ISAL 2025 | Retrospective validation of a computational method for assessing MRD in AML

We presented a poster at this meeting, which continues a publication of ours in Communication Medicine of the Nature Springer catalogue. In the publication we actually propose a methodology to achieve a computational approach to the analysis of flow cytometry-based measurable residual disease assessment. And what we did in that publication is we developed a method of which we think it’s important that it’s an explainable method. So it allows in the end, if we make a prediction on measurable residual disease in a patient, it will allow us to go from the predicted results back to actually the original data. So along the process we really know very well why a patient is predicted to be MRD positive or MRD negative. So the methodology works as follows. So we have actually a quite small set of control samples as a reference. And the importance of this control set is that these are healthy samples and a mixture of regenerating bone marrow samples. So samples that have been exposed to chemotherapy, but are otherwise free of leukemia. And it’s important because these samples provide a background on the immunophenotypes that are assessed by flow cytometry and allow us to discover cells with aberrant protein expression – the leukemic cells. So we use those references, those control references, to actually learn the algorithm what’s normal or regenerating, so without leukemia. And if you do that you can estimate the probability of a cell to be either normal, so close to that reference, or far away from that reference more likely to be leukemic. The advantage of doing this approach is that it overcomes two issues that we had when automating measurable residual disease assessment in AML and that is one, the disease is very heterogeneous so it would be very difficult and require a lot of data to train models to learn leukemia features and therefore it’s more easy to learn what deviates from a control. And the second aspect is of course that in MRD we should be able to assess rare disease cells, meaning that after therapy most of the leukemic cells have been declined and only a very small proportion of cells reside and that’s the MRD that will in the end may cause the relapse. So the rare cell aspect is also important because rare cells don’t form clusters. So for us it was very important to develop a method that can actually classify not at the sample level but at the cell level. So we classify each cell as being leukemic or non-leukemic. So having done that we showed actually that we can obtain a very high concordance with the expert-based MRD assessment of flow cytometry which is sort of like the gold standard right now. And we achieved a very high concordance, so 97% of the cases that we did actually were in concordance with manual analysis. And in the cases that it wasn’t the case, those were very special acute myeloid leukemia samples with, for example, CD45 negativity. So that’s what we did in the publication but then of course the question is how well does this perform in clinical practice. So we did a retrospective validation on the HOVON132 flow cytometry MRD data and we show in the poster that we actually have a much stronger outcome association of the computationally predicted MRD when compared to the expert-based MRD. And that was not necessarily what we expected because we thought in advance that it would be very complex to achieve this computational approach. So we were very happy now to have seen that actually both in terms of relapse-free survival, cumulative incidence of relapse and overall survival, that we do better doing this computationally than expert-based. Now why is that important? Flow cytometry-based MRD assessment is really a complex thing. It requires a lot of knowledge on hematopoiesis and to recognize the many different aberrant immunophenotypes in the flow cytometry data really requires a high degree of expertise and that’s why some people say it’s like flipping a coin – not every laboratory can build this expertise. So, because we… And then comes another advantage of having to use a very low amount of control samples to train the algorithm. This means that it’s also more easy to transfer this to other laboratories that maybe use other panels when compared to what we do. Allowing it to be broadly implemented. And for us this is a next step in further standardization of MRD assessment. So a lot of efforts have been put by the European LeukemiaNet, collaborators of ours, to standardize and harmonize the MRD assays. And this will allow us also to standardize the analysis of the data. So, meaning that also when laboratories have less expertise, we know that the algorithm works, we are also going to validate it prospectively and other types of data, but it means that other people can also use it without the requirement of that extensive expertise. So, yeah, this is basically the main result of the poster that we presented in this conference. And I think a very important step in automating MRD assessment.

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