Cancer mutation prediction using artificial intelligence

Share :
Published: 16 Sep 2024
Views: 28
Rating:
Save
Dr Carlo Bifulco - Providence Cancer Institute, Southfield, USA

Dr Bifulco talks to ecancer at ESMO 2024 about GigaPath - an open-weight billion-parameter AI model based on a novel vision transformer architecture which can be used for cancer mutation prediction and tumour microenvironment analysis.

GigaPath excels in long-context modelling of gigapixel pathology slides, by distilling varied local pathological structures and integrating global signatures across the whole slide.

The team compared GigaPath H&E molecular prediction with competing methods HIPT, CtransPath, REMEDIS, across three tasks: lung adeno 5-gene (EGFR, FAT1, KRAS, TP53, LRP1B), pan-cancer 5-gene, and tumour mutational burden prediction.

Dr Bifulco notes that GigaPath could potentially be applicable to broader biomedical domains for efficient self-supervised learning from high-resolution images, including applications leveraging the long-context modelling features of the model to deconvolute emerging spatial biology datasets to drive a personalized and comprehensive characterization the tumour microenvironment.

To accelerate research progress in digital pathology, the team made GigaPath fully open-weight, including source code and pretrained model weights

Cancer mutation prediction using artificial intelligence

Dr Carlo Bifulco - Providence Cancer Institute, Southfield, USA

I’m presenting here at ESMO Prov-GigaPath. Prov-GigaPath is one of the very first large foundation models for pathology, a little bit of a revolution in the pathology field and in the diagnostic imaging field in general. It’s basically AI technology which is making a qualitative jump, learning lessons from the large language models that you see applied for things like GPT4 which are transforming the AI field. We are basically bringing that kind of approach into diagnostics. This enables us to do new things in AI and image analysis that couldn’t be done before and also do them in a more robust, more generalisable kind of fashion. It’s one of the very first of these big AI training efforts and there will be more to come following this initial start.

What are some of the concerns with AI?

As for everything else that we do in AI, I think it’s important to keep the human in the loop. These technologies are transformative and are going to be incredibly powerful but we see them as complementary to the human interpretation. It will also be important to keep an eye on them from a safety and security perspective. We will need to make sure that the models don’t drift so there’s going to be a need for a lot of thoughtful approach in how we implement them in the clinical space. But the promises are very significant and my expectation is it’s going to be transformational.

What might be the long term effects?

Long term I expect these kind of technologies to affect every single step of a patient journey. So we are currently focussed really on the diagnostic side and, as I mentioned before, I do expect that they will really dramatically improve both the quality of what can be done currently but also dramatically improve the things that can be done. We will be able to predict outcomes, tailor therapies, predict genomic biomarkers from images, things that are currently not typically done. I also want to mention that these kinds of technologies will be multimodal so not only will we be focussing on the pathology images but also we’ll incorporate radiology and any kind of other diagnostic imaging inside a single toolbox.

Text will also become a big part of this body of knowledge if you’re going to interact dynamically. So you’re going to be able to have conversations with the images, you’re going to be able to ask questions of the images and you will get answers from the large language models. So it’s going to impact dramatically the diagnostic space.

But also we expect that it will impact the way that the patient journey evolves and the way that the oncology actually interacts with the patient through the journey at every single step. So it will affect the therapy decision making processes, management of the patient, management of toxicities, the whole stack.

What challenges are you currently trying to overcome?

I would say that this is the beginning. So these are super-early days, we published recently in Nature, I think the paper came out in June. We’re going to have more coming out soon but many other groups are publishing very similar results so we’re getting confirmation, independent confirmation, of what we’re doing from other groups as well. It’s a very competitive arena but the cool thing is that I expect that things will evolve very rapidly.

I’ll give you a little bit of a feeling of how early we are. If you think about the large language models in terms of size and parameters that are used, if you want to get a level that is close to the early iterations of Chat GPT I think you are around 300 billion parameters. The current models are in the one billon parameters; the pathology models are currently in one billion parameters at best. So there’s a long way to go, I think they can get much better in the future.

Currently we do not know what the limitations are to this scaling. The most common opinion is that there is no clear visible limitation to the scaling of these models yet. I’m not sure that that is necessarily correct but my impression is that there is room to grow, at least in the immediate future. So things are going to get very interesting and I suspect they’re going to get very interesting very soon.