducing the Size of Breastfeeding Reference Arms in Infant Nutrition Studies Using Prognostic Covariate AdjustmentReducing the Size of Breastfeeding Reference Arms in Infant Nutrition Studies Using Prognostic Covariate Adjustment
Luca LavalleLuca Lavalle
Click to view abstract
Abstract
Breastfed (BF) infants are commonly included as a non‑randomized reference group in pediatric nutrition trials comparing experimental and control formulas, providing essential context for interpreting growth and other health outcomes. However, recruitment of breastfeeding participants can present ethical and operational challenges. This project investigates the use of the Prognostic Covariate Adjustment (PROCOVA) method – digital twin concept – on legacy data to reduce the size of BF reference arms in future pediatric nutrition studies. Legacy demographic and anthropometric data from 9 completed clinical studies, comprising 1,275 breastfed infants, were pooled to develop a machine‑learning‑based prognostic model for infant weight at 4 months of age. Multiple modeling approaches, imputation strategies, and hyperparameter combinations were evaluated using cross‑validation, and the final model was validated on an independent out‑of‑sample dataset (the CLARA study). The prognostic model demonstrated consistent performance, with an out‑of‑sample Pearson correlation coefficient of 0.58, corresponding to 96% of the in‑sample correlation and meeting PROCOVA validation requirements. The PROCOVA model was subsequently applied retrospectively to the CLARA study to assess potential sample size reduction for the primary endpoint, weight gain from enrollment to 4 months of age. Compared with a classical sample size approach, prognostic covariate adjustment reduced the required sample size from 110 to 74 infants per arm, representing an approximate 33% reduction while maintaining 90% statistical power. These results demonstrate the potential of PROCOVA to optimize study design and reduce participant burden, leveraging legacy data.
Chair: Antonello Maruotti
Aula Gini
Chair: Francesco Branda
SS1: AI and Computational Methods for Medical Informatics - Part II
12:00 - 13:15
12:00 – 12:15PaperMEDINFPredicting Alzheimer’s Progression Over Time Using Sheaf Neural Networks and Brain GraphsPredicting Alzheimer’s Progression Over Time Using Sheaf Neural Networks and Brain Graphs
Annamaria Defilippo, ...Annamaria Defilippo, Giulia Avolio, Pierangelo Veltri, Pietro Liò, Pietro Hiram Guzzi
Click to view abstract
Abstract
Predicting the longitudinal progression of Alzheimer’s disease (AD) from multimodal neuroimaging data remains a challenging task for machine learning. Clinical observations are acquired at irregular time intervals, structural brain measurements are high-dimensional and often incomplete, and interactions among brain regions may differ across anatomical systems. This study introduces a spatio‑temporal deep learning framework that represents each patient as a sequence of brain graphs and performs end‑to‑end next‑visit diagnosis prediction among Cognitively Normal (CN), Mild Cognitive Impairment (MCI), and Dementia (D) stages. The spatial component is modelled through a Sheaf Neural Network (SNN), which enables flexible, region-specific information exchange between connected brain areas. Short-range intra-lobar connections are distinguished from long-range inter-lobar connections, while temporal edges link the same anatomical region across consecutive visits and encode the elapsed time between observations. Visit-level graph representations are subsequently processed by a Long Short-Term Memory (LSTM) network to capture longitudinal disease trajectories. The proposed pipeline is evaluated on the Alzheimer’s Disease Neuroimaging Initiative (ADNI) dataset, integrating clinical variables with FreeSurfer-derived morphometric features.
12:15 – 12:30PaperMEDINFOntology-Aware Candidate Reranking for SNOMED CT Entity Linking in Clinical NotesOntology-Aware Candidate Reranking for SNOMED CT Entity Linking in Clinical Notes
Luka Blašković, ...Luka Blašković, Nikola Tanković, Ivo Ipšić
Click to view abstract
Abstract
Biomedical entity linking normalizes clinical mentions to controlled medical terminologies, but short and context-dependent mentions remain difficult to resolve with lexical matching alone. This paper studies ontology-aware candidate reranking for SNOMED CT concept normalization in clinical notes. We rank ten gold-contained candidate SNOMED CT concepts per mention using either preferred labels alone or ontology-enriched concept text containing labels, synonyms, and UMLS semantic types. On 75,491 mention-level examples, we compare fuzzy matching, TF-IDF, BM25, SapBERT, and fine-tuned MiniLM cross-encoders. The ontology-aware MiniLM reranker achieves 90.4% Recall@1 and 94.2% MRR, outperforming the label-only reranker by 2.64 percentage points in Recall@1 and 1.84 points in MRR; the gain persists on a note-disjoint sensitivity subset.
12:30 – 12:45PaperMEDINFEvaluating Locally-Deployable Large Language Models on Free-Text Data in an Italian Obstetric ContextEvaluating Locally-Deployable Large Language Models on Free-Text Data in an Italian Obstetric Context
Pierluigi Reali, ...Pierluigi Reali, Giulio Steyde, Gianluca Carta, Mark James Carman, Maria Gabriella Signorini
Click to view abstract
Abstract
Electronic health records often contain unstructured free text characterized by acronyms, jargon, and inconsistent syntax, making automated research difficult. Due to strict privacy regulations and hardware constraints in clinical environments, this paper evaluates the performance of locally deployable open-weight Large Language Models (LLMs) for correcting and extracting structured information from noisy, non-English medical text. A modular automated pipeline was developed to test seven open-weight model families (Gemma3, Llama3, MedGemma, Mistral, GPT-OSS, Qwen3, and DeepSeek), ranging from 4 to 70 billion parameters, and executed locally through the Ollama framework. Models were evaluated on two tasks: correcting typos and expanding Italian obstetric acronyms, and extracting 28 clinical fields, including Apgar scores and blood gas values. Performance was assessed against an expert-validated gold standard of 100 notes using both string-based metrics and semantic-embedding similarity. We also compared several prompt-engineering strategies, including few-shot, positive instructions, domain-specific acronym lists, and prompt repetition. Results show that local LLMs can effectively clean and structure clinical data, although performance depends strongly on model architecture and prompt design. Injecting domain knowledge into prompts significantly improved correction accuracy, outperforming few-shot prompting and larger model sizes. In extraction tasks, smaller models such as Qwen3 8B matched or exceeded larger counterparts, indicating that instruction-following ability can outweigh parameter count. Overall, Mistral and Gemma3 achieved the best balance between semantic accuracy and inference latency, demonstrating that small local LLMs can unlock valuable clinical data without cloud-based APIs or expensive fine-tuning.
12:45 – 13:00PaperMEDINFDecoding HIV-1 Antibody Escape with Interpretable Protein Language ModelsDecoding HIV-1 Antibody Escape with Interpretable Protein Language Models
Fahsai Nakarin, ...Fahsai Nakarin, Pawin Taechoyotin, Kayla G. Sprenger
Click to view abstract
Abstract
Broadly neutralizing antibodies (bnAbs) offer promise for HIV-1 prevention, yet their efficacy is often undermined by distributed viral escape mechanisms. Understanding how HIV-1 evades best-in-class bnAbs requires decoding the context-dependent escape fitness landscape of the viral envelope (Env). We fine-tune a protein language model (PLM) on naturally circulating HIV-1 Env sequences paired with bnAb VRC01 neutralization phenotypes, redirecting the model from general evolutionary knowledge toward antibody-specific viral escape fitness. By integrating gradient-based attribution with attention-derived residue communication maps, we derive sequence-specific fitness scores that resolve viral escape into interpretable residue-level and epistatic-coupling contributions. Our results show that these scores strongly correlate with VRC01 neutralization phenotypes, enabling escape predictions with both residue position- and amino acid substitution-level accuracy. The model further reveals dynamic epistatic rewiring of viral escape, demonstrating how residues' functional impact changes with sequence context. Together, our work establishes a scalable foundation for decoding how viruses evolve under antibody pressure, offering a principled route to anticipate viral escape from the immune system.
13:00 – 13:15PaperMEDINFAutoEncoder based approach for the identification of genomic regions responsible for poorly described oligo-patient disordersAutoEncoder based approach for the identification of genomic regions responsible for poorly described oligo-patient disorders
Joanna SzydaJoanna Szyda
Click to view abstract
Abstract
Chronic fatigue syndrome is a clinically poorly defined disorder with a possibly complex aetiology, posing significant challenges for genetic analysis due to its rarity and poor clinical definition. Traditional Genome Wide Association Studies often rely on large sample sizes, which are not feasible for such clinical conditions. We propose an AutoEncoder-based approach to identify anomalous genomic patterns in affected individuals, leveraging deviations from healthy populations of matched origin. As a result, genes (CHRNE, CNTNAP3, SPATA31A7, MYO5BP3, SPATA31A, MAN1B1, ATP8A2, HOXB3, TPTE) and noncoding regulatory elements (lncRNA and an enhancer) were marked as candidates for further consideration. For the trained models, SNP genotype classification accuracies varied between 0.72 and 0.88, while the residual mean of SNP genotype reconstruction varied between 0.02 and 0.59, depending on the chromosome.
Aula Magna
Chair: Duccio Medini, Giuseppe Maccari, Francesco Branda
SS10: Computational tools for the effective design of mRNA-based therapeutics
12:00 - 13:15
12:00 – 12:15PaperMEDINFPrediction of treatment response in Multiple Sclerosis with Machine Learning modelsPrediction of treatment response in Multiple Sclerosis with Machine Learning models
Ariadna Masot-Llima, ...Ariadna Masot-Llima, Francisco Aparicio-Serrano, Agustín Pappolla, Susana Otero-Romero, René Carvajal, Álvaro Cobo-Calvo, Manel Alberich, María Jesús Arévalo, Georgina Arrambide, Cristina Auger, Joaquín Castillo, Manuel Comabella, Ingrid Galán, Daniel Hernández, Carlos Nos, Jordi Río, Breogán Rodríguez-Acevedo, Jaume Sastre-Garriga, Ángela Vidal-Jordana, Ana Zabalza, Àlex Rovira, Xavier Montalban, Mar Tintoré, Xavier Lladó, Deborah Pareto, Carmen Tur
Click to view abstract
Abstract
Predicting treatment response (TR) in patients with multiple sclerosis (pwMS) at the time of treatment initiation remains a major clinical challenge. Current TR clinical approaches require one year of treatment exposure, delaying optimised therapy selection while disability irreversibly accumulates. This paper aims to evaluate whether machine learning (ML) models trained on baseline clinical and imaging variables can predict unfavourable outcomes in pwMS initiating moderate-efficacy (Interferon Beta, Glatiramer Acetate) disease-modifying therapies (DMTs). pwMS from Barcelona (N=471) and the BENEFIT clinical trial (N=296) were used as development and validation cohorts, respectively. Four ML classifiers: Logistic Regression (LR), Random Forest (RF), Support Vector Machine (SVM), and XGBoost (XGB) were trained to predict 2-year outcomes (relapses, confirmed disability worsening [CDW], new T2 lesions [NL], and evidence of disease activity [EDA]: any new lesions, relapses or CDW events). Models were evaluated via 1000-iteration bootstrap. Baseline (BL) models were compared to extended models incorporating 1-year follow-up data (BL+1Y). BL models achieved AUCs of 0.56--0.77 with substantial variability across outcomes. LR achieved the highest AUC for EDA (0.77) and CDW (0.68), while RF achieved the highest AUC for NL (0.75) and relapses (0.67); no single model type was uniformly superior. BL+1Y models did not significantly outperform BL in most comparisons (McNemar p$>$0.05). External validation revealed meaningful cohort shift (Standardised Mean Difference of past relapses (SMD) = 1.91; SMD disease duration = 1.07), with AUC degradation observed across all models and outcomes; relapse prediction generalised most consistently, while new lesions and EDA showed the largest declines. Baseline-only ML models display discriminative ability for 2-year TR without requiring follow-up data, though no single algorithm consistently dominated across outcomes. External validation highlights generalisability challenges.
Chair: Francesco Branda
12:15 – 12:30PaperBIOINFSecondary structure design for efficient translation initiationSecondary structure design for efficient translation initiation
Tobias von der Haar, ...Tobias von der Haar, Corina Keller
Click to view abstract
Abstract
Controlling secondary structure is potentially useful in the design of mRNA Therapeutics, since structural features can balance protein synthesis and RNA degradation in vivo, and can affect shelf life during storage of formulated therapeutics. Different measures of RNA secondary structure have been targeted in different design approaches for different purposes, but which measure(s) to target for optimal therapeutic efficacy is not yet fully understood. Here we ask whether optimisation of secondary structure around the start codon can be specifically used to enhance translation initiation efficiency. We find that secondary structure formation between untranslated regions and the open reading frame can be observed in many therapeutic sequences, and the effect of such structures can be usefully quantified via their opening energies, which are strong predictors of translation initiation efficiencies measured in single molecule studies. We are currently testing whether the opening energy around the start codon is a useful parameter to include to optimise therapeutic efficacy in vivo.
Chair: Duccio Medini, Giuseppe Maccari
12:30 – 12:45PaperBIOINFManufacturability of mRNA Therapeutics: Native vs. N-1-Methyl-PseudouridineManufacturability of mRNA Therapeutics: Native vs. N-1-Methyl-Pseudouridine
Roland Huber, ...Roland Huber, Kuo Chieh Liao, Yue Wan
Click to view abstract
Abstract
The rapid clinical validation of cell-free mRNA therapeutics has underscored the necessity for robust, scalable manufacturing platforms. However, the manufacturability of mRNA in vitro is significantly influenced by the specific sequence of the desired product, frequently exhibiting two distinct failure modes: low overall yield and early truncation of the transcript. While common biotechnological RNA polymerases are evolutionarily adapted to utilize native nucleotides, resulting in highly processive transcription under natural conditions, therapeutic mRNA utilizing exclusively native nucleotides remains fundamentally limited by innate immunogenicity. The standard industry mitigation, complete replacement of uridine with 1-methyl-pseudouridine (m1Ψ), effectively circumvents immune detection but introduces additional complexity to both the manufacturability and subsequent translatability of the product. In this study, we rigorously investigate the yield and premature termination characteristics of template-based mRNA manufacturing, directly comparing native uridine and m1Ψ incorporation. Our analysis reveals that while bulk yield profiles remain highly comparable between the two chemistries, their early termination characteristics diverge significantly. Native nucleotides demonstrate stronger, more consistent early termination in response to specific sequence motifs. Conversely, m1Ψ offers a highly favourable, distinct profile, being overall less sensitive to sequence-induced abortive transcription. We identify a catalogue of structural motifs that uniquely trigger early truncation in native nucleotide sequences, compared to limited subset of motifs that specifically terminate m1Ψ elongation.
Chair: Duccio Medini, Giuseppe Maccari
12:45 – 13:00PaperBIOINFOptimizing RNA yield using deep neural networks coupled to massively parallel screeningOptimizing RNA yield using deep neural networks coupled to massively parallel screening
Adrien VillainAdrien Villain
Click to view abstract
Abstract
Messenger RNA (mRNA)-based therapeutics have emerged as a powerful platform for vaccines, protein replacement therapies, and cancer immunotherapy. A critical bottleneck in mRNA development is manufacturing large quantities of RNA economically, as measured by RNA yield emerging from an in vitro transcription (IVT) reaction. However, how promoter-adjacent DNA sequences influence RNA yield remains poorly characterized. Here, we present an integrated deep learning framework that leverages massively parallel next-generation sequencing (NGS) assays to measure RNA yield across large sequence spaces. A library of 10⁵ randomized oligonucleotide sequences was designed to systematically explore sequence diversity within a defined structural context. DNA and RNA abundances were quantified in parallel using Illumina sequencing, enabling high-resolution measurement of sequence-to-yield relationships at scale. Sequences were one-hot encoded and used to train deep learning models, using a convolutional neural network architecture. The model achieved a Pearson correlation of 0.94 between predicted and experimentally measured RNA yield on a held-out test set, demonstrating strong generalization across diverse sequence contexts. Importantly, the trained model can be deployed in a production environment to score and rank novel RNA sequence designs by predicted IVT yield, enabling cost-effective, pre-experimental prioritization of the most manufacturable candidates. This framework establishes a scalable, data-driven approach to DNA and RNA sequence optimization, with broad applicability to vaccine antigen design, therapeutic protein delivery, and synthetic biology. By integrating high-throughput experimentation with advanced deep learning modeling, it significantly reduces screening costs and accelerates RNA engineering cycle times.
Chair: Duccio Medini, Giuseppe Maccari
13:00 – 13:15PaperBIOINFMAP-Net: An Interpretable Deep Learning Framework for Universal mRNA ManufacturabilityMAP-Net: An Interpretable Deep Learning Framework for Universal mRNA Manufacturability
Giorgio CianoGiorgio Ciano
Click to view abstract
Abstract
The growth of mRNA therapeutics is limited by bespoke manufacturing processes. To overcome this barrier to access and innovation, we introduce an AI-driven framework that decouples sequence design from manufacturing, analogous to the universal design principles that revolutionized the semiconductor industry. We performed a large-scale screen, quantifying the in vitro transcription (IVT) efficiency of one million diverse sequences. We then trained an interpretable deep learning model that accurately predicts manufacturability from sequences alone and learned underlying molecular mechanisms. An algorithm using this model prospectively improved the IVT yield of a vaccine and a gene-editing therapeutic by over 7.5 fold. Our AI-driven framework establishes a universal design paradigm with the promise to democratize and accelerate the development of mRNA medicines, potentially unlocking a new era in biotechnology.
Chair: Duccio Medini, Giuseppe Maccari
Lunch
Terrazza del RettoratoComputational metagenomics to unravel person-to-person microbiome transmission
Francesco Asnicar
BIOINF Main Track 2
Aula MagnaChair: Graziano Pesole14:15 – 14:30PaperBIOINFSequence-Derived Physicochemical Features and Stacked Learning for SUMOylation-Site Prediction
Junyan Li, Qiaobin Yao, Danda Rawat, Jiang Li, Shaolei Teng
Click to view abstract
Abstract
SUMOylation regulates protein stability, localization, and interaction networks, but computational prediction of SUMO-acceptor lysines remains difficult because many true sites do not strictly follow the canonical Ψ-K-x-E motif. We developed a sequence-based prediction framework that systematically evaluates AAindex-derived physicochemical descriptors for human SUMOylation-site prediction. The dataset comprised 5,888 experimentally verified SUMOylated lysines and 415,918 unlabeled lysines from 176 human proteins collected from dbPTM and UniProt. For each candidate lysine in sumoylated proteins, we generated lysine-centered windows of multiple lengths and encoded each window with 566 AAindex properties. We first ranked individual AAindex features using single-feature XGBoost performance on the full dataset, then benchmarked several classifiers (logistic regression, SVM, random forest, and XGBoost) across window sizes. Based on these results, we trained a XGBoost model and a stacked ensemble (XGBoost + random forest + logistic regression with a logistic regression meta-learner) using a balanced training subset. The best setting used an 11 (±5 from center lysines)-residue window and ~350 top-ranked features. Under this configuration, the stacked ensemble achieved an AUC-ROC of 0.9426 on the test set, substantially outperforming the best single model. Motif-level analysis confirmed the presence of the canonical Ψ-K-x-E consensus, but the strongest predictive signals arose from flexible, coil-promoting sequence regions flanking the lysine. Overall, our results show that rich physicochemical encoding combined with stacking can markedly improve SUMOylation-site prediction and helps highlight property patterns that are most informative around sumoylated lysines.
14:30 – 14:45PaperBIOINFLLM-Guided Integer Linear Programming for Antibody Library Design
Mikel Landajuela Larma
Click to view abstract
Abstract
Designing a library of antibody variants requires choosing which mutations to introduce while balancing competing goals such as binding-related scores and human-likeness. We study whether a large language model (LLM) can help a combinatorial solver make better choices. In the proposed pipeline, the LLM reads a compact summary of the antibody--antigen structure (derived from a Protein Data Bank entry) together with per-mutation scores, and returns strictly structured hints: mutations to avoid, mutations to suggest, avoid-combinations, and additive preference scores. These hints are mechanically translated into linear constraints and objective coefficients of an integer linear program (ILP), which then returns a diverse batch of designs. The design decouples reasoning (handled by the LLM, which is good at high-level priors but bad at satisfying many interacting constraints at once) from search (handled by the ILP, which guarantees feasibility and can be solved to global optimality for each scalarization). We evaluate on two antibodies, trastuzumab and D44.1, and report multi-objective diversity and quality metrics; for trastuzumab we additionally report the success rate of a learned oracle classifier. LLM-guided ILP improves the strongest previous baseline on every reported metric while reliably delivering a full library of 1{,}000 unique sequences; direct LLM generation, in contrast, under-produces and collapses in diversity. Improvements are consistent across both antibody targets and across LLM backbones (gpt-4o, gpt-5), showing that the structural priors surfaced by the LLM translate directly into better designs. A key practical benefit of the approach is that those priors are emitted as strictly structured JSON, so domain experts can inspect, edit, or veto individual constraints before the ILP is re-run --- a level of auditability that end-to-end neural generators do not offer.
14:45 – 15:00PaperBIOINFCell Senescence Identification Using Single-Cell Transcriptomics Foundation Model
Yongqi Zhao, Fan Tong, Xiangwen Zheng, Yuanyuan Ma, Huajian Mao, Yuting Zhou, Peixiang Yang, Dongsheng Zhao
Click to view abstract
Abstract
Background: In recent years, single-cell foundation models have achieved good results in tasks such as cell type annotation, and how to use single-cell foundation models to identify senescent cells is a scientific question worth investigating. Data and Methods: This study selects 4 outstanding single-cell foundation models including scGPT, CellPLM, tGPT and UCE, combining multilayer perceptron as simple downstream networks, and uses 10 consensus senescence datasets e.g. GSE119807, GSE102090, and GSE94980 to train (80%) and test (20%) their performance on the task of cell senescence identification, preliminarily exploring the performance on their AUC evaluation. At the same time, using approaches such as decrease of training dataset, dimensionality reduction clustering analysis, and simplification of downstream models, the ability of the foundation model's cell embeddings to distinguish cellular ageing is evaluated. Result: The experimental results show that, except for a very few datasets, the constructed model generally achieved an AUC value of 99%, higher than the traditional domain models - SenCID (mean 95%) and hUSI (mean 92%), and comparable to DeepScence. When the training data is reduced to 60% or the downstream model is simplified to logistic regression, the AUC only slightly decreases (the mean AUC still exceeds 98%), and dimensionality reduction clustering analysis of embedding representations indicates that the foundation model has a strong ability to distinguish cell condition. Conclusion: The single-cell foundation model gains intrinsic knowledge related to cellular ageing through pre-training on massive sequencing data without expert feature engineering, providing a competitive new approach for research in cell senescence identification.
15:00 – 15:15PaperBIOINFOmicsGPT: Multi-Agent Bioinformatics Orchestration for Evidence-Grounded RNA-seq Interpretation
Patrick Roney, James Li
Click to view abstract
Abstract
Large language models can summarize oncology evidence, but unsupported stan- dalone use is vulnerable to hallucinated druggability, acronym collisions, and failure to respect tissue and assay context. We present OmicsGPT, a LangGraph-based multi-agent prototype that converts RNA-seq counts, metadata, optional context files, and a disease-specific prompt into an auditable oncology interpretation report. The workflow applies PyDESeq2/DESeq2-style differential-expression analysis and preranked GSEA, then routes selected targets through a planner-executor architecture querying PubMed, Open Targets, OncoKB, STRING, and GTEx- style tissue context. Role-specific agents flag lineage-mismatch artifacts, distinguish mutation- level from expression-level actionability, and filter acronym collisions before a report-writer agent synthesizes a source-traced report with evidence-bounded follow-up. As a public software- architecture test, we used GEO series GSE164641, an RNA-seq dataset profiling normal breast tissue from women at high or average risk for breast cancer. In a representative breast-risk run, the target roster included OLAH and BTN1A1; OmicsGPT treated these RNA-level find- ings as biomarker-hypothesis signals rather than direct therapeutic recommendations. Com- pared with a standalone LLM baseline, OmicsGPT preserved count-model statistics, exposed database provenance, separated RNA expression from DNA-level actionability, and retained weak or negative evidence when no direct claim was supported. The current system is intended for research interpretation rather than direct clinical decision support. Its main contribution is an interpretable glass-box architecture for evidence-grounded bioinformatics analysis and on- cology hypothesis generation.
15:45 – 16:00PaperBIOINFIdentifying functional drivers of Hepatoblastoma outcomes via agent-based modeling and transcriptomics
Alessandro Ravoni, Yuanhua Liu, Stefano Cairo, Filippo Castiglione, Christine Nardini
Click to view abstract
Abstract
Hepatoblastoma (HB) is the most common pediatric liver cancer and represents a major clinical challenge, due to the lack of effective therapies for advanced stages and disease relapse. In this work, we use the results of a previously HB-tailored agent-based model of the immune system to investigate whether model-derived variables can be of use in the prediction of patients’ outcomes. To this aim, we apply factor analysis to the results of a simulated cohort of HB patients, to identify combinations of key immunological variables able to discriminate disease outcomes in the simulator, and we then assess the coherence of such predictions with independent results of differential expression and enrichment analyses on HB transcriptomics. Our analysis proposes that the ability of immune cells, particularly natural killer and CD8+ cytotoxic T cells, to recognize tumor-associated antigens and exert cytotoxic activity is essential for disease control following treatment.
Keynote — Francesco Asnicar
Aula MagnaPoster Session A / Coffee Break
Museo dell'Arte ClassicaChair: Davide ChiccoELIXIR-IT: the Italian Research Infrastructure for Big Data in Life Sciences
Graziano Pesole
Welcome Cocktail + MUSA Jazz Concert
Terrazza del RettoratoChair: Annamaria CarissimoAI-assisted Single-Cell transcriptomics reveals persistent malignant CD4+ activity in Sézary Syndrome patient
Domenico Palumbo, Viola Melone, Luigi Palo, Carlo Ferravante, Giulia Salvatore, Cristina Cristofoletti, Maria Grazia Narducci, Roberta Tarallo
Click to view abstract
AI-assisted Single-Cell transcriptomics reveals persistent malignant CD4+ activity in Sézary Syndrome patient
Domenico Palumbo, Viola Melone, Luigi Palo, Carlo Ferravante, Giulia Salvatore, Cristina Cristofoletti, Maria Grazia Narducci, Roberta Tarallo
Abstract
Sézary syndrome (SS) is a rare and aggressive leukemic variant of cutaneous T-cell lymphoma characterized by erythroderma, lymphadenopathy, and circulating malignant CD4+ T cells. Despite several therapeutic advances such as extracorporeal photopheresis (ECP), the disease remains difficult to treat. Here, we applied single-cell RNA sequencing to investigate transcriptional changes in CD4+ T cells from an SS patient undergoing ECP at three timepoints (T1: 2 cycles, T2: 20 cycles, T3: 74 cycles). Exploiting the Illumina Single Cell 3’ RNA Prep Kit, we analyzed output data by using a custom pipeline with an AI-based annotation tool. Across all timepoints, CD4+ T cells were predominant, with an increasing CD4+/CD8+ ratio, indicating persistent disease spread. The AI sub-clustering identified a single active CD4+ population, which was selected and analyzed for differential gene expression. We identified 42, 82, and 151 differentially expressed genes in T1-T2, T2-T3, and T1-T3 comparisons respectively, indicating progressive transcriptional changes. Pathway analysis revealed a good activation of the inflammatory response, even under ECP therapy. These results suggest that, in this patient, ECPs were not enough to prevent transcriptional evolution of malignant CD4+ T cells in SS, showing an ongoing immune activation and disease progression. For this reason, a perturbation analysis was performed to highlight possible drugs to be used in combination with ECP to maximize its effect. This study demonstrates the utility of single-cell transcriptomics and AI-based annotation for monitoring therapeutic response in CD4+ cell population, causative of the pathological state, and understanding the SS pathophysiology.