Transcript: The emergence of generative AI for biomedicine
ASCO 2024: A high-level overview
Dr Ben Gallarda
All transcripts are created from interview footage and directly reflect the content of the interview at the time. The content is that of the speaker and is not adjusted by Medthority.
Also in the opening session, Dr. Jonathan Carlson from Microsoft gave a exciting overview of all of the ways artificial intelligence might impact cancer care in the years to come. He listed three advances in artificial intelligence models, particularly large language models, that allow them to vault far beyond what was previously possible in artificial intelligence and computer input into oncology.
The first is there are new math models of attention. Helping these models analyse large amounts of data and pay attention to the particularly important features was the first of these. The second is self supervision, where the models can train on increasingly new data and large amounts of data in a self supervised way, not requiring the input of the researchers. And finally, and perhaps most importantly, scaling laws that the larger you build the model, the larger the resources behind training the model, the better it performs. And this is simply a matter of scale.
Dr. Carlson also talked about two ways artificial intelligence, particularly generative artificial intelligence, is impacting oncology. The first is technology substitution and the second ecosystem transformation. Some examples of the technology substitution were brought up in the previous presentation by Dr. Schefter, automating repetitive tasks. Helping physicians with their workload, suggesting treatments, matching patients to clinical trials. These are all things that can be integrated rather easily into existing oncology care workflows. However, he also emphasised ecosystem transformation, perhaps designing new trials, analysing pathology images, and also exploring protein biology, both understanding why existing small molecule proteins might work to treat a given cancer.
And secondly, even suggesting new potential proteins based on deep learning analysis of protein folding and protein biology. He concluded his talk by pointing out that these new tools, these generative AI tools are going to be implemented in oncology workflows. And because they are so accessible now, because every physician in the room is able to, on their phone or laptop access one of these generative AI tools, Dr. Carlson encouraged the audience to explore these tools, find ways they might integrate into workflows, and use them to improve their management of patients.