Two years ago, when Microsoft partnered with OpenAI, owning the exclusive licence of the GPT-3 language model, little did we know its full potential. A lot has changed since then. The company has been aggressively looking at utilising GPT-3, working alongside OpenAI teams to figure out possible solutions and impacts across sectors – and healthcare and life science is definitely on top of that list. 

“Medical knowledge doubles every 73 days,” said Junaid Bajwa, chief medical scientist at Microsoft, in an exclusive interaction with Analytics India Magazine, at Global AI Summit, Riyadh. 

He said it is truly about the richness of information – the data coming from publications on medical research, particularly related to ailments and treatments for various diseases and medical conditions across the globe. Looking at the rate at which research papers are published, he estimates that it has the potential to double every three days in a few years. 

In this backdrop, Bajwa believes that GPT-3 could be the answer to solving various challenges in healthcare, offering more specific, personalised, and result-backed healthcare solutions, treatments, and consultations to people across the globe. 

Bajwa is a chief medical scientist at Microsoft Research and a practising physician in the NHS (the UK’s National Health Service). Prior to this, he was the global lead for strategic alliance and solutions for the Global Digital of Excellence at Merck Sharp & Dohme. He is also a clinical associate professor at University College London. 

Bigger picture 

Bajwa thinks that GPT-3-like models could be used in creating a medical consultation/treatment recommendation platform for medical practitioners across departments. Giving an example, he said, “If a loved one of ours is in consultation, trying to manage cancer, they will have a specific history, social demographic, ethnicity, etc. But, the treatment for that person is currently based on competitive paths that happened over time.”

“What if at that moment in time, you could translate all the knowledge in the universe and say, for this person, who is 55-years old, from a South Asian background, who lives in this particular postcode, etc., this is the treatment regimen for them, these are the side effects, and understanding what works for them or not, and this is the best plan that we can go ahead with,” shared Bajwa, saying that taking that kind of information to deliver amazing insight will be truly revolutionary. 

“We are not there today, but are on the journey to make that happen,” he added.

Microsoft GPT-3 play 

Last year, Microsoft unveiled its first features in a customer product powered by GPT-3, which helps users build apps without knowing how to write computer code or formulas. 

The same year, in July 2021, GitHub-owned by Microsoft, in collaboration with OpenAI, also launched a revolutionary new feature called Copilot, an AI-based pair programmer that plugs into the editor and offers real-time coding suggestions. One year later, in June 2022, the company made it available to all developers. 

These are just a few instances of how Microsoft uses GPT-3 to unleash new possibilities. Microsoft CTO Kevin Scott said that the scope of commercial and creative potential that can be unlocked through the GPT-3 model is profound. 

Further, he said that in aiding human creativity and ingenuity in areas such as writing and composition, describing and summarising large blocks of long-form data (including code), and converting natural language to another language – the possibilities are endless. 

Bets Big on Biomedical NLP 

Bajwa said that there are a few fascinating research happening in the area of Biomedical NLP, led by Hoifung Poon, senior director of Biomedical NLP at Microsoft, where the team is looking at how they apply that to real-world evidence, different datasets, and sources to understand better what is happening in the life sciences, healthcare, etc. 

Microsoft is betting big on Biomedical NLP, where the team is looking at accelerating progress in precision medicine. Leveraging deep research assets in deep learning and biomedical machine reading, the team focuses on advancing self-supervised learning, like conducting task-agnostic biomedical language model pretraining and proposing a general framework for task-specific self-supervision. 

Over the years, Microsoft has made some exciting new progress in deep collaboration with Microsoft partners such as JAX, Providence, and now OpenAI. Some of their work in the past includes the creation of BLURB, a comprehensive benchmark and leaderboard for biomedical NLP, and PubMedBERT models. 

Meanwhile, Google is also working on creating new machine-learning tools and discovering opportunities to increase the availability and accuracy of healthcare globally. Some of its research work, particularly in NLP, includes Underspecification in Scene Description to Depiction Tasks, Retrieval-guided Counterfactual Generation for QA, Source-summary Entity Aggregation in Abstractive Summarisation, PaLI: A Scalable Approach to Joint Modeling of Language and Vision, and InnerMonologue others. 

Protein structure prediction models

Citing one of the MIT articles on ‘Analysing the potential of AlphaFold in drug discovery,’ Bajwa said that AlphaFold has given us new knowledge, a new paradigm shift. “But, the journey from what the initial discovery is, to translating into real molecules, and taking those molecules into the real world, there is still a massive journey for us to undertake,” shared Bajwa. 

He said, at Microsoft they have been working on similar things. However, they are working on an end-to-end pipeline, beyond the drug discovery aspect, where they have partnered with Novartis, Novo Nordisk, and others to apply to the real-world scientific research breakthrough and, ultimately, put it to real patients at the end of it. “Not just a paper publication, but something that delivers robust impact,” he concluded.