The health system is producing data at an unrestrainable speed; data that can mean personalized therapy, patient safety, personalized cancer prognoses, better prevention, and monitoring of epidemics. We show how such data can be exploited, with a series of innovative prototype projects.

Personalized cancer therapies: Modelling cancer drugs sensitivity and synergy in in-vitro screening

The analysis of whole genomes of pan-cancer data sets provides a challenge for researchers, and we have developed a novel rank-based Bayesian clustering method to tackle this unsupervised problem. The advantages of our method are the integration and quantification of all uncertainties related to both the input data and the model, the probabilistic interpretation of final results to allow straightforward assessment of the stability of clusters leading to reliable conclusions, and the transparent biological interpretation of the identified clusters since each cluster is characterized by its top-ranked genomic features. A paper is published.

Cancer pharmacogenomic screens profile cancer cell lines versus many potential anti-cancer drugs to identify new combinations of drugs that have a high probability to work on individual patients. We work with data generated by our partners at Oslo University Hospital and public data to guide therapy based on the statistical prediction of how drugs will behave for individual tumor samples. To improve predictions, we are exploring both structured penalised regression models and structured priors in multivariate Bayesian models to incorporate prior knowledge about the dependence structure between drugs and between multi-omics profiles of cancer cell lines. In particular, we developed the mix-lasso model that introduces an additional sample group penalty term to capture tissue-specific effects of features on pan-cancer response prediction. The mix-lasso model takes into account both the similarity between drug responses (i.e., multi-task learning), and the heterogeneity between multi-omics data (multi-modal learning). When applied to large-scale pharmacogenomics dataset from Cancer Therapeutics Response Portal, mix-lasso enabled accurate drug response predictions and identification of tissue-specific predictive features in the presence of various degrees of missing data, drugdrug correlations, and high-dimensional and correlated genomic and molecular features that often hinder the use of statistical approaches in drug response modeling. Two papers are published.

For combinatorial treatments, prediction of likely synergistic effects is crucial to suggest efficient combinations. We have proposed PIICM, a probabilistic framework for dose-response prediction in high-throughput drug combination datasets. PIICM utilizes a Permutation Invariant version of the Intrinsic Co-regionalization Model for multi-output Gaussian Process regression, to predict dose-response surfaces in untested drug combination experiments. Coupled with an observation model that incorporates experimental uncertainty, PIICM is able to learn from noisily observed cell-viability measurements in settings where the underlying dose-response experiments are of varying quality, utilize different experimental designs, and the resulting training dataset is sparsely observed. One manuscript is submitted for publication

Healthcare safety management

There is an extreme amount of information available in electronic health records that can be used to learn the behaviour of healthcare institutions, make predictions, guide treatment choices and so on. We have been working on electronic health record data from Akershus University Hospital (AUH) on a project to explore patients’ movements within the hospital and how these may affect the risk of spread of infections. We have been using network models and focused on the evolution of- and differences between networks, in time and space. The project is mostly descriptive and aims to inform decision makers. One paper is published and the PhD student defended her thesis in 2022.

Exploring clonal heterogeneity in blood cancers for personalised treatment.

Our goal is to develop a data-driven modelling framework to improve treatment strategies in blood cancers. BigInsight has strong clinical and experimental collaborations in blood cancer at OUS as well as access to unique datasets. One major obstacle to developing personalized medicine is the presence of cellular heterogeneity within the cancer cell population of each patient. This can lead to a common scenario where a therapy initially succeeds at reducing disease burden, but the cancer eventually rebounds due to the outgrowth of a minor but drug-resistant clone. To address this obstacle, we have developed a new method to estimate and quantify the heterogeneity present in each particular cancer. We use available high-throughput drug screening data to infer the subpopulation substructure. Our statistical platform, called DECIPHER, estimates the number of distinct clones present as well as how these clones respond to a specific drug, based on drug screens of patient samples. This information then feeds into evolutionary models of drug response to therapy, to predict the effect of a drug. We use a combination of mathematical modelling and inference for mixture models. Successful implementation of our method will potentially greatly aid in the management of different types of blood cancers, and potentially also solid cancers.

Mathematical models and Bayesian inference in personalised breast cancer therapy

Current personalized cancer treatment is based on bio-markers which allow assigning each patient to a subtype of the disease, for which treatment has been established. Such patient classification represent a first important step away from one-size-fits-all treatment. However, the accuracy of disease classification comes short in the granularity of the personalization: it assigns patients to one of a few classes, within which heterogeneity in response to therapy usually is still very large. In addition, the combinatorial explosive quantity of combinations of cancer drugs, doses and regimens, makes clinical testing impossible. Our strategy for personalised cancer therapy is in silico, based on producing a copy of the patient’s tumour in a computer, and to expose this synthetic copy to multiple potential therapies. We show how mechanistic mathematical modelling, patient specific inference and simulation can be used to predict the effect of combination therapies in a breast cancer.

The model accounts for complex interactions at the cellular and molecular level and is able of bridging multiple spatial and temporal scales. The model is a combination of ordinary and partial differential equations, cellular automata, and stochastic elements. The model is personalised by estimating multiple parameters from individual patient data, routinely acquired, including histopathology, imaging, and molecular profiling. The results show that mathematical models can be personalized to predict the effect of therapies in each specific patient. The approach is tested with data from breast tumours collected in a recent neoadjuvant clinical phase II trial at OUS. This year, we have been able to develop a numerical algorithm that allows the simulation of a full biopsy, exploiting parallel computing. This study is possibly the first one towards personalized computer simulation of breast cancer treatment incorporating relevant biologically-specific mechanisms and multi-type individual patient data in a mechanistic and multiscale manner: a first step towards virtual treatment comparison.