The discussion touched on several major themes involving AI and its impact on the life sciences, from training up the next generation of scientists in AI to data-sharing across systems from different vendors. The panel even proposed that competitors share data on negative studies in a dedicated journal, for which Charles Fracchia, CEO of data automation firm BioBright, proposed the tongue-in-cheek name Journal of Negative Results.
Barriers to adoption: talent, data, and legacy systems
One of the major barriers to integrating AI into the life sciences industry is finding people who have both the scientific and AI skill sets that are required. AI skills from other domains are not necessarily transferable to the life sciences, according to Fracchia.
"If you're doing -- let's say machine learning for automated driving or image-based analysis -- that's one thing," Fracchia said. "But if you're trying to apply it to drug development, discovery, and subsequent selection or speeding up the whole process, it requires a deep knowledge of that environment. We need to think about even just the entire pipeline from the very beginning on how to blend those two disciplines together."
Fracchia characterized the current availability of talent as a "very tangible shortage," adding that "obviously salaries and hiring is becoming more and more complicated."
Those best equipped to integrate AI into their experimental processes are those who understand both AI and experimentation on a "deep level," observed Josh Kangas, PhD, an assistant teaching professor in the computational biology department at Carnegie Mellon University.
"They need to understand the ways in which the choices that they make during the experimentation affect the analysis that they do in AI," Kangas said. "If you understand both aspects at a deep level, you'll be in a very strong position to implement AI."
Another barrier to the application of AI cited by the panel is the problem of legacy lab devices running outmoded technologies. Kangas noted that in these environments, AI integration is still doable but takes "some hacking to make it work."
Daniel Rines, PhD, senior director, R&D strategy, at biotechnology company Strateos, cited an example of a legacy system he worked on that was "very standalone, and they didn't give you all that data that you needed in order to be able to do the downstream learning or processing."
"It's a really important part for us to work closely with the hardware vendors," said Rines, adding that ideally one would gain access to the vendor's application programming interfaces or software to ensure successful integration with their devices.
The need for data
A third barrier addressed by the panel was data -- specifically, where to get it, how to share it, and how to know when you have enough to train a machine-learning system that will produce good results.
Fracchia shared his experience dealing with patient records, noting that if your focus is on a single disease or condition, you often need much more data than you think.
"Once you get to a high enough level of definition that you would need for a disease, even if you have the world's largest dataset, you end up with a very, very small N -- the number of records that you can actually train or classify on," Fracchia said. "You quickly realize that in some cases, you just don't have the data, and so you may need to run experiments just to acquire that data."
Kangas lamented the lack of sharing across industry and company data silos.
"I've kind of wondered if that's a business problem," Kangas said. "If you think about how to incentivize people depositing data -- because if I put data out there and that helps someone discover the next billion-dollar drug -- I'm going to be kind of bummed if I don't get anything out of that."
Hot off the press: the Journal of Negative Results?
As the panel discussed the issues around acquiring data, a side discussion ensued on the importance of data acquired from failed experiments.
"A Journal of Negative Results would be fantastic -- we need to know that things that didn't work too," said Fracchia, adding that he is not the first person to propose the idea.
"On top of that, if I can put a dream out there, [publish] it in machine-readable format," added Fracchia. "But, of course, there's a lot of infrastructure that needs to happen for that to come true."