Two days, one venue, five challenges, five runners up , one winner!
Join our community to advance the use AI/Deep Learning in life sciences and healthcare!
From the video above and the story below you can hear about our hackathon, the first step in us showing the potential of AI and Deep Learning to transform aspects of our industry.
The Pistoia Alliance now want to help our industry bring the life sciences and healthcare industry and AI/Deep Learning community together to accelerate the potential of deep learning to aid drug discovery and bring new life saving treatments to the world.
We have started a Community of Interest to understand the barriers to making this happen with the aim to build a cross industry collaboration to overcome these challenges together as an industry instead of as individuals.
For more information contact Nick Lynch
For the background on the hackathon, here’s the story:
The Story of the Hackathon
Pistoia Alliance wants to lower the barriers to R&D in the life sciences / healthcare industry. Our innovation related events aim to be the starting point for developing new pre-competitive collaborations within this space.
We decided to run our first ever hackathon to demonstrate how new technologies can drive our industry forward. We chose deep learning as the first theme because it has the ability to transform many parts of modern life as recent innovations show.
The Pistoia Alliance wanted to bring the deep / machine learning community together with the life-science and healthcare industry to demonstrate the potential of deep learning to aid drug discovery and bring new life saving treatments to the world.
The result was a truly high energy event showing just some of the areas deep learning can be brought to.
If you want to know more, have ideas for future hacks or wish to get involved contact David Proudlock
Here’s the story:
The challenges were provided from our sponsors at Elsevier, ExCAPE consortium, Promeditec and Microsoft. They aimed to show the breadth of opportunity to apply deep and machine learning to life science R&D, they were:
- Demonstrate the ability of deep learning to help Findacure, a UK-based charity, accelerate treatment and clinical research for Friedreich’s Ataxia. Thanks to Elsevier, they had access to a heterogeneous set of data related to the disease: biological pathway analysis, associated chemical compounds and bioactivities, potential candidates for drug re-purposing, full-text scientific literature, and clinical trial data.
- Using the ExCAPE consortium's machine learning ready dataset, propose an innovative Deep Learning Pipeline for a new model that can accurately predict the in vitro activity of compounds based on their chemical structure.
- Promeditec asked: is it possible to use deep learning to reconstruct a 3D model of aorta from digital slices? The aim is to support surgeons in detecting and treating Thoracic Aortic Aneurysm. Such a reconstruction in tandem with computer simulation could lead to better detection and diagnosis of this potentially fatal condition.
- Microsoft set a challenge to train a decision tree model to predict clinical significance of missense mutations from at least three features, including:
- BLOSUM score,
- minor allele frequency
- sequence alignment entropy
Access a reddit social media forum on asthma and show how you could extract insight on patient behaviour, treatment, medication and anything else that can be derived from the forum.
Team In Too Deep won the event with their neural network tackling the ExCAPE dataset and their approach of building a prototype cloud implementation to distribute the neural network and data. The idea would be to enable others to securely upload their data against the neural net without exposing their datasets to potential competitors. Thus allowing the model to develop and improve over time as the new data joins the existing data for the model.
There were 5 runners up:
Team Find Me A Drug - tackling Elsevier’s and Findacures datasets to scour data across scientific journals, posts and R&D data to help preliminary compound identification that may warrant further investigation.
Team Tuna Melt: taking on the ExCAPE challenge utilised autoencoder to embed data into the neural network and then implemented a multitask learning approach. The intention is to then tackle the unbalanced labels through controlled dropout and use matrix factorisation to complete the label matrix prior to training.
Team The CT Guys identified an initial prototype model using detection of the lungs in the image slices combined with machine learning to identify the aorta and order the slides for a 3D reconstruction.
Team Cambridge Stars built a preliminary mobile app to help people searching the asthma stream for advice. The machine learning provided a guide to posts appearing to give better advice and then learns from user feedback on their opinion of the suggested post. Essentially the app would help train the AI.
The Mutation team took on Microsoft’s DNA mutation challenge utilising Keras to prototype and Microsoft’s Azure Machine Learning Studio to deploy a pathological mutation classifier.
What happens next?
All the contact details of the teams have been passed to the challenge providers, already:
- In Too Deep are in early discussions as to pursuing their idea further
- Find Me a Drug have been invited to present their ideas to Elsevier directly.
- The CT Guys have already been contacted by Promeditec who have shared their presentation with their partners to see if this is a potential new collaboration.
- Team Tuna Melt intend to put their work onto GitHub subject to agreement with the ExCAPE team to enable further development of their work.
What we’ve learnt.
The power of hackathons
The hackathon brought a huge amount of energy and passion to our door. This model is a great way to generate and test new ideas. Bringing a diffuse set of skills together really helps look at established problems in a new way.
We will definitely pursue future hackathons to bring more of our Pistoia members together to advance collaboration in our industry.
Teamwork / Collaboration:
Watching strangers become highly interactive and effective team in under 2 days was inspiring, from this it was clear that for success:
- Keep teams small (the Amazon pizza test”)
- Have all their focus, not juggling many different projects as is common in large corporations
- People work in different ways: from flip charts to single laptops to a team building a Slack group to continue discussions overnight. The environment gave them great flexibility.
Deep Learning Opportunities
Standards and clarity:
As an industry we need to build standard, structures and documentation that can make our data more easily usable for deep learning. The better we can describe our datasets will help bring the wider data science experts into our industry. The clearer our meta data and descriptions are, the more we can leverage community tools in the vein of Kaggle and others.
There’s a Deep Learning community out there waiting for us!
Openness and Transparency
The more open and transparent we are with our data and algorithms will help more innovation in our industry.
There's opportunity in unstructured data
Unstructured data may well have a wealth of insight on how patients are coping with their disease, in terms of their medication, the impact on their lifestyle and general well being. There are many sources of health and wellbeing open data waiting to be accessed, the challenge is to extract it and consider how it is relevant in a regulated environment.
Events like this hackathon and a collaborative approach across the industry to pool resources and lower costs instead of a silo’d one will be key to making this happen.