laboratory informatics


chemical, clinical and epidemiological data through technologies such as next generation sequencing, imaging and informatics tools has also been key. ‘Feed your algorithm with high-quality


Dr Nick Lynch oversees participants at The Pistoia Alliance’s Deep Learning Hackathon in London (left) and Team ‘Find Me A Drug’ tackles Elsevier’s and Findacure’s datasets to search for ways to potentially repurpose existing drugs to treat the rare genetic disease Friedreich’s ataxia


➤


scanner, so when a patient comes in for triage, they can be diagnosed early, at this first point of contact,’ Halabi added. Essentially, it’s about transferring the expert’s


knowledge into an AI platform. Any field that uses imaging – whether that be pathology, radiology ophthalmology, dermatology, cardiology or, in the pharmaceutical space, high- content and 3D screening, is a prime candidate for AI, Halabi suggests. Te potential for AI to inform on therapy and


particularly personalised medicine is enormous, Halabi believes. In brain cancer, for example, therapies exist that target tumours with specific mutations, but currently the only way to test for those mutations is to drill into the skull and take a brain tumour biopsy. Work by one Mayo Clinic team could make


this invasive biopsy procedure redundant. Te researchers have trained a neural network on patients’ MRI scans and tumour sequencing information, and the network learned how to identify genetic mutations in a brain tumour by analysing features in the associated MRI image. ‘It’s incredible that the system has been able


to do this, with 95 per cent accuracy. In fact, we still don’t actually know what those features are that the algorithm is ‘seeing’ in the scans,’ Halabi admits. In another area of cancer diagnosis, a Brigham


and Women’s research team is exploiting deep learning to identify the correct region of a prostate tumour from which to take a biopsy. ‘Te computer effectively draws a circle to


identify the area to be biopsied, and then even communicates with the surgical robot that is carrying out the procedure. Tis ability to use AI to talk to equipment, and make robots and scanners smarter, will be a key area for development.’ Take the predictive abilities of AI one step further, and it may be possible to predict which diseases a currently healthy individual will eventually develop, Halabi believes. ‘Researchers at Mount Sinai have fed more than 70,000 patient records and laboratory data and other health-related data – excluding imaging –


20 SCIENTIFIC COMPUTING WORLD


into a neural network. Te resulting model is now more effective at predicting disease development than the physician, and can forecast, with about 80 per cent accuracy, whether an individual will develop any of about 80 diseases within a year. In the case of predicting liver cancer the system is up to 90 per cent accurate,’ Halabi concluded.


Defining good data Machine learning is, at is most basic, concerned with defining the subject area in which you are interested, using numerical values, notes Dr Jabe Wilson, consulting director for text and data analytics at Elsevier. ‘Once you have defined something in numerical values, you can then take those values as features, and look for a formula that defines the patterns in those numeric values.’ Te key to solving such specific and highly


variable problems in life sciences is to identify the right feature set, and create the right data points whether you are looking at gene expression, target or activity data, or chemical structural information, for example, notes Jabe Wilson, consulting director for text and data analytics at Elsevier. ‘However, the key here is to remember that ‘if you put rubbish in, you get rubbish out. Yes, you have to identify the right feature set, but you also must have confidence in the reliability of the data that you are feeding in.’ Te ability to extract semantic information


from the scientific literature is also a whole industry in itself, he suggests. ‘We need quality information that has been cleaned up, so that it is relevant and can be mined. And we need a set of ontologies and taxonomies to ensure that correct terms and relationships are used as you go through scientific literature to pull out your statements. Tis is a key area of expertise for Elsevier.’ Huge advances in algorithm complexity and


quality, and access to cost-effective computing power have been key to stimulating the use of AI within the healthcare and life sciences sectors, the Pistoia Alliance’s Lynch suggests. Te ability to derive and analyse a much


greater depth and breath of genomic, biological,


data and it will perform better when it starts to encounter things that it has not seen before. Tis will improve performance dramatically,’ said Lynch. ‘Over recent years the availability of broader, but more complex and detailed data has been critical to unlocking the true value of AI in these areas,’ Lynch added. Lynch concurs with Halabi that although AI


may not completely supersede human experience in every clinical field, it will play a major role in multiple areas such as image analysis and epidemiology. ‘Google and Facebook already use deep


learning to derive and analyse key trends from the photos that we put on social media and the things


RESEARCHERS AT MOUNT SINAI HAVE FED MORE THAN 70,000 PATIENT RECORDS AND LABORATORY DATA AND OTHER HEALTH-RELATED DATA – EXCLUDING IMAGING – INTO A NEURAL NETWORK


that we look at. Te major difference for the healthcare industry will be the need to work with the regulators to develop solutions that deliver the value of AI, but that are reproducible and reliable enough to be used in a clinical setting,’ said Lynch. Te potential for deep learning to solve some


very real challenges in pharmaceutical R&D was demonstrated by the Pistoia Alliance’s first Hackathon, staged in London in March.


Encouraging AI education More than 40 previously unacquainted scientists from academia and industry came together and formed small teams that worked together to address challenges in areas including image processing, text analytics and structure activity relationships, using deep learning techniques, in just a couple of days. One of the trials in the London Hackathon


was put forward by Elsevier, which challenged scientists to use machine learning to help the UK charity Findacure identify existing drugs that may be repurposed for treating the rare genetic disease Friedreich’s ataxia, (FA). ‘Elsevier provided access to a huge,


heterogeneous set of disease-related data including biological pathways and pathway ➤


@scwmagazine l www.scientific-computing.com


Page 1  |  Page 2  |  Page 3  |  Page 4  |  Page 5  |  Page 6  |  Page 7  |  Page 8  |  Page 9  |  Page 10  |  Page 11  |  Page 12  |  Page 13  |  Page 14  |  Page 15  |  Page 16  |  Page 17  |  Page 18  |  Page 19  |  Page 20  |  Page 21  |  Page 22  |  Page 23  |  Page 24  |  Page 25  |  Page 26  |  Page 27  |  Page 28  |  Page 29  |  Page 30  |  Page 31  |  Page 32