The integration of deep learning (DL) with multi-omics data has significantly advanced our understanding of biological systems, particularly in cancer research. DL enables the analysis of high-dimensional datasets and the discovery of novel disease mechanisms and biomarkers, contributing to improved patient treatment and management. This review provides a detailed overview of recent developments in deep learning models applied to genomics data, with a focus on cancer type classification, driver gene identification, survival analysis, and drug response prediction. We introduce the foundational concepts of machine and deep learning and explain the characteristics of multi-omics data, addressing a broad and interdisciplinary audience. Methods published since 2020 are systematically reviewed, including their model architectures, datasets, and key innovations.

Data sharing among different institutions represents one of the major challenges in developing distributed machine learning approaches, especially when data is sensitive, such as in medical applications. Federated learning is a possible solution, but requires fast communications and flawless security. Here, we propose SYNDSURV (SYNthetic Distributed SURVival), an alternative approach that simplifies the current state-of-the-art paradigm by allowing different centres to generate local simulated instances from real data and then gather them into a centralised hub, where an Artificial Intelligence (AI) model can learn in a standard way. The main advantage of this procedure is that it is model-agnostic, therefore prediction models can be directly applied in distributed applications without requiring particular adaptations as the current federated approaches do. To show the validity of our approach for medical applications, we tested it on a survival analysis task, offering a viable alternative to train AI models on distributed data. While federated learning has been mainly optimised for gradient-based approaches so far, our framework works with any predictive method, proving to be a comparable way of performing distributed learning without being too demanding towards each participating institute in terms of infrastructural requirements.

PURPOSE: Chronic myelomonocytic leukemia (CMML) is a rare myeloid neoplasm characterized by clinical heterogeneity and is associated with poor outcomes. To date, limited molecular information has been incorporated into disease classification and risk stratification. We aimed to integrate genomic features into the clinical decision-making process for CMML. PATIENTS AND METHODS: We analyzed a retrospective cohort of 3013 patients with CMML (training set) and a prospective population of 516 patients (validation set). Using an innovative framework for multimodal data analysis, we developed molecular-based disease taxonomy and prognostication. RESULTS: Unsupervised clustering identified nine entities with distinct genomic features and outcomes ( P < .001), including splicing machinery, transcription factors, signal transduction and tyrosine kinase pathways aberrations, and high-risk molecular signatures. Notably, 15% of patients showed molecular/clinical overlap with other myeloid neoplasms. We integrated molecular and clinical information to build the international CMML Prognostic Scoring System (iCPSS), incorporating mutations in nine genes together with hematologic parameters and cytogenetic abnormalities. The iCPSS identified five groups with distinct probability of overall and leukemia-free survival in both training and validation cohorts ( P < .001), outperforming existing prognostic models. Importantly, 55% of patients were reassigned to higher or lower risk groups by the iCPSS. Decision analysis demonstrated that iCPSS could refine the optimal timing of allogeneic transplantation at the individual level; compared with conventional prognostic tools, iCPSS-based decision modeling changed transplantation strategy in 31% of cases, resulting in a significant gain-in-life expectancy for eligible patient population ( P < .001). A federated learning platform was implemented to enable continuous, privacy-preserving model update across multiple centers. CONCLUSION: Molecular information improves CMML classification and prognostication, supports more effective clinical decision making, and potentially refines the design of clinical trials.

Survival analysis often relies on Cox models, assuming both linearity and proportional hazards (PH). This study evaluates machine and deep learning methods that relax these constraints, comparing their performance with penalized Cox models on a benchmark of three synthetic and three real datasets. In total, eight different models were tested, including six non-linear models of which four were also non-PH. Although Cox regression often yielded satisfactory performance, we showed the conditions under which machine and deep learning models can perform better. Indeed, the performance of these methods has often been underestimated due to the improper use of Harrell's concordance index (C-index) instead of more appropriate scores such as Antolini's concordance index, which generalizes C-index in cases where the PH assumption does not hold. In addition, since occasionally high C-index models happen to be badly calibrated, combining Antolini's C-index with Brier's score is useful to assess the overall performance of a survival method. Results on our benchmark data showed that survival prediction should be approached by testing different methods to select the most appropriate one according to sample size, non-linearity and non-PH conditions. To allow an easy reproducibility of these tests on our benchmark data, code and documentation are freely available at https://github.com/compbiomed-unito/survhive.

Several computational pipelines for biomedical data have been proposed to stratify patients and to predict their prognosis through survival analysis. However, these analyses are usually performed independently, without integrating the information derived from each of them. Clustering of survival data is an underexplored problem, and current approaches are limited for biomedical applications, whose data are usually heterogeneous and multimodal, with poor scalability for high-dimensionality. We introduce VAE-Surv, a multimodal computational framework for patients' stratification and prognosis prediction. VAE-Surv integrates a Variational Autoencoder (VAE), which reduces the high-dimensional space characterizing the molecular data, with a deep survival model, which combines the embedded information with the clinical features. The VAE embedding step prioritizes local coherence within the feature space to detect potential nonlinear relationships among the molecular markers. The latent representation is then exploited to perform K-means clustering. To test the clinical robustness of the algorithm, VAE-Surv was applied to the Genomed4all cohort of Myelodysplastic Syndromes (MDS), comparing the identified subtypes with the World Health Organization (WHO) classification. The survival outcome was compared with the state-of-the-art Cox model and its penalized versions. Finally, to assess the generalizability of the results, the method was also validated on an external MDS cohort. Tested on 2,043 patients in the GenomMed4All cohort, VAE-Surv achieved a median C-index of 0.78, outperforming classical approaches. In addition, the latent space enhanced the clustering performance compared to a traditional approach that applies the clustering directly to the input data. Compared to the WHO 2016 MDS subtypes, the analysis of the identified clusters showed that the proposed framework can capture existing clinical categorizations while also suggesting novel, data-driven patient groups. Even tested in an external MDS cohort of 2,384 patients, VAE-Surv achieved a good prediction performance (median C-index=0.74), preserving the interpretability of the main clinical and genetic features. VAE-Surv enables automatic identification of patients' clusters, while outperforming the traditional CoxPH model in survival prediction tasks at the same time. Applied to MDS use case, the obtained genetic-based clusters exhibit a clear survival stratification, and the application of the clinical information allowed high performance in prognosis prediction.