Automated phenotyping to accurately infer functional variants in clinical genetics

Rare diseases are numerous - so much so that as a group they are very common. One person in 17 has a rare genetic disorder, but most fail to receive a genetic diagnosis. Diagnosing the genetic cause of a disorder, even when there are only a handful of cases in the world, represents the first of many steps in finding effective treatments. Even though new genetic tests promise to assist in diagnosing some of these patients, the tests are expensive and currently only available to a few people in the wealthiest countries.

For the past 65 years expert clinical doctors have been matching a diagnosis to patients based on facial features and follow up clinical tests. We are developing algorithms through which a computer will learn and apply these skills objectively.

Identifying patients with the same genetic disorders allows comparisons to be made between them. In turn, this can improve estimates of how the disease might progress and allow direct therapeutic benefits, for instance by showing which symptoms are caused by the genetic disorder and which symptoms might be caused by other clinical issues that can be treated.

Using the latest research in computer vision and machine learning the algorithm automatically analyses patient photographs and finds their place in "Clinical Face Phenotype Space" (CFPS). Patients that share a specific dysmorphic disease or syndrome, will cluster together in CFPS. The CFPS model is created and shaped using ordinary, family album photos and accounts for variations between images that are not disease relevant (such as lighting, image quality, background, pose, age, gender, ethnicity, and facial expression).

In the present application we seek funding to develop methods for clinical geneticists to query CFPS for clinically relevant information. The work will develop means by which a patient's similarity to other patient groups can be visualised, explored and tested through robust statistical modelling.

Furthermore we will make it possible to overlay a patient's DNA with CFPS to identify disease causing mutations. This will improve our understanding of how rare diseases disrupt the normal functioning of the body, and in turn influence decisions in treatment strategies.

A clinician should, in future, be able to take a smartphone picture of a patient and query CFPS to quickly find out which genetic disease the person might have. For diseases unknown to medical science, CFPS will find if there are any other patients around the world that might have the same disease.

CFPS will learn from our faces to help diagnose rare diseases.

We will pursue three primary methodological approaches required to allow us to leverage CFPS together with genotype data to aid clinical diagnoses.
1. Develop methods to allow clinicians to explore and interpret data in CFPS.
We will build on the foundation of the CFPS pipeline to return clinically meaningful data interpretations. This will require a 2nn based and Bayesian Risk framework for estimating the probability that two faces show similarity. Using dimensionality reduction approaches such as Stochastic Neighbour Embedding we will create navigation and visualisation software to explore informative vectors in CFPS. We will develop a likelihood based method for assessing consistency within groups to detect endophenotype stratification within syndromes.

2. Infer causative genotype-phenotype associations by integrating functional gene networks within the CFPS.
Next, we will develop methods to integrate DNA variation into CFPS. This will allow the automatic analysis of craniofacial phenotypes to predict the likelihood of functional pathways being causally disrupted in syndromes. We will use functional genetic pathway information to improve the predictive power of CFPS for clinical genetics and, vice versa, we will also use CFPS to enhance the power of functional gene networks in computational biology. To explore relationships between facial similarities (CFPS) and gene functional similarities we will use machine learning to alter the structure of informative vectors in CFPS to create a honed Functional Genetic CFPS (fgCFPS).

3. Create a robust statistical framework to infer perturbations in functional genetic pathways from facial phenotypes.
Finally, we will develop methods for querying fgCFPS to aid the understanding of complex genetic disorders with unknown aetiology. From the new fgCFPS we will develop methods that assign likelihood estimates for the contributions of functional pathways to syndromes where causative genetics are unknown or complex.

This project is highly collaborative both within the UK and internationally. This is also translational and interdisciplinary science, criss+crossing the fields of computer vision, computational biology, and clinical genetics. Consequently it provides several direct pathways to impact. The most important of these is that our work will aid in the clinical diagnosis worldwide of the genetics of rare disorders. We have established and will be expanding an international network of collaborations to apply and develop our approach using patient data. We expect this project in 3 years' time to have progressed to being in the early stages of implementation in clinical genetics counselling, initially in the UK and Europe and then elsewhere.

In the space of three years the following directions will lead to measurable impacts:

The methods to interrogate CFPS together with genotype will add statistical power to the inference of genetic diagnoses. In collaboration with clinical sequencing efforts our approach will be applied to patient image data to add value to these existing datasets and on-going projects. This will result in research participants receiving diagnoses that otherwise they would not.

As part of the larger research effort we will set up partnerships with patient support groups for syndromes with identified genetic origins. We will present the science and the project as a means of public engagement, and offer the opportunity to contribute photos (under an informed consent model) for the further refinement of the CFPS. This is public engagement of the best and most interactive kind, which can be measured by the number of images of individuals (>10,000) and syndromes (>200) that we hope to collect over the coming years.

We will expand collaborations to additional clinical geneticists across the world. For clinicians the practical utility would be to evaluate the approach as a means of narrowing down hypotheses for the clinical diagnosis of known syndromes. Their results will also directly feed back to the project, as they will expand the number of images of clinically defined syndromes that will iteratively improve the CFPS - the more people that use it the better it gets.

We anticipate that extending contacts within the MRC Units in The Gambia and Uganda would be a beneficial means of geographically widening our research to efficiently test and apply CFPS in the future.