Bioinformatics

This is the first algorithm to be applied and it mines the data distilling out the key biomarkers that can predict the clinical question being asked.  This is achieved through the production of a ranking of biomarkers based on the error of predictions for an unseen data set averaged across multiple repeats.  The findings made in this approach feed into other approaches described below.

Expression array studies are frequently underpowered if considered in isolation.  This approach applies the data mining algorithms described above across multiple data sets.  The top ranked probes are then cross compared to find commonality (Figure 2.).  Probes that are found in common between multiple data sets are unlikely to have occurred by random chance.  This approach utilizes this statistical principle to increase the certainty of the markers discovered.  This approach is used to Increase the statistical power of the marker set discovered and reduce the risk of false discovery.  Furthermore, the approach increases the generality of markers discovered so they are more likely to be able to predict for the general population.  Integration is achieved at the probe level for the same array platform and at the gene level for different array platforms. The approach has been used in the Abdel fatal (2016, Lancet Oncology) study to identify a key set of markers for proliferation in breast cancer by integrating datamining across 4 datasets.  Here the biomarkers identified had a false discovery probability of 2x10^-74 and have now been validated using immuno-histo-chemistry in over 15000 cases.

By using the above algorithm in a stepwise additive fashion on large molecular datasets, we can build an optimized panel using the best subset of markers that have the greatest sensitivity and specificity.  If this is coupled with the concordance analysis a very robust diagnostic panel can be developed.  Classifier panels are assessed for both seen and unseen data to assess their suitability as a diagnostic for the wider population and performance is evaluated based on Receiver Operating Characteristic (ROC) Curves.  The population is then characterized and ranked based on the probability of disease for a given individual.

This approach adopts a systems biology and pathways analysis approach. A set of markers derived from the core analysis described above are used in network inference algorithms to identify a network of interactions. This network is analysed to identify the key molecular drivers and the most influential in a given system. The approach has been used in a commercial contract with Syngenta to identify transcriptomic regulators of ripening in tomato. Pan Y. et al, Plant Physiology, 2013.

The approach was also used in the Abdel fatal (2016), Lancet Oncology study. This use of a systems approach goes further than a simple list of markers because biology is defined by the interaction of molecular markers. The approach refines the marker set identified and can be used to identify molecular based disease processes that differentiate between a healthy population a diseased population (Therapeutic target identification In Silico), processes associated with therapeutic response or pharmacodynamics.

Knowledge of biological pathways is essential for drug discovery and development. However, our knowledge of the biology of pathways and what drives them is scant. Often pathways have been modelled based on simple reductionist experiments in the lab and the results do not reflect the nature of the pathway in a given disease or reflect a broader evaluation of the whole transcriptomic or proteomic nature of the pathway.  In short there are numerous gaps and omissions in our knowledge of pathways, yet they are the basis of many expensive drug discovery strategies.

We have, based on the success of this work, developed a new systems biology and pathway interrogation methodology (Intellomx Pathway Miner). This approach first uses our patented Distiller algorithm to mine transcriptomic data for selected diseases using existing pathway knowledge as the framework for data interrogation. This approach identifies new disease specific pathway features from across the whole proteome/transcriptome, creating in effect an augmented pathway specific to the disease being investigated.  The extensive cross validation conducted in distiller ensures that the markers we identify and robust having been validated across multiple data sets.

Once the augmented pathway has been created, new and existing features are added to the Driver algorithm. This uses the network inference and systems biology algorithms to identify the strength and nature of the influence of each molecule on the pathway. In this way the most influential and the most influenced molecules in a given pathway for a given disease can be identified. These influential molecules, given their relationship to phenotype as described above, are the most likely druggable target contenders.

Through the interrogation of Drug Gene Interaction Databases, we can identify existing drugs that will interact with these molecules. Thus, we can identify using In silico methods potential repurposing targets.

Our ability to identify new potential targets is demonstrated by the results of our R&D programme where we have analysed a number of high quality Public Data Sets for lung cancer and through application of our Pathway Miner methods, identified a number of novel targets for known drugs and potential therapies. Below we present some of the results for unknown but influential genes for enriched MEKK pathway in lung cancer (See Figure 6). We have also applied our approaches to a wide range of other cancers and pathways generating a pathway driver IP portfolio.