Englander Institute for Precision Medicine’s Director Olivier Elemento, Ph.D. on Artificial Intelligence and Precision Medicine. #AIWeek
How is AI used at the EIPM?
We are using AI for many different types of basic research and clinical projects, such as assessing genomes to better understand the functions of mutations. For basic research, for example, we use AI models to assess the mutations we see when we sequence someone’s DNA. We can now leveraging tools like AlphaFold (an AI system developed by Google DeepMind that predicts a protein’s 3D structure from its amino acid sequence) to be able to understand if the mutation is creating a new protein structure that may be exploitable to treat patients.
We’re using AI to assess images and videos in radiology and pathology. AI is well-suited for this work by segmenting, comparing, and analyzing the structure and architecture of images. We have been using AI to assess data from the fertility clinic here at Weill Cornell Medicine, to better understand and predict success with in vitro fertilization based on the images of embryos.
We use AI to analyze tissue at single cell resolution. We have technologies that allow us to assess what I call the battlefield of disease, where we can see individual cells battling each other and to better understand what they do in real time. We use AI to unravel the complex ecosystem of disease and understand the communication between cells, with the goal of disrupting communications and destabilizing disease cells.
We are also integrating AI into reading medical records data, while also protecting patient privacy. We are now working to re-structure the way medical information is gathered so it can be more searchable and that patterns in disease and transmission that might have been missed by the human eye can be quickly and accurately identified and analyzed by AI.
So, AI is really being used broadly across the entire research enterprise at Englander Institute for Precision Medicine and is really giving us great results both in basic research and in the clinic with patients. There is so much potential with AI.
Do you have concerns about relying on AI for research?
Bias is the major concern for AI, mostly stemming from the fact that a lot of the data used for training AI models has been collected in the context of clinical care, which can underrepresent or overrepresent certain populations of patients.
There can be some bias attached to the data that is potentially preserved in the context of the AI model--essentially the AI model is learning the biases of the data that it’s being trained upon. We must be very careful about this because we want to create AI models that are equitable in terms of the information that they provide to patients and their risk predictions.
We are involved in multiple projects, some funded by the National Institutes of Health like the Bridge2AI project, which is trying to rethink how data is collected for AI by making the data as unbiased as possible at the point of collection. When we collect the data, we’re trying to make sure the biases don’t apply by selecting patient populations, for example, that are representative of the communities and the patients we care for. This approach could potentially correct for the biases, but really the critical part is to make sure there is no bias in the AI models.
Can you talk more about how AI is being used at the EIPM in reproductive health to help couples have healthy babies?
Yes, our colleague Dr. Iman Hajirasouliha had some success in using AI for assessing IVF imaging data. Iman and his colleagues have come up with AI models to assess videos and images from embryos, in collaboration with the fertility clinic at Weill Cornell Medicine, to make a prediction about quality of the embryos at fertilization. The goal would be to optimize IVF treatment. This is a very exciting collaboration we have with the fertility clinic, and one of the most advanced and important AI-related projects that we work on. The fertility clinic has a lot of image data because they do so many IVF cycles, so it’s really a project that was very ripe for using AI. This work has already been a great success.
Where do you see AI moving in the future?
The critical next part is to fill the gap between the research lab and the clinic by assessing the clinical utility of AI models. To do this we have to invest in randomized clinical trials to assess the real-world context of the clinic, the utility of AI, how it can help the clinicians be more successful and productive and improve outcomes.
We need to do this work to be able to demonstrate how helpful the AI models are in the clinic. This is what I’d love the EIPM to be more closely involved in, to promote the idea of randomized clinical trials for AI. Almost thinking of AI as a way to treat patients, in the same way medications are treatment vehicles for patients. We need to think in the same way for AI: What is the utility of AI, how does it help clinicians and patients? It’s only when you have this information in hand can you make the case that AI should be used in the clinic on a routine basis.
Is there a specific instance of AI proving potentially very useful that has surprised you?
One thing that AI has become good at is folding proteins from sequence. This is quite remarkable and critically important. Scientists thought, before using AI, that it might take another 100 years from sequencing of amino acids and proteins to folding them in three dimensions. Researchers were trying to simulate physics to do the folding. But a very clever group of Google engineers showed that by using deep learning AI models you could accurately predict the folding of these proteins from sequencing! This significant advance didn’t take 100 years, it is quite feasible now using AI. And this advance opens all kinds of possibilities in terms of folding proteins with very specific mutations to help us better understand cancer mutations, for example, which are very often modifying sequences of genes. And now we can use AI to better understand the real consequences of having those mutations.
At its core, the gift of AI is its ability to look at multiple different signals, combine them in smart ways, and produce predictions driven by the data that are effective in the success of therapies or treatment procedures for patients.
# # #