Everyone in the life science industry – from the C-suite in the boardroom to scientists at the bench – are talking about the potential of AI, especially now they’re starting to see pharma companies using the technology. And we are beginning to see researchers mixing diverse kinds of data together – images, text and numeric data to name a few – to learn from.Watching how this data interacts is helping researchers widen their learning outcomes.
But working with big data and AI is as much an art as a science. While the media continues to portray the technology as mathematical precision, it actually often requires some artistry to apply it well. This can mean we end up with the implementation of AI ranging from the good and the bad, to the ugly.
In the life science industry, we’ve seen applications from early preclinical drug discovery all the way through to selecting precision treatments for individual patients. Not only has the number of companies experimenting with and implementing AI grown (107 start-ups and 30 pharma companies use AI in drug discovery according to BenchSci), we are also starting to see successes. For example, Google’s use of AI in pathology has helped with detecting cancer.
But along with the good examples, the more widespread adoption of AI is highlighting some major pitfalls. First, the lack of data. In some instances, researchers are finding they lack the necessary data to effectively train AI systems. Second, it has been recognized that we are still very much in the experimental stage of AI in drug discovery – throwing everything at the system and seeing what successfully sticks. As the field progresses, we see how some uses of AI work better than others.
A lack of data isn’t the worst finding so far from AI in the life science industry. Recent reports of large life science AI initiatives show failure to deliver on expectations, highlighting there are still some significant pitfalls that need to be overcome in the application of AI and big data.
My colleague, Dr Jabe Wilson believes one of the greatest risks of using AI in the drug discovery and development process is the inherent biases introduced. He believes bias can be introduced at the point where data sets are chosen for training data, or built right into the chosen models themselves. The worst possible outcome from this is the development of drugs which work for one patient group but not another as a result of their genetic background.
Two good starting points for overcoming the challenge of providing usable outcomes are: implementing basic standards for data classification, as well as ensuring that data mining comes from multiple data sources. With tech companies leading the way with AI, cross-industry collaboration will be important to achieve the promise of AI in the life science industry.
Dr Matthew Clark, Director of Scientific Services at Elsevier will be presenting at R&D 100 Conference in Orlando on 15th November. You can find out more about his talk here.
You can read his piece for R&D Magazine – AI, Machine Learning & Big Data for Life Sciences: The Good, the Bad and the Ugly – here.