Nicolò Lazzaro
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Nicolò Lazzaro has a total of 23 co-authors. Here are the top co-authors by number of shared publications:
- Gianluca Leonardi
Affiliations Data Science for Health, Fondazione Bruno Kessler, Trento, Italy, Department of Cellular, Computational and Integrative Biology, University of Trento, Trento, Italy
4 shared publications - Raffaele Marchesi
Affiliations Data Science for Health, Fondazione Bruno Kessler, Trento, Italy, Department of Mathematics, University of Pavia, Pavia, Italy
4 shared publications Flavio Ragni
Affiliation Data Science for Health, Fondazione Bruno Kessler, Trento, Italy
2 shared publicationsGiuseppe Jurman
Affiliations Data Science for Health, Fondazione Bruno Kessler, Trento, Italy, Department of Biomedical Sciences, Humanitas University, Milan, Italy
2 shared publications- Giuseppe Jurman
Humanitas University Milan Italy
2 shared publications - Marco Chierici
Fondazione Bruno Kessler
2 shared publications
Latest Publications
- Open Access
- 16 April 2026
- Published by Public Library of Science (PLoS) in PLOS Computational Biology
- Vol. 22 (4) , e1013455
Coherent cross-modal generation of synthetic biomedical data to advance multimodal precision medicine
Integration of multimodal, multi-omics data is critical for advancing precision medicine, yet its application is frequently limited by incomplete datasets where one or more modalities are missing. To address this challenge, we developed a generative framework capable of synthesizing any missing modality from an arbitrary subset of available modalities. We introduce Coherent Denoising, a novel ensemble-based generative diffusion method that aggregates predictions from multiple specialized, single-condition models and enforces consensus during the sampling process. We compare this approach against a multi-condition, generative model that uses a flexible masking strategy to handle arbitrary subsets of inputs. The results show that our architectures successfully generate high-fidelity data that preserve the complex biological signals required for downstream tasks. We demonstrate that the generated synthetic data can be used to maintain the performance of predictive models on incomplete patient profiles and can leverage counterfactual analysis to guide the prioritization of diagnostic tests. We validated the framework’s efficacy on a large-scale multimodal, multi-omics cohort from The Cancer Genome Atlas (TCGA) of over 10,000 samples spanning across 20 tumor types, using data modalities such as copy-number alterations (CNA), transcriptomics (RNA-Seq), proteomics (RPPA), and histopathology (WSI). This work establishes a robust and flexible generative framework to address sparsity in multimodal datasets, providing a key step toward improving precision oncology.
- 16 April 2026
- Published by Public Library of Science (PLoS) in PLOS Computational Biology
- Vol. 22 (4) , e1013455
Open AccessCoherent cross-modal generation of synthetic biomedical data to advance multimodal precision medicine
Integration of multimodal, multi-omics data is critical for advancing precision medicine, yet its application is frequently limited by incomplete datasets where one or more modalities are missing. To address this challenge, we developed a generative framework capable of synthesizing any missing modality from an arbitrary subset of available modalities. We introduce Coherent Denoising, a novel ensemble-based generative diffusion method that aggregates predictions from multiple specialized, single-condition models and enforces consensus during the sampling process. We compare this approach against a multi-condition, generative model that uses a flexible masking strategy to handle arbitrary subsets of inputs. The results show that our architectures successfully generate high-fidelity data that preserve the complex biological signals required for downstream tasks. We demonstrate that the generated synthetic data can be used to maintain the performance of predictive models on incomplete patient profiles and can leverage counterfactual analysis to guide the prioritization of diagnostic tests. We validated the framework’s efficacy on a large-scale multimodal, multi-omics cohort from The Cancer Genome Atlas (TCGA) of over 10,000 samples spanning across 20 tumor types, using data modalities such as copy-number alterations (CNA), transcriptomics (RNA-Seq), proteomics (RPPA), and histopathology (WSI). This work establishes a robust and flexible generative framework to address sparsity in multimodal datasets, providing a key step toward improving precision oncology.
- Open Access
Coherent Cross-modal Generation of Synthetic Biomedical Data to Advance Multimodal Precision Medicine
Integration of multimodal, multi-omics data is critical for advancing precision medicine, yet its application is frequently limited by incomplete datasets where one or more modalities are missing. To address this challenge, we developed a generative framework capable of synthesizing any missing modality from an arbitrary subset of available modalities. We introduce Coherent Denoising, a novel ensemble-based generative diffusion method that aggregates predictions from multiple specialized, single-condition models and enforces consensus during the sampling process. We compare this approach against a multi-condition, generative model that uses a flexible masking strategy to handle arbitrary subsets of inputs. The results show that our architectures successfully generate high-fidelity data that preserve the complex biological signals required for downstream tasks. We demonstrate that the generated synthetic data can be used to maintain the performance of predictive models on incomplete patient profiles and can leverage counterfactual analysis to guide the prioritization of diagnostic tests. We validated the framework’s efficacy on a large-scale multimodal, multi-omics cohort from The Cancer Genome Atlas (TCGA) of over 10,000 samples spanning across 20 tumor types, using data modalities such as copy-number alterations (CNA), transcriptomics (RNA-Seq), proteomics (RPPA), and histopathology (WSI). This work establishes a robust and flexible generative framework to address sparsity in multimodal datasets, providing a key step toward improving precision oncology. Author Summary: To make precision medicine a reality, doctors need to understand a patient’s status from many angles, using different data types like genetic information (omics) and tissue slide images (histopathology). The problem is that most patient records are incomplete, with one or more of these data types missing, which can limit the effectiveness of powerful predictive tools. We have built a generative AI system designed to learn the complex biological patterns that connect all these different data types. By looking at the patient data that is available, our system can then generate a realistic, synthetic version of any missing piece. We developed a novel method called Coherent Denoising to do this, which is flexible and helps protect patient privacy. We validated this approach on a large dataset of over 10,000 cancer patient profiles. We show that our AI-generated data is high-fidelity and can successfully complete these sparse patient profiles, allowing AI models for crucial tasks like cancer staging and survival prediction to work at their best even with incomplete patient data. We also demonstrate how this tool can be used to evaluate the potential impact of new tests, helping to prioritize which expensive diagnostic tests would be most beneficial for a patient.
Open AccessCoherent Cross-modal Generation of Synthetic Biomedical Data to Advance Multimodal Precision Medicine
Integration of multimodal, multi-omics data is critical for advancing precision medicine, yet its application is frequently limited by incomplete datasets where one or more modalities are missing. To address this challenge, we developed a generative framework capable of synthesizing any missing modality from an arbitrary subset of available modalities. We introduce Coherent Denoising, a novel ensemble-based generative diffusion method that aggregates predictions from multiple specialized, single-condition models and enforces consensus during the sampling process. We compare this approach against a multi-condition, generative model that uses a flexible masking strategy to handle arbitrary subsets of inputs. The results show that our architectures successfully generate high-fidelity data that preserve the complex biological signals required for downstream tasks. We demonstrate that the generated synthetic data can be used to maintain the performance of predictive models on incomplete patient profiles and can leverage counterfactual analysis to guide the prioritization of diagnostic tests. We validated the framework’s efficacy on a large-scale multimodal, multi-omics cohort from The Cancer Genome Atlas (TCGA) of over 10,000 samples spanning across 20 tumor types, using data modalities such as copy-number alterations (CNA), transcriptomics (RNA-Seq), proteomics (RPPA), and histopathology (WSI). This work establishes a robust and flexible generative framework to address sparsity in multimodal datasets, providing a key step toward improving precision oncology. Author Summary: To make precision medicine a reality, doctors need to understand a patient’s status from many angles, using different data types like genetic information (omics) and tissue slide images (histopathology). The problem is that most patient records are incomplete, with one or more of these data types missing, which can limit the effectiveness of powerful predictive tools. We have built a generative AI system designed to learn the complex biological patterns that connect all these different data types. By looking at the patient data that is available, our system can then generate a realistic, synthetic version of any missing piece. We developed a novel method called Coherent Denoising to do this, which is flexible and helps protect patient privacy. We validated this approach on a large dataset of over 10,000 cancer patient profiles. We show that our AI-generated data is high-fidelity and can successfully complete these sparse patient profiles, allowing AI models for crucial tasks like cancer staging and survival prediction to work at their best even with incomplete patient data. We also demonstrate how this tool can be used to evaluate the potential impact of new tests, helping to prioritize which expensive diagnostic tests would be most beneficial for a patient.
- Open Access
Toward Reliable Synthetic Omics: Statistical Distances for Generative Models Evaluation
Background: Synthetic data generation is emerging as an approach to overcome the limitations of real-world data scarcity in omics studies, especially in precision medicine and oncology. Omics datasets, with their high dimensionality and relatively small sample sizes, often lead to overfitting, especially in deep learning models. Generative models offer a promising way to generate realistic synthetic data preserving the original data distribution. However, there is still no objective consensus on how to evaluate their performance. In this study, we set out to validate generative networks for transcriptomics data generation by using statistical distances as robust evaluation metrics. Results: We observe that statistical distances enable simultaneous evaluation of global and local data fidelity of generated synthetic data. Because these distances satisfy the properties of true metrics, they also enable formal hypothesis testing to assess whether generative models have in fact converged or are merely approaching the reference distribution. Crucially, optimizing for these distances was found to implicitly select models maximizing other widely used metrics of generative performance, providing evidence of their broad applicability. Overall, our findings indicate that the adoption of these metrics can play a key role in guiding the development of generative models across a wide range of domains.
Open AccessToward Reliable Synthetic Omics: Statistical Distances for Generative Models Evaluation
Background: Synthetic data generation is emerging as an approach to overcome the limitations of real-world data scarcity in omics studies, especially in precision medicine and oncology. Omics datasets, with their high dimensionality and relatively small sample sizes, often lead to overfitting, especially in deep learning models. Generative models offer a promising way to generate realistic synthetic data preserving the original data distribution. However, there is still no objective consensus on how to evaluate their performance. In this study, we set out to validate generative networks for transcriptomics data generation by using statistical distances as robust evaluation metrics. Results: We observe that statistical distances enable simultaneous evaluation of global and local data fidelity of generated synthetic data. Because these distances satisfy the properties of true metrics, they also enable formal hypothesis testing to assess whether generative models have in fact converged or are merely approaching the reference distribution. Crucially, optimizing for these distances was found to implicitly select models maximizing other widely used metrics of generative performance, providing evidence of their broad applicability. Overall, our findings indicate that the adoption of these metrics can play a key role in guiding the development of generative models across a wide range of domains.
- Open Access
Signature Distance: Generalizing Energy Statistics
Comparing empirical distributions is central to generative model evaluation, hypothesis testing and data augmentation in high-dimensional biological data. Established methods such as energy distance summarize each point's relationship to the opposing distribution through a single expected distance, providing sensitivity to location shifts but not to local density or topological structure. We introduce Signature Distance (SD), a metric that compares empirical distributions through the mean absolute difference of their sorted pointwise distance profiles. SD is a structural generalization of energy distance and matches its $\mathcal{O}(n^2)$ computational complexity. On TCGA pan-cancer transcriptomic data, we show that (1) SD detects density changes that energy distance is insensitive to; (2) the per-point SD loss landscape reveals the geometric mechanisms behind known limitations of energy distance as a generative objective; (3) linearly interpolated biological samples that are not detected by energy distance are correctly penalized by SD; (4) SD provides a direct differentiable potential energy for model-free Langevin data expansion, with a bootstrap resampling protocol that stabilises the stopping epoch; and (5) SD is directly usable as a differentiable generative training loss.Open AccessSignature Distance: Generalizing Energy Statistics
Comparing empirical distributions is central to generative model evaluation, hypothesis testing and data augmentation in high-dimensional biological data. Established methods such as energy distance summarize each point's relationship to the opposing distribution through a single expected distance, providing sensitivity to location shifts but not to local density or topological structure. We introduce Signature Distance (SD), a metric that compares empirical distributions through the mean absolute difference of their sorted pointwise distance profiles. SD is a structural generalization of energy distance and matches its $\mathcal{O}(n^2)$ computational complexity. On TCGA pan-cancer transcriptomic data, we show that (1) SD detects density changes that energy distance is insensitive to; (2) the per-point SD loss landscape reveals the geometric mechanisms behind known limitations of energy distance as a generative objective; (3) linearly interpolated biological samples that are not detected by energy distance are correctly penalized by SD; (4) SD provides a direct differentiable potential energy for model-free Langevin data expansion, with a bootstrap resampling protocol that stabilises the stopping epoch; and (5) SD is directly usable as a differentiable generative training loss.