ASCO 2026 | AI-predicted NK cell fitness to guide dara & transplant strategies in newly diagnosed myeloma

To address that question, we’ll have to go back earlier to another question which we tried to address. Can we predict subtypes within the myeloma space using histopathology images? Now, there’s plenty of histopathology images sitting in pathology labs across the world, which has been untapped. And the community is aware of this for a few years now. What we tried to address was what we published in ASH last year, American Society of Hematology, which is to say, could we predict cytogenetics using these images? And what we saw was that, you know, these images actually offer a lot of information where they actually look at heterogeneity across many, many different slides. So we were able to look at many different subtypes, and we found 12 such subtypes, which was incredible. 

To give you an idea of what that was, there was an earlier model which we published in JCO in 2024, where we took NGS, like next-generation sequencing samples, and fed it into a model, which we called IRMA at the time, to predict individualized risk for patients. So that was the first attempt of using machine learning and AI to actually give individual predictions to patients. We found 12 subtypes there as well, which is what fascinated us towards moving in the direction of histopathology images because it’s ubiquitous, it’s resource-less intensive, and many more centers can use that. 

Now, we actually addressed the question of finding, you know, how we could actually tackle the other question of predicting response to therapies. Now, obviously, daratumumab has been extensively used in the community for the last some time. And transplant has also been a question of how do you give it to the right patients? Now, we actually tried to address that right after ASH, where we said there has to be a way that these images can offer insights into certain immune profiles that probably would have been unrecognized by a human eye. So we found that obviously CD16 is very, very commonly tested for daratumumab because it goes hand-in-hand with the immune response. So we tried to see if the model could predict that pretty much based on a picture, right? 

And what we found was that, obviously, yes, the model is capable of finding response to daratumumab in patients where they should get it, if they have high fitness in their immune system or even low, actually. Transplant or no transplant did not matter for those patients. But if you look at VRD without daratumumab, obviously there, if you were high immune versus low immune, clearly mattered according to the predictions of the model. So that’s when we recognize that just with a single picture has a lot of information and richness to actually predict who will eventually do better. This was a question we tried to address 10 years ago, but it’s taken us 10 years to get here. 10 years ago at Sloan Kettering, Memorial Sloan Kettering Cancer Center, we actually tried to build a model just to predict cancer or no cancer. And that’s what we were trying to get, like, a 90-plus percent accuracy. It took us some time to actually break through that, and our group published in Nature Medicine. I think that was a revolutionary paper, which started driving the whole field forward and researchers started pouring their time, money, data into it because it clearly worked. Now, no cancer or cancer to then subtypes. The field then moved towards subtypes. And now to precision medicine. This is the story that I think a lot of the folks in the community have been trying to see for the last 10 years and we’re at a point now where it’s actually coming to reality. This is one such work. I’m sure many people will take to it. And I hope the community looks in the direction of trying to not just build another model, but utilize the resources that’s already available sitting in labs to actually improve the patient care. Because everyone in the cancer space is talking about how AI can be beneficial to them. You know, I think this conference this time around ASCO is actually putting AI as one of the top important priorities for researchers as well. So now is the time where I think as a community, we can actually take this question and use the resources we have, not to necessarily displace or maybe, you know, like replace anything, but to add more information into the knowledge graph or the knowledge system that people themselves are trying to build.

This transcript is AI-generated. While we strive for accuracy, please verify this copy with the video.