Biomedicine in the Age of Generative AI

Summary

Dr. James Zou from Stanford University presents on using generative AI to advance biomedicine across the research pipeline. He demonstrates generating novel antibiotic drug candidates with AI and testing successful syntheses against dangerous pathogens. He then shows optimizing clinical trial eligibility with an AI system to enable more inclusive, efficient trials. Finally, he presents a pathology image chatbot to assist clinician diagnosis. Throughout, he stresses the need to quantify and mitigate biases in medical AI systems.

Takeaways

• 😀 Generative AI models like GPT can be used to help explore complex design spaces like drug molecules, resulting in new discoveries.
• 👍 Applying generative AI to guide molecular design led to the creation and validation of 6 new antibiotic drug candidates.
• 💡 Using generative AI models to computationally emulate clinical trials allows more rapid optimization of trial eligibility criteria.
• 📈 Optimized clinical trial designs from the AI system enabled enrolling over 2x more patients while maintaining efficacy.
• 👍 The optimized trials also enabled more enrollment of women, minorities and elderly without harm, increasing diversity.
• 🤖 AI-powered chatbots can provide interpretive assistance to pathologists, improving diagnostic accuracy.
• 🔍 Large collections of pathology images and clinician discussions from social media provide rich training data.
• 😕 Generative AI models still exhibit harmful biases and stereotyping that need to be addressed.
• 💡 Generative AI provides powerful tools for design and exploration, but human oversight is still critical.
• 🎓 These advances create opportunities to rethink biomedical research education and leverage AI tools responsibly.

Q & A

What is the main goal of the biomedical data science seminar series?

The main goal is to bring together AI platform developers and university educators to better understand the power, promise, and potential pitfalls of the generative AI explosion in biomedicine.

How does the synIO algorithm work to discover new antibiotic molecules?

It uses a tree search algorithm to explore the space of potential molecules, scoring each one based on properties like efficacy, safety, novelty and ease of synthesis. It generates explicit recipes to synthesize the top-scoring molecules for experimental validation.

How does Trial Pathfinder use AI to design more inclusive clinical trials?

It computationally emulates millions of clinical trials using real-world EHR data to evaluate the impact of different eligibility criteria. It aggregates this information to provide data-driven suggestions to relax restrictive criteria and enroll more diverse participants safely.

What is the purpose of the Pathology Image Language Model (PLIP)?

PLIP is an AI chatbot that can describe pathology images and answer clinician questions to assist with diagnosis and treatment planning. It can also retrieve similar textbook/EHR cases to provide additional interpretive context.

What are some key ethical considerations around generative AI tools?

We must be aware of and mitigate issues like stereotyping, bias, and potential model hallucination. Responsible use in education and clear human oversight are important.

Why is discovering new antibiotics so crucial nowadays?

Antibiotic resistance is a growing crisis causing over a million deaths annually. Very few new classes of antibiotics have been introduced in decades, so pathogens are catching up to existing drugs.

How did the authors validate the antibiotic molecules designed by synIO?

They synthesized 58 of the 70 proposed molecules and tested them against 4 dangerous 'ESCAPE' pathogens in the lab. Six were highly effective at low dosages with minimal observed toxicity or resistance.

How are generative AI models different from traditional machine learning?

Instead of making simple predictions, generative models produce highly complex and rich outputs like images, text, code, and molecules. This allows for expansive design space exploration.

Why is clinical trial enrollment an important challenge to address?

Overly restrictive eligibility leaves out many patients needing treatment. This slows recruitment, limits generalizability, and reduces access to the latest therapies.

Where does the PLIP pathology image dataset come from originally?

It curates hundreds of thousands of high quality pathology image Twitter posts where clinicians openly discuss challenging case images to crowdsource expertise.