NeurIPS 2019 fact-sheets: overviews of the research we presented, and its future impact
Despite all the advances in cancer therapeutics seen over the past few decades, one problem still remains abundantly clear: with up to 64% of patients still not responding to their first-line therapy, what can we do to better target patients with a therapy that will work for them?
The advent of “precision oncology” (applying targeted therapies to a specific individual’s biomarkers) promises to provide huge impacts in cancer therapeutics. Through comprehensive genomic profiling, oncologists are able to match patients to entirely personalised regimens of targeted therapies based on their unique DNA variants. The cornerstone of such an approach is in the correct identification of these DNA variants, and the challenge is in doing so from the vast genomic sequencing data produced from an individual’s tumour sample. This process, known as “variant calling”, is a complex classification task to distinguish true DNA variants from sequencing noise, which in some cases needs to be performed billions of times, across every position in an individual’s genome.
A complex big-data classification task such as this is well-matched to the latest deep learning AI approaches, which could provide great impact in the successful matching of patients to the best targeted therapies. However, adoption in a clinical setting requires these so-called “black-box” algorithms to be far more understandable, safe, and robust before they can be deployed at a wider scale. At this year’s NeurIPS conference for machine learning, we presented our work on Bayesian neural networks, and the utilisation of these in ensuring oncologists have much greater confidence in decisions made by such machine learning approaches to variant calling.
The application of powerful machine learning models, such as neural networks, to the task of variant calling has the disadvantage of simply providing an oncologist with a binary “yes” or “no” output as to the presence of a DNA variant. The latest methods from our team at CCG.ai demonstrate a novel approach to neural network variant calling that additionally provides the oncologist with an indication of how much confidence the model has in the call. This is accomplished through the use of a Bayesian neural network: augmenting a Bayesian inference approach with the power provided by standard neural networks. Instead of dealing with point estimates, Bayesian neural networks learn entire probability distributions to provide a measure of uncertainty. This in turn makes the overall network more robust in its predictions by giving it the ability to say “I do not know” if a given variant is particularly noisy. This is a hugely important property for decision making, and especially so when selecting which variant in a tumour is the best candidate to target with a therapy.
Analysing and understanding the exact molecular dynamics underpinning an individual’s tumour is key in getting each patient the right treatment, at the right time, to beat their cancer. The latest advancements in machine learning and AI algorithms could be highly impactful in this setting, but bringing the power of these into the clinic requires bespoke development to adapt these tools to cancer genomics. The work presented here is part of a continued mission by CCG.ai to combine the power of AI with precision oncology and, ultimately, have a profound impact on cancer therapy selection and patient response rates.
This blog gives a high level overview of a paper presented at the NeurIPS 2019 workshop: Safety and Robustness in Decision Making.
We published 5 papers in total at NeurIPS 2019. Check out our press release to learn about our other Machine Learning advances.