High throughput imaging and quantitative phenotyping for behavioural genomics

Lead Research Organisation: Imperial College London
Department Name: Institute of Clinical Sciences

Abstract

Many behavioural traits, including those underlying psychiatric disorders, are heritable, but identifying the associated genes remains challenging. With the rapid development of genome sequencing technology, the bottleneck is no longer knowledge of gene variants, but an insufficient description of their impact on phenotype. Advances in imaging, computer vision, and machine learning are now converging and will make it possible to perform automatic and quantitative phenotyping at a scale commensurate with genomics. We are taking a combined experimental and computational approach using the nematode worm C. elegans as a model that will reveal genes and gene networks that control behaviour that have been missed in manual screens.

Students in the lab will have the opportunity to use cutting edge imaging and analysis methods to pursue their interests in behavioural genomics and/or to contribute to the development of the next generation of phenotyping technology. There is an emphasis on discovering the molecular mechanisms of nervous system function through novel analyses of behaviours including spontaneous locomotion, feeding, collective behaviour, and decision making. Students will also have the opportunity to develop skills in molecular biology, microscopy, and/or quantitative genetics.

Publications

10 25 50

Studentship Projects

Project Reference Relationship Related To Start End Student Name
BB/P504683/1 03/10/2016 02/04/2021
1810437 Studentship BB/P504683/1 03/10/2016 31/12/2020 Adam McDermott-Rouse
 
Description A lot of work developing new drugs as pesticides is done through symptomology, whereby a person will expose a pest (bacteria/fungus/insects etc.) to a drug and record any effects on its behaviour, for example changes in speed or if it's lethal. This is usually subjective in that different operators will note down on a scale, for example 1-5 of how slow the pest is. The issue here is that the data is low quality in that it doesn't give you the actual speed or a percentage reduction of the pest. The first goal of the project was to see if we can use video imaging to take away the operator from these experiments and get data that is quantifiable and would give us a behavioural fingerprint for drugs.

We developed a method for exposing the model organism C. elegans (which shares a lot of genetic background with pest nematodes) to pesticides and recording high quality videos of their behaviour. We used a piece of software (Teirpsy) that was already developed by the lab (not part of this project) to extract behavioural features from the videos. These features (<3000) could then be used to build up a behavioural fingerprint for the drug affects. We therefore have behavioural data from C. elegans treated with pesticides (110 drugs in total).

With this data we can then test the hypothesis that if two drugs target the same pathway/structures, called the mode of action (MoA), will they induce similar behavioural changes. If true, we can then use this to try and predict the MoA of a drug from the behaviour it induces alone without any prior knowledge about how it functions. We were able to use a method called multiple instance learning which takes data we have labelled as belonging to a certain MoA (A) and tries to identify what makes that group similar in terms of features (e.g. reducing speed). We can do this with the other MoA B, C, D etc. and we then have a system which has classified the different MoA by the features that affect the same.

The key finding, I believe, is that we now we have the MoA's classified, we can now test blinded drugs (ones in which we do not know the MoA) and ask the system to try and place these into one of our MoA classifiers, thereby predicting the MoA from the behaviour alone. When we did this the system had an accuracy of 61% in giving the correct MoA. While if you ask it for its best two predictions it improves this accuracy to 100%. In other words, in our data set with the test drugs we could say from the behavioural fingerprint that it is either this MoA or this one. Something which would take many different experiments if you were to do it with traditional symptomology.
Exploitation Route The outcomes of this research could help improve the drug discovery process in industry (something that our collaboration with Syngenta will work on). The idea is that the earlier you identify a drugs mode of action (MoA) the better. For example, more and more pesticides are needing to be more specific with the goal of no off species affects. If you identify the MoA of a drug belonging to a pathway that is shared among many species you would not like to pursue this for further development. By identifying this early you will save time and money that would have been lost developing a drug that would not be able to be used in the field. This also applies to resistance as if a MoA has known resistance in the field developing a drug that targets the same pathway would also not be affective as the resistance is already there.

The idea of using highly detailed behavioural data along with prediction analysis could also be expanded on with other drugs that aren't necessarily pesticides. We used pesticides as they had a high chance of inducing observable affects but it would be good to test these finding on drugs that don't particularly target our model organism.
Sectors Agriculture, Food and Drink,Pharmaceuticals and Medical Biotechnology