Matching the right drug with the right patient does not only increase the treatment success, but also saves the patient from unpleasant side-effects of suboptimal treatments . For delivering personalised treatments, we use as models cancer cell lines in large-scale pharmacological screens and their genomic profiles . These studies enable the identification of genes that separate patients into responders versus non-responders that are commonly known as biomarkers of drug response. Identifying biomarkers is a statistically challenging task and we focus on the development of novel methods. For example, we developed machine learning methods that do not only leverage the genomic aspect, but also the chemistry of drugs  to predict the drug response .
Footprint methods for improved drug prediction
A main challenge in modelling is that the number of samples is relatively small when compared to the number of molecular features to consider, such as genomic variants or gene expression. We, thus, develop and apply various footprint methods that leverage mechanistic knowledge of the underlying signaling pathways and regulatory modules to limit the number of features [5,6]. The lower number of features increases statistical power and at the same time provides understandable and actionable biomarkers .
Linking drug target and pathway activity with multi-task learning
We use matrix factorization to generalize from individual drugs and genes and explore the interactions between the drug targets and signaling pathways’ activation, thereby providing robust and mechanistically interpretable insights . A typical insight would be: “Activation of pathway Y confers sensitivity to a drug targeting protein X”. These interactions can guide a tissue-specific combination treatment strategy, for example suggesting to modulate a certain pathway to maximize the drug response for a given tissue. Such an analysis of interactions across tissues might help target discovery, drug repurposing and patient stratification strategies.
Macau factorization model: (A) The drug response (IC50) is computed by 2 latent matrices. Each of them is being sampled by a Gibbs sampler. In presence of additional information (side information), the latent matrix is predicted by a multiplication of a link matrix and the side information matrix. Arrows in this figure indicate the matrix multiplication. (B) By multiplying the 2 link matrices, we obtain the interaction matrix, which is the interaction between the features of the drugs with the features of the cell lines.
Personalize treatments with network-based approaches and multi-omics integration
Targeted therapies and personalised treatments are promising assets to treat cancer. Each patient develops cancer in its unique way. We take advantage of the flexibility of network modelling and the power of patient molecular features (such as genomics and transcriptomics) to find actionable causal biomarkers using Omnipath as prior knowledge network for our causal network reconstruction tool CARNIVAL.
Work-flow to obtain causal-reasoning networks embedding molecular features in the scaffold network. From each cell line or patient sample, the mutational load and the expression are used to create sample-specific scaffold networks. From a fully signed and directed network obtained from Omnipath, the molecular features constrain the network to make it sample-specific. The non-expressed proteins are removed from the initial scaffold; the genomic variants are translated to their functional impact using IMEx and mutfunc resources. This translation allows us to remove specific interactions that are disrupted, or proteins that are not functional.
Drug combination and repositioning
To increase therapeutic options and to overcome drug resistance, cancer researchers have been actively investigating drug combinations. The rationale is that by targeting multiple mechanisms simultaneously, the potency of the treatment is increased and tumor cells are less likely to develop resistance and become refractory to treatment. We were also involved in an evaluation of computational approaches to predict drug combinations in the context of the DREAM challenges .
Pharmacogenomics is based on (A) pharmacological screens of cell lines and (B) their deep molecular characterisation. (C) Both different data sources enable the identification of sensitivity biomarkers, e.g. a cancer somatic mutation might render cell lines sensitive to treatment X. (D) Shows a computational approach, which additionally to the deep molecular characterisation of cell lines, also considers the chemistry of the compounds [adapted from ]. (E) Using a microfluidic platform, we can screen cells derived from tumor samples resected from patients. Cells are co-encapsulated along with single drugs and drug combinations and resulting data are used to analyze drug interactions .
Microfluidics for ex vivo drug screenings
State of the art screening technologies allow high-throughput screening of a large panel of drugs and cell lines but they cannot be applied to primary tumors freshly resected from patients, mainly due to the limited amount of malignant cells that can be recovered from a biopsy or a resection. For these reasons, we are collaborating with Christoph Merten at EMBL to develop a droplet-based microfluidics platform to perform combinatorial drug perturbation screening on patient samples. The main advantages of using microfluidics are that: 1. it allows to perform automatic combinatorial drug screening experiments with low amounts of reagents and 2. it requires only a reduced number of cells per droplet (thus per experiment) allowing the application also to primary cells, including patient-derived samples, without need for intermediate culturing steps. We have applied this technology to identify combination therapies for pancreatic cancer, for which very limited therapeutic opportunities exist, in collaboration with Thorsten Cramer (Molecular Tumor Biology and Department of Surgery, RWTH Aachen University Hospital) . We use our logic-based methods to build patient-specific signaling models that in turn can be used to predict new combination .