Shailesh and Alan continue their discussion on companion diagnostics and start thinking about how AI and Machine Learning are impacting new technologies and how companion diagnostics are utilized.


  • Shailesh Maingi, Founder and CEO, Kineticos


  • Alan Wright, CMO, Roche Dx

Kineticos: There are a lot of companies out there that are putting effort into AI. On the diagnosis side, and maybe on the treatment side as well, the disease is so complex we need efficient algorithms that can help physicians with things that are not obvious. There’s a limited capacity of how many variables a human being can process on both the research and treatment side. What are your thoughts about AI and the practicality of it?

AW: It is inevitable, and it will be embraced. I am an internist, and when I trained years ago, I carried calipers to measure PR intervals, QRS intervals and mapped things out.  Then, an EKG machine was introduced that interpreted EKGs. For a while, I checked the machine’s performance with my calipers, and now, I just read what the interpretation is. For tissue interpretation, the uptake of AI won’t be primarily diagnosis, and initially, there’s going to be some substantial resistance to AI interpretation for primary interpretation. There will be less resistance, by the pathology community, for AI support of the secondary reads of all these stains that will be used for precision diagnostics. AI enables significant multiplexing of staining in a tissue specimen because computers don’t get confused as the number of stains increases, and you can layer many different stains and have different algorithms running.

Kineticos: That’s a really interesting insight of the analogy that you gave. Basically, it won’t feel like computers are taking over. These will be tools that will be available to a physician and once we understand how this works, we’ll begin to trust these things more as they become better tools for us.

AW: Absolutely, we’ll incorporate it into practice. When I lecture medical students now, I ask them to hold up their calibers. Nobody has them anymore. The same thing is going to happen with AI and secondary image interpretation. There will be a lot of counting involved in percentage calculation, all of which is difficult for pathologists today.  Computers can take over some of the tedious work of analyzing biomarkers in tissue diagnostics.

The treatment of cancer is local. Every general hospital in the U.S. treats lung cancer. The idea that it is all going to be sequencing and it’s all going to tertiary referral centers, is not going to happen. If anything, treatment is going to be democratized and flow out into the community. In lung cancer, the average community pathologist might see 1-2 cases of lung cancer a month.  It’s infrequent. Often you don’t have huge, sub-specialty pathology practices. Those pathologists will certainly need support in interpretation of increasingly complex companion diagnostics. When you find out you have cancer, turnaround time and time for treatment will be a quality and marketing issue for hospitals and something that patients will increasingly insist upon. The speed of treatment will become an issue and pathology is a big component of that. You can’t be sending specimens far and wide for different types of testing.

Kineticos: Treatment and outcomes at rural hospitals are different than university, research-based hospitals, for oncology in particular. In the context of Malcom Gladwell’s 10,000 hours, you have to possess a certain amount of proficiency and see enough patients in order to really understand the nuance of a disease as complicated as cancer. It’s one word but its hundreds of diseases. I wonder if AI can help with that. The way you are talking about it in terms of these tools being available, it will democratize the treatment of it, and we can expect better care, at least in the developed areas is quite interesting.

AW: Getting back to the EKG analogy, it is all now in an 18×24 box that people don’t even acknowledge the expertise that used to be.

Kineticos: The box used to tell you that it is a long QT. The doctor used to tell you, but that is what the box said.

AW: The box can actually list the drugs that caused that.

Kineticos: AI is already part of society. When we see it more in society, we will become more comfortable with in the medical community. Both from a patient and clinician perspective.

AW: Our conversation has moved away from the actual science and technology into the community implementation of precision medicine and companion diagnostics. That is the next frontier for pharmaceutical companies. The idea of rapid deployment and day 1 readiness when the drug is approved. How many people can actually run your test? That will become much more of a strategic concern. We commonly talk to pharma companies, and there can be the assumption that tests are ambient in the marketplace. When you need a test, it will be available and with novel biomarkers for precision medicine, that is not the case. That can be a startling revelation for pharmaceutical project managers.

Kineticos: What are some interesting technologies and practices in the machine learning and AI space as it relates to precision medicine?

AW: One of the most common approaches to AI is this taking of genome and exosome sequencing and using AI either to identify a novel biomarker or several novel biomarkers and making that piece of software available. Another interesting approach is that companies are developing AI across platforms. We’ve done work in using cell phone GPS, accelerometers, or various wearable technologies to assess performance. There’s also a company that is looking at radiology to determine that the architecture of a tumor visible on a cat-scan or MRI is determined by the driving mutation of the tumor. A tumor with one driving mutation has a different architecture than a tumor of another driving mutation.

Radiology usually, almost always, precedes biochemical and tissue diagnosis. You can use those to stage and prioritize different types of diagnostic intervention. Other uses of AI would be taking complicated diagnostics and developing correlation with simpler diagnoses. For instance, using PD-L1 staining and using AI to interpret H&E and finding correlations between PD-L1 and the appearance on the slide of H&E.  There are a wide variety of uses.

Subscribe to the Kineticos Research Institute to receive weekly emails containing insightful content.