MATHEMATICAL & COMPUTATIONAL MODELLING (Theme-4) Although ML/DL approaches such as Support Vector Machine (SVM), Random Forest (RF), as well as DL approaches, e.g. Convolutional Neural Network (CNN) and DeepCC, and subtype classification using a single type of omics or imaging data, have been successfully applied in cancer diagnosis and subtype classification, current existing methods may not be adequate to address the challenge of integrating genotypic data from patients, multi-omics data from drug treatments, pharmacology data from bioactivity and toxicity assays, as well as radiomics imaging data for diagnosis and classification of brain tumours, therefor, developing new algorithms are highly demanded. Pioneer studies have been conducted to integrate scRNAseq, miRNAseq, and DNAm data from TCGA using ML/DL for pan-cancer classification and survival prediction where multi-omics and clinical data are both considered, this Action aims to extend such strategies by integrating genotype and radiomics data. Specifically, omics-specific DL models including radiomics model will be built-up for a single type or multiple types of brain cancer depending on the data availability, then the VCDN approach will be utilized to integrate these models for classification, and novel approaches will be developed for optimisation. Biomarkers can be identified by their importance to classification performance which has the potential to interpret the results and understand the underlying cancer biology and pathology, especially when the correlation between radiomic and otheromics biomarkers is detected (radiogenomics study). To make ML/DL explainable and enhance its success in clinical practice, combining domain expertise with AI algorithms is essential. Therefore, this Action also aims to encourage clinicians, neuro-oncologist, radiologists, etc. to participate in the early development of AI algorithms and the whole training and validation process.
Moreover, five mathematical models will be developed to understand and overcome resistance to therapies by integrating multi-omics data as well as biomarkers discovered in liquid biopsies:
1) to optimise the scheduling/therapy combinations allowing to mitigate or overcome the problem of the emergence of resistances; 2) to simulate tumour heterogeneity influence the response to drug therapies in pre-clinical models developed during this Action; 3) to reveal “macroscopic” heterogeneity metrics correlate with tumour cell biology (genotypes and/or phenotypes i.e., invasive vs angiogenic growth); 4) to estimate prognosis in brain metastasis and understand the response of these secondary tumours to radiotherapy focusing on lung and breast cancers. 5) to develop a tumour growth model (TGM) by estimating cell density.
Finally, open-source big data processing and bioinformatics solutions — e.g., cBioportal, tranSMART, and Arvados — will be deployed for creating a pan-European brain cancer database implementing the FAIR Data Principles during this Action and be publicly accessible after this Action. Although the general limited data/sample size due to the cost in biomedical research is one of the key challenges for the application of ML/DL, it is believed through data integration/combination along with more and more data becoming available, it may not be a significant concern in the future.