Multimodality powered by AI

On extracting the wealth of information from data:

Digital eyes make human, expert eyes look weak. Each of these images is a treasure chest, and we're only beginning to recognize now what's in there. It turns out there's all this opportunistic medical scan capability - that is, you order a scan for one thing and there's all this other stuff that's left on the table. There is so much information there that needs to come out.

Eric Topol, Scripps Research

On maximizing therapeutic quality:

The challenge for making really high-quality therapeutic molecules today is you have to optimize along many, many dimensions. It has to be a tight binder, it has to be very stable, it can't aggregate - and the truth of the matter is you could just keep listing more and more properties that you care about. And the AI tech is incredibly enabling throughout that whole stack.

Mark DePristo, BigHat Bio

Watch: Beyond the hype: advancements in protein modeling, digital pathology, and human virtual models

On getting the right patients into clinical trials:

If we think about the oncology space and how those trials are run, even as we get better at identifying potential patients through diagnostics for trials, it then creates a new problem of how do you find those patients? That's where we can start to use large datasets, EMR, and other types of real-world data to identify those patients - especially when it's multimodal and includes clinical data, genomic data, imaging data. We can really triangulate and find those patients and then hopefully pull them into the trials.

Kate Sasser, Tempus

Watch: AI's bold leap in clinical trials: rewriting the rules

On modeling biology at larger scales:

When we think about modeling cells, compared to proteins, the scale is just enormously different. A protein is very small compared to the size of an entire cell. How can we possibly bridge those scales? There's a lot of emergent properties in biology and historically we've been very bad at modeling those. We need a lot of multimodal data captured and multimodal models - but also multi-scale models. Because if we want to understand biology, it's about the spatial scales.

Emma Lundberg, Stanford University

Watch: Beyond the hype: advancements in protein modeling, digital pathology, and human virtual models

On combining data types to identify medical deserts:

Once you have this foundational data, we could couple it with other layers of data: night light data, population density, nurse density, road maps - you name it, you can couple it, and you can start seeing things. It's actually really tricky because it's multimodal. We're using image, geospatial and numeric data all together to get to this answer.

Joan LaRovere, Virtue Foundation

Watch: AI-driven precision in global health: from predictive models to practical impact

On going beyond human imagination:

In very high dimensions when we have a lot of variables, humans are just terrible. We just do not function in high dimensions well. It's particularly relevant in drug design because the ultimate goal is personalized medicine and precision medicine, and I think we're hitting an age where the dimensionality of the data permits it but our imagination can't imagine it. What's been shocking to me is a few simple, non-linear regression techniques - that we call machine learning - are able to find all sorts of simple stuff that is way more powerful than what humans can find.

Usama Fayyad, Institute for Experimental AI

Watch: Hyper-performing our world with data and analytics

These quotes were edited for length and clarity.

Public link

Please use the above public link if you want to share this noodl on another website.