Our work in collaboration with Imperial College on modelling of disease subtypes based on proteomic expression in asthma has been published in The Journal of Allergy & Clinical Immunology. The publication discusses how novel types of analysis have the potential to unlock research and treatment options for patients by applying machine learning, bolstering scientific understanding and improving patient care.
Many chronic conditions have complex and differing presentations, and physicians rely on simple heuristics to differentiate between them to make appropriate treatment decisions for patients.
Combining biological and clinical datasets currently used one dimensionally in the treatment of patients can give a more holistic view of the underlying state of the patient. This helps to ultimately find the best course of action to optimise clinical outcomes. By doing this, we also head towards true personalised medicine, where interventions can be developed and administered to more closely match an individual, and we’ll start to see that broad medical conditions such as “asthma” are actually clusters of closely related syndromes.
There are many biological indicators that can aid these diagnoses, including genetic data, proteomic data, clinical history and imaging data. These are often impractical or costly to obtain in a care setting, and combining all different types of indicators rapidly exceeds human capability. This is where machine learning methods can benefit research and healthcare by incorporating multimodal datasets.
Leveraging high dimensional data to aid in our understanding of health and disease is far from simple. Sophisticated computational techniques are usually required to identify the complex patterns, clusters, and associations that lie hidden in data like this. In this case we used the Uniform Manifold Approximation and Projection for Dimension Reduction (UMAP) technique to handle the data’s high dimensional nature and a Gaussian Mixed Model to identify clusters for us, combining this with the use of Shapley value analysis to provide insight into what biochemical features were driving the probabilistic cluster membership decisions made by this model. These approaches allowed us to pull important insights from data despite its complex nature.
“These new types of analysis will become incredibly important in the biomedical field – revealing new associations in the data that aren’t obvious from traditional methods of inspection which can enable more research and ultimately new therapies”
Professor Mohamed Shamji, Faculty of Medicine, National Heart & Lung Institute
Faculty believes that these types of machine learning led multimodal analysis, will continue to unlock our understanding of human biology and help to drive advances in personalised medicine in diseases as diverse as Alzheimer’s, diabetes, cancer and heart disease.
You can read the full published journal article here: Multidimensional endotyping using nasal proteomics predicts molecular phenotypes in the asthmatic airways
Special thanks to the Faculty team Giulia Vecchi, Tara Ganepola, Ben Bedford and the team under Professor Shamji at Imperial College