Workshop Program


October, 10
October, 11

09:30 Welcome

slot

10:00 Neuroscience Session

10:00 Life Sciences Session

Coffee break

Coffee break

Lunch break

Marianna La Rocca

Scale-driven Methodology for Analyzing and Recognizing Post-Traumatic Epilepsy alterations

An estimated 10 million individuals experience a traumatic brain injury (TBI) each year and up to 53% of TBI subjects develop post-traumatic epilepsy (PTE) over the course of their lifetimes. PTE is diagnosed if one or more unprovoked seizures occur at least one week after TBI. The trauma variability and the complexity of the PTE mechanism make it challenging to study biomarkers and identify at-risk subjects for the experimentation of anti-epileptogenic therapies. The Epilepsy Bioinformatics Study for Antiepileptogenic Therapy (EpiBioS4Rx) is an international multi-center project that is collecting many high quality multimodal data with the aim to understand epileptogenic mechanisms and discover PTE biomarkers. Recent studies have shown that graph theory is a valid mathematical tool to detect changes due to neurodegenerative diseases. The graph theory formulation describes the brain as a set of nodes (i.e. brain regions) linked by edges (i.e. their functional connectivity or their structural similarity). The limit of these approaches is the inability to study what happens to the same nodes as their interactions change. Multiplex networks are innovative tools that can overcome this limit. In this talk, I will present how I and my research team are going to develop a novel multiplex network approach to manage multi-scale, multimodal and longitudinal data of EpiBioS4Rx patients to mine important features that can be modeled from artificial intelligence to predict seizure onset in TBI patients. Our aim is to achieve three main points that extend beyond the scope of EpiBioS4Rx: (i) multiplex network combined with machine learning techniques can be a powerful instrument to predict seizure onset after TBI; (ii) multiplex networks are able to model the relationships among multimodal and longitudinal data and allow the detection of subtle changes related to PTE that standard methods cannot; and (iii) multiplex networks make it possible to find the anatomical regions related to the PTE onset.

Dominique Duncan

Navigating the complexity of post-traumatic epilepsy using AI tools

The Epilepsy Bioinformatics Study for Antiepileptogenic Therapy (EpiBioS4Rx) is a multi-site, international collaboration that integrates data from human studies and animal models to better understand post-traumatic epilepsy (PTE) following traumatic brain injury (TBI). Epileptogenesis is a complex, multifactorial process, influenced by diverse modalities such as MRI, EEG, and molecular biomarkers. The challenge lies not only in the biological intricacies of this condition but also in the vast amounts of heterogeneous data generated across global research centers. Artificial Intelligence (AI) is needed to uncover new insights within this complexity. By developing and applying advanced machine learning methods, we aim to identify novel biomarkers of epileptogenesis across imaging, electrophysiology, and molecular data. Our centralized data platform standardizes and harmonizes these data, enabling seamless integration and analysis. Through this platform, researchers worldwide can access shared datasets, AI-driven tools, and sophisticated analytics to deepen our understanding of the mechanisms behind PTE, ultimately paving the way toward effective therapeutic interventions. This talk will explore how AI is revolutionizing epilepsy research, from data standardization to biomarker discovery.

Daniele Marinazzo

Higher-order behaviors in complex systems: insights from information decomposition

Systems composed of many units, whose behavior goes beyond the sum of the individual behaviors of the singles, are ubiquitous. Examples relevant to what we do are the brain, the body as a whole, and the social systems we live in. When it comes to analyzing collective behavior we are often stuck with pairwise dependencies (often correlations). In this talk, I will describe a framework rooted in information theory to mine multiplets of variables sharing common information about the variability of complex systems, and provide some examples in neuroscience, physiology, machine learning, and psychometrics.

References: [1]

Marina De Tommaso

The Exploration of EEG Connectivity in Chronic Pain: Clinical Significance in Migraine and Fibromyalgia

Abstract

Loredana Bellantuono

Liquidity of gene co-expression networks highlight the role of the perinatal GABA switch in genetic risk for schizophrenia

Network approaches have demonstrated utility in neuroscience. However, disorders with genetic and neurodevelopmental components, such as psychiatric conditions, require dynamic approaches to account for network changes across the lifespan. This work is focused on a complex network approach to analyze the time evolution of gene co-expression patterns in different types of brain tissue, comparing the behavior of genes related to different multifactorial diseases, including schizophrenia (SCZ), Alzheimer’s disease (AD), and major depression disorder (MDD), that occur as a result of many genomic variants.

Gene co-expression networks were constructed using a dataset of 800 postmortem brains from neurotypical individuals collected at the Lieber Institute for Brain Development. The tendency of genes to vary their co-expression relations in time was quantified by the liquidity metric. Results highlight a liquidity hierarchy, in which risk genes tend to have significantly less volatile co-expression patterns than not-risk genes, only for specific brain tissues and disorders. Actually, in the case of SCZ, the greater time stability of connections involving risk genes compared to those involving not-risk genes emerges much more markedly in the co-expression networks from the dorsolateral prefrontal cortex than in those obtained from other tissues. Similarly, the study highlights the well-established role of the hippocampus in the context of AD since, in this tissue, the difference between the liquidity distributions of risk and not-risk genes is much more significant compared to the one obtained for networks related to other tissues.

Moreover, we analyzed gene co-expression trajectories across the lifespan in neurotypical individuals with relatively high and low polygenic risk scores. Our findings revealed increased co-expression stability across the lifespan in the dorsolateral prefrontal cortex for subjects with higher genetic risk for SCZ, indicating altered GABAergic regulation and delayed GABA-A receptor maturation. These insights underscore GABA-signaling disruptions as critical contributions to translating SCZ risk into altered developmental trajectories.

Pierluigi Selvaggi

Biomarkers for mental health

A considerable amount of individuals with severe mental health disorders such as Schizophrenia, Bipolar Disorder, and Major Depression, do not respond to conventional pharmacological treatment. Unfortunately, despite the efforts over the last few years, psychiatry still does not have biological predictors (i.e. biomarkers) that could be used in clinical practice to help clinicians in delivering tailored treatment to patients. Neuroimaging techniques have been largely exploited for this task, with difficulties in making accurate predictions at the individual level. In this talk I will present the state-of-the-art of biomarkers research in psychiatry and discuss the challenges ahead. In addition, I will present novel computational approaches that, applied to neuroimaging data, might help to better parse heterogeneity that charachterise patients with mental health disorders to deliver effective biomarkers in psychiatry.

Giuseppe Magnifico

Introduction to Tensor Network methods and Quantum Computing

In the rapidly evolving landscape of modern physics, quantum technologies promise to revolutionize computation opening new possibilities for solving complex problems. This talk aims to provide an introduction to the world of tensor networks and quantum computing.
We will begin by exploring the core concepts behind these methods and technologies. In particular, we will focus on tensor networks, a powerful quantum-inspired computational framework essential for simulating and understanding complex quantum systems and quantum circuits. We will discuss the fundamental structure of tensor networks, their role in efficiently representing quantum states, and their application in quantum information, machine learning.

Flavia Esposito

Low-rank approximation methods for real data analysis and integration

Over the years, low-rank approximation models have gained significant attention due to their effectiveness in analyzing real data.The key idea is that real data has a structured form (such as vectors, matrices, or tensors) and admits a low-rank representation. A data matrix X ∈ Rn x m, with n samples and m features, can be represented as a product of two factors W ∈ Rn x r and H ∈ Rr x m, with r < min(m, n), such that X ≈ WH.

The problem of finding such a pair (W,H) can be mathematically formulated as a penalized optimization task:

minW,H∈C Div(X,WH) + μ1J1(W) + μ2J2(H) + μ3J3(W,H)

where Div(·,·) is a divergence function that evaluates the quality of the approximation, C is a feasible set that encodes structural or physical information about the data, Ji are the penalty functions that enforce additional properties on W and H, and μi are the penalty hyperparameters, balancing the bias-variance tradeoff in approximating X and satisfying factor properties.

In this talk, we review some theoretical and computational issues related to specific low-rank approximation models and numerical methods defined on the set C of nonnegative matrices.

We address several mathematical challenges, including the selection of an appropriate divergence function tailored to the specific data domain, and the proper definition of Ji to integrate domain-specific prior knowledge. We also emphasize real-world applications, particularly in the biomedical and environmental fields. Morover, we also investigate how additional constraints encoded by the peculiar form of Ji can be advantageously handled using manifold optimization techniques.

References: [1], [2]

Samuele Cavinato

What happens when radiotherapy meets quantum computing?

Quantum computing promises unprecedented advantages in solving complex optimization problems, and its application to radiotherapy could lead to significant improvements in efficiency. In fact, modern intensity-modulated radiotherapy (IMRT) techniques, including fixed-gantry IMRT, volumetric modulated arc therapy (VMAT), and Helical Tomotherapy, enable the precise delivery of radiation doses with high conformity to the target, while simultaneously sparing surrounding healthy tissues and reducing the risk of side effects. However, implementing IMRT treatments requires solving an optimization problem known as inverse-planning, which typically involves optimizing cost functions with hundreds to thousands of variables to achieve the optimal modulation of radiation beams’ fluence and meet a set of dose objectives.

In this talk, I will present recent collaborative work between the Medical Physics Department of the Veneto Institute of Oncology (IOV-IRCCS) and the Department of Physics and Astronomy of the University of Padua. Our work focused on developing and validating novel strategies to adapt classical beamlet-fluence optimization problems to meet the characteristics of quantum hardware. I will discuss some of the current challenges in this area and the solutions we have developed to address them. Finally, I will present the results of two proof-of-concept studies in which these strategies were combined with Tensor Network Methods and applied to solve two distinct inverse-planning problems for prostate cancer treatment.

Timo Felser

A view on Tensor Networks from simulating quantum systems to explainable AI

In this talk, we will present the use of Tensor Networks (TN) as novel approach to explainable AI. TNs are established since over 3 decades in particular for data representation to simulate quantum systems on classical computers. They were developed specifically with the aim to offer direct accessibility to key physical information to understand complex systems and are state-of-the-art methods in the development of quantum computers and quantum machine learning.

At Tensor AI Solutions, we transferred this method to explainable AI systems. Training machine learning (ML)-models based on TNs allows us to learn highly complex patterns from data while still maintaining access to key information learned by the ML-model, such as learned feature relevance, learned correlations, or feature contributions to model decisions. In doing so, we can assess the learned patterns when testing the AI system and, in addition, tune the model performance after learning. In real-time applications, we can control the accuracy with respect to the prediction time without the need for retraining.

Over the last years, we applied TN-based machine learning to a multitude of data-based problems from research and industry to demonstrate its practicality, robustness and transparency advantages in challenging, real-world scenarios. The use-cases entail: high-risk systems, such as credit scoring where unbiased decision-making in each single prediction is compliance-critical; real-time systems, such as particle classification at the large hadron collider in CERN, and reliability-critical systems, such as autonomous driving, where (in both applications) deeper understanding of the data patterns used by AI models is critical for understanding the physics behind the data, improving the efficiency of AI models and enabling trustworthy real-time predictions.

Marco Trenti

Tensor Network Machine Learning and xAI in Industrial Applications

Tensor Network (TN) methods, renowned for their powerful applications in quantum computations and simulations, also serve as a versatile tool for optimal information processing. Tensor Network Machine Learning is an emerging field at the intersection of physical sciences, high-performance computing, and data science, promising a future-proof learning paradigm. TNs have been successfully applied to both supervised and unsupervised machine learning tasks, offering a linear, transparent, and explainable model without sacrificing expressivity. In industrial applications, the explainability of models, not only enhances performance, resiliency, and robustness but crucially ensures compliance with the forthcoming European AI Act, which restricts the use of opaque models like neural networks.

Domenico Pomarico

Grokking as an entanglement transition during training dynamics of MPS machine learning

Generalizability is a fundamental property for machine learning algorithms, detected by a grokking transition during training dynamics. In the quantum-inspired machine learning framework we numerically prove that a quantum many-body system shows an entanglement transition corresponding to a performances improvement in binary classification of unseen data. Two datasets are considered as use case scenarios, namely fashion MNIST and genes expression communities of hepatocellular carcinoma. The measurement of qubits magnetization and correlations is included in the matrix product state (MPS) simulation, in order to define meaningful genes subcommunities, verified by means of enrichment procedures.

Sebastiano Stramaglia

Higher-order interactions in living systems

Higher-order interactions are well-characterized in biology, including protein complex formation and feedback or feedforward loops. These higher-order relationships are better represented by a hypergraph as a generalized network model. I will describe a novel approach which enables the quantification of the relative importance of high-order effects compared to pure two-body effects in information transfer between two processes, while also highlighting the processes that contribute to building these high-order effects alongside the driver.

References: [1]

Maria Colomba Comes

Revolutionizing Oncology: The Synergy between Digital Pathology and Machine Learning

Digital pathology (DP), a revolutionary approach in the field of pathology, leverages digital imaging and advanced computing techniques to transform traditional slide-based analyses into high-resolution, easily accessible digital formats, known as whole slide images (WSIs). When combined with machine learning (ML), these images can be analysed with a scale and precision beyond human capability. This powerful integration holds great promise for the diagnosis and prognosis of diseases, including both solid tumors and hematologic disorders. 

Our research group is actively investigating DP-ML application in three cancer research areas: melanoma, breast cancer and myeloproliferative neoplasms (MPNs). Within the melanoma setting, we developed a ML-based model exploiting WSIs to predict recurrence in negative sentinel lymph-node melanoma patients. For breast cancer, we explored the DP-ML combination capabilities to early predict response to neoadjuvant chemotherapy. In the field of MPNs, we applied ML to analyze WSIs representing bone marrow biopsies to provide a detailed tissue analysis for MPN diagnosis. 

Finally, these approaches could identify subtle patterns and anomalies, enabling earlier and more accurate diagnoses, personalized treatment plans, and better patient outcomes. Hence, DP-ML integration not only enhances the efficiency of oncological practices but also opens new avenues for research and innovation in cancer care. 

Roberto Cazzolla Gatti

Artificial Intelligence to model complex systems in One Health

In an epoch of global challenges, interdisciplinary scientists are requested to explore the relationships between Species, Ecosystems, and Human Health (One Health) with the help of Artificial Intelligence (A.I.), which can provide great support to research fostering the understanding of patterns and dynamics among big environmental and biomedical data. With the help of A.I. there we can foster outstanding progress in scientific discoveries that will benefit the whole planet, improving the health of our and other species, ecosystems, and ultimately of Earth systems. In this talk, I will provide some examples of research studies that bring together several human intelligences from the fields of ecology, conservation biology, biomedicine, physics, informatics, veterinary, etc. and merge them with artificial intelligence to speed up the advancement of science using the most advanced technologies to answer cutting-edge questions. Moreover, I will present the activities of the newly established One Health Artificial Intelligence (OHAI) Research Hub, which is exploring the interactions between at least two of the three Health domains (i.e. Species, Ecosystems, and Human) and is supported by Artificial Intelligence.

Raphaël Mourad

Machine Learning for Genomics and Genetics

Predicting molecular processes using deep learning is a promising approach to provide biological insights for non-coding SNPs identified in genome-wide association studies. However, most deep learning methods rely on supervised learning, which requires DNA sequences associated with functional data, and whose amount is severely limited by the finite size of the human genome. Conversely, the amount of mammalian DNA sequences is growing exponentially due to ongoing large-scale sequencing projects, but in most cases without functional data. To alleviate the limitations of supervised learning, we propose a paradigm shift with semi-supervised learning, which does not only exploit labeled sequences (e.g. human genome with ChIP-seq experiment), but also unlabeled sequences available in much larger amounts (e.g. from other species without ChIP-seq experiment, such as chimpanzee). We present two algorithms: DeepGNN (BMC Bioinformatics 2023) and DeepSSL (Briefings in Bioinfo, in revision). Our approach is flexible and can be plugged into any neural architecture including shallow and deep networks, and shows strong predictive performance improvements compared to supervised learning in most cases. Moreover, small models trained by DeepSSL showed similar or better performance than large language model DNABERT2.

Ester Pantaleo

Crowdsourced prediction of preterm birth with non invasive microbiome analysis

The microbiome, or the collection of all microbes that naturally live on our bodies and inside us, strongly contributes to human health and wellness. In this talk we will present the contribution of our group to the Microbiome preterm birth DREAM Challenge, a crowdsourcing approach to advance preterm birth research organised by the University of California San Francisco. For this challenge we proposed an ensemble machine learning approach based on Random Forests and oversampling of the minority class to predict preterm and early preterm birth, the latter scoring second over 121 submissions.

Andrea Astolfi

Decoding the Protein Folding Complexity to Design Folding Interfering Degraders: The Prion Protein Case Study

Years of research have established the cellular prion protein (PrPC) as a promising pharmaceutical target for prion diseases, fatal neurodegenerative disorders caused by the transformation of physiological PrPC into a misfolded, infectious isoform known as PrP scrapie (PrPSc). [1] 

Various strategies have been proposed over time to address this target, including traditional drug discovery approaches like identifying small molecules that can relocate PrPC from the cellular membrane to intracellular endosomes, and binders that prevent its conversion to PrPSc. However, no therapy is currently available, and prion disease remains an unmet medical need. [2] 

Recently, we have applied a novel drug discovery approach devoted to lowering PrPC levels by hampering a complete folding process. We refer to this strategy as Pharmacological Protein Inactivation by Folding Intermediate Targeting (PPI-FIT). [3] 

Our study identified a metastable intermediate of the PrP folding pathway with a druggable pocket through all-atoms MD simulation. Virtual screening of a commercial small molecule library led to the identification of potential binders, four of which could selectively reduce PrP load on the cellular membrane and promote its degradation. 

The most promising compound, SM875, was able to reduce PrP loads in various cell lines without decreasing PrP mRNA, selectively promote the degradation of nascent PrP molecules by the autophagy-lysosomal pathway, and exhibit dose-dependent anti-prion activity in prion-infected mouse fibroblasts. These findings strongly suggest that SM875 acts by targeting a folding intermediate of PrP, opening the door to hit-to-lead optimization efforts.

Lydia Siragusa

In Silico approaches based on Molecular Interaction Fields: a tour through algorithms, applications and case studies

The availability of a huge amount of protein data deposited in Protein Data Bank offers a unique occasion to explore the mechanism beyond pharmacology. Furtheremore, in silico methods allow to expand the knownledge from a structural standpoint. The structural comparison of protein binding sites is increasingly important in drug design; identifying structurally similar sites can be useful for techniques such as drug repurposing, and also in a polypharmacological approach to deliberately affect multiple targets in a disease pathway, or to explain unwanted off target effects. Once similar sites are identified, identifying local differences can aid in the design of selectivity. Such an approach moves away from the classical “one target one drug” approach and toward a wider systems biology paradigm. We developed a semiautomated approach, called BioGPS [1], that is based on GRID Molecular Interactions Fields (MIFs) [2] and pharmacophoric fingerprints. BioGPS comprises the automatic preparation of protein structure data, identification of binding sites, and subsequent comparison by aligning the sites and directly comparing the MIFs. Chemometric approaches are included to reduce the complexity of the resulting data on large dataset, enabling focus on the most relevant information. A database of about 900K proteome cavities, enabling large scale protein-protein and ligandprotein virtual screening, is integrated in BioGPS. Here we will focus on 2 examples of great interest in the current pharmaceutical research scenario: 1. ELIOT, the E3 ligases pocketome [3], aiming to find new ligases capable of hijacking the UPS system, to extend the applicability of PROTAC molecules, and 2. CROMATIC, the Coronaviruses proteins pocketome [4], to expand the chemical variability of scaffolds for pan-coronavirus antiviral drugs.

Daniela Trisciuzzi

GRID-based strategies to spot drug-like binding pockets of protein interaction systems

The dysfunction of protein interaction systems is acknowledged as responsible of the onset of severe pathologies such as cancer and neurodegeneration. In this respect, looking for putative druggable cavities can be a viable strategy for understanding how to modulate such complex biological systems. Based on an in house non-redundant collection of high-quality 3D crystallographic structures, we have firstly explored the property space of peptide-protein complexes by generating interpretable molecular interaction fields. The energetic distributions of the most frequent peptide-protein interface residue pairs have been analysed in order to understand the extent to which peptides can interfere with protein interaction systems [1]. A k-medoid clustering available from BioGPS software [2] has thus been employed to spot interacting druggable areas on the peptide-protein interface. Our model proved to be highly predictive and could be also employed to run peptide-protein and protein-protein virtual screening campaigns [3]. 

Further investigations are now in progress for featuring key protein pockets in the protein-protein networks; this is of great importance to better understand the rationale behind the molecular mechanisms of tumorigenesis with the aim of developing more effective diagnosis, prognosis, and treatment strategies. 

Nicola Gambacorta

Web-platforms based on Explainable Artificial Intelligence for Comprehensive Health Assessments

In our pursuit of innovative tools to address rare diseases and predict toxicological outcomes, we have developed three distinct public platforms: PLATO [1], TISBE [2] , and CIRCE [3] . PLATO (Polypharmacology pLATform for predictiOn) is a ligand-based predictive platform focused on polypharmacology. It is designed to facilitate target identification and bioactivity prediction, with the dual aim of finding potential protein drug targets and calculating bioactivity affinity values. Using a multi-fingerprint similarity search algorithm, PLATO is particularly effective for reverse screening in drug repurposing, making it a valuable resource for identifying potential treatments for rare diseases. TISBE (TIRESIA Improved on Structure-Based Explainability) is a powerful platform for predicting developmental toxicity, a key factor in protecting maternal and child health. TISBE offers four major innovations: a comprehensive manually curated dataset, a transparent explainable AI (XAI) framework using fragment-based fingerprints, a novel consensus classifier combining five independent machine learning models, and a new applicability domain method with a dual top-down approach for improved accuracy. CIRCE (Cannabinoid Iterative Revaluation for Classification and Explainability) employs a multi-layer machine learning framework to predict selective and unselective CB1/CB2 cannabinoid receptor binders. Shapley values [4]  are used to explain model predictions, enabling insights into feature importance. CIRCE aids in the intuitive design of CB1/CB2 binders and supports drug repurposing efforts.

Nicola Amoroso

Chemical space networks: a novel perspective for in silica toxicology

Recent works demonstrated how Chemical Space Networks (CSNs) can unveil meaningful chemical patterns extremely useful for drug design and model toxicological endpoints, just to mention a few examples. Here, CSNs are shown to effectively characterize the toxicity of chemicals towards several human endpoints, such as developmental toxicity, hepatoxicity, carcinogenicity, mutagenicity, androgenicity, estrogenicity, chromosomal aberrations and skin irritation. Thanks to CSNs it is possible to learn enhanced embeddings which, for all the considered endpoints, allow an accurate and robust distinction of toxic from non-toxic chemicals. In fact, using embeddings registered, on average, an increment of all considered metrics. Particularly, an increase of +12% was detected for the area under the curve ROC, which takes into account the prediction of toxic compounds. Moreover, through a dedicated eXplainable Artificial Intelligence (XAI) framework, a direct interpretation of results is provided. Hence, the proposed approach paves the way to the improvement of state-of-the-art in silico classification systems for toxicological endpoints and, more importantly, it provides novel insights about toxicophores which could lead to breakthrough innovations for the design of safer chemicals and drugs.