Summary of M19-M36 of iPC

14 April, 2022

The treatment of paediatric cancers presents specific challenges distinct from the treatment of adult cancers. iPC focus on developing a comprehensive computational set of models & software to address some of the most challenges that clinicians face in their daily treatment of paediatric patients. iPC is combining a multitude of knowledge-base, machine-learning, and mechanistic models to predict optimal standard and experimental therapies for each child. Our approach is based on the development of virtual patient models, i.e. in-silico avatars who resemble the molecular and clinical landscape of the child patient and can be used to investigate computationally personalised diagnostics and recommend treatments. iPC is developing a computational platform that also allows caregivers to query models and infer benefits and drawbacks for specific treatment combinations for each child.

Work performed from July 2020 – December 2021

WP1: Data collection and generation

The collection of molecular and clinical data from past and ongoing trials is progressing and one ongoing trial for HB is now actively testing our predictive models. Datasets that are specific to our proposed tumour types or that could be used to build and test predictive models are being generated by multiple partners. Work to collect and harmonise multiple datasets is proceeding as planned. Efforts have been devoted to generate animal models necessary for drug testing, as well as to conduct genetic perturbation experiments where the readouts are assessed by single-cell RNA sequencing readout.

WP2: Platform Implementation

A first version of the iPC’s Computational Platform was released in Sept 2021, which included a data catalogue system integrated to an analysis platform (openVRE). Efforts have been made to implement a data access framework operating on top of the current infrastructure and a prototype version of this framework has already been released. The platform is based on a microservices architecture, and the different services are attached to CI/CD for automated testing, delivery & deployment to production environments, and therefore, easing not only the collaboration between development teams, but also the reusability of the different platform’s components. Importantly, (meta)data from different partners as well as analysis tools (e.g. DoRothEA, PROGENy, COSIFER, DpFrEP) have been included to the platform.

WP3: Big data analysis of available data sources

Successful integration of pipelines for the text mining analysis of biomedical publications (INtERAcT, MelanomaMine and LimTox). We worked on the development of a computational pipeline for the deconvolution of multi-omics paediatric datasets using matrix factorization approaches. Development of interpretable deep learning models. 2 applications have been investigated: the prediction of protein properties from sequence, e.g. binding between a T cell receptor and an antigen, and the cell-type classification from single-cell data.

WP4: Network-based meta-analysis of multi-omics and text-mined data

We worked together on application of network-based computational methods to jointly analyse a large corpus of multiomics datasets on several types of paediatric cancers: medulloblastoma, Ewing sarcoma, neuroblastoma. A computational pipeline for construction of molecular networks and patient similarity network has been developed and applied to several tens of mutliomics data. We developed a novel computational method for dimensionality reduction, based on multilayer network community detection, aiming to facilitate the molecular characterization of patient stratification. A totally redesigned version of NaviCell web-based tool for interactive analysis and visualisation of multi-omics data using biological networks has been implemented.

WP5: Blending machine learning and mechanistic models

Development: The algorithm and its scalable implementation for probabilistic bilinear/linear tensor factorization for graph learning. A computationally efficient variational criterion for active structure and parameter learning on synthetic and real-world data. An augmented CTBN model with local clocks at each node that captures non-exponential waiting times between state change events. A method for polypharmacy side-effects prediction based on relational graph embeddings. A modular approach to pseudotime reconstruction from scRNA-seq data that improves accuracy on simulated and empirical data. A deep neural network based inference method for detection of nonlinear dependencies that assigns graph structures to data.

WP6: Mechanistic and agent-based models

Worked on the extension of the ModCell model to capture all relevant mutations. Using this model, we performed in-silico drug screenings of hepatoblastoma patient samples. Furthermore, a module for automatic report generations was implemented and extended for immunotherapy-based simulations. The optimization pipeline was established and benchmarked on synthetic “ground truth” data and the tool “scycle” for analysing cell cycle trajectories from scRNAseq was developed.

WP7: Transcriptomic models for clonal deconvolution, intratumoural heterogeneity and non-coding elements

An optimised and validated workflow for deconvolution of bulk RNA-sequencing data from AML and NB patients through integration of matching single cell and bulk RNA-sequencing data of AML samples has been developed. UGent & BCM established a computational framework to map regulatory networks of non-coding RNA transcripts, by integrating computational target predictions through the LongHorn algorithm with intron and exon coverage data from total RNA-sequencing data.

WP8: Multi-omics and metabolomics models for drug discovery

We developed a computational method for predicting small molecule inhibitors of cancer cell growth. Selected drugs for HB patients are being validated and we identified molecular targets for high-risk AML and HB patients. Developed: a pipeline to run CARNIVAL in a multi-omic approach to create sample-specific networks and find causal links, and a metabolic model of intratumoral heterogeneity within Ewing sarcoma tumours. We derived both a liver and an HB metabolic model. We are combining single-cell proteomics and transcriptomics data, and a proteomic study involving has led to identification of deregulated pathways and 3 prognostic proteins in HB patients.

WP9: Use cases and validation

Collaboration on high-spectral flow cytometry and data analysis of samples from AML patients. We established an arrayed CRISPR-interference screening platform for high-throughput gene perturbation with serial cellular and molecular phenotyping. We are collaborating on the development of a machine learning pipeline for the identification of biomarkers in HB. We established an “HBpatient platform” with more than 150 specimens from retrospective and prospective collections and we performed RNA seq and whole-exome seq of tumours from 32 TCF3-PBX1 BCP-ALL patients at different stages and 14 PDXs. We identified that ALK and Hedgehog pathway may constitute potential therapeutic targets in HB. Co-culture systems have been optimised and we built an in vitro testing platform for HB.


Progress beyond the state of the art, expected results until the end of the project and potential impacts (including the socio-economic impact and the wider societal implications of the project so far)

The integration of the multiple data modalities collected for each patient holds the key to precise personalised models for diagnosis, prognosis and improved treatment for a multitude of diseases, including paediatric cancers. However, the volume of its complexity has so far prevented its widespread translation into clinical practice. By combining expertise ranging from computational modelling to clinical patient management, iPC is building innovative clinic-ready software tools to address the real needs of paediatric tumour patients. In doing so, iPC is making possible for clinicians and health professionals to readily interpret multi-level data into actionable information for each patient.