How generative AI is expanding what is possible in research

Historical limitation

Scientists have spent decades trying to map the structure of individual proteins based on the sequence of their amino acids. Previous approaches were trial-and-error and required multimillion-dollar specialized equipment (e.g., nuclear magnetic resonance and x-ray crystallography) to map just a handful of proteins.

AI innovation

AlphaFold, powered by Google’s DeepMind, is an AI system that can accurately predict protein structures almost instantly. It was taught by showing the sequences and structures of around 100,000 known proteins. AlphaFold has already been recognized as a solution to the grand challenge of protein-folding by CASP (Critical Assessment of protein Structure Prediction), a community for researchers to share progress on their predictions against real experimental data. The AlphaFold database now contains more than 200 million protein structures and has been accessed by over a million users in 190 countries.

This system has potentially saved the research world up to one billion years of progress—and trillions of dollars.

AlphaFold enables research that helps scientists to understand and tackle diseases that cause millions of deaths globally. A team at the University of Cambridge uses the system to search for a more effective Malaria vaccine. Researchers at the University of Colorado, Boulder are studying antibiotic resistance—a problem which causes 2.8M infections in the US alone each year.

Historical limitation

There are over 3.4 million children in America requiring speech and language services and a stark shortage of speech-language pathologists to diagnose and provide these services. Even when a child is able to see a speech pathologist, there are limits to what humans are capable of observing and capturing.

AI innovation

The AI Institute for Exceptional Education, located at the University at Buffalo (UB), is working to develop artificial intelligence (AI) systems to aid young children facing speech and/or language processing challenges. Researchers are developing two systems that use AI to capture and analyze children’s speech patterns in real time. There are initial plans to deploy prototypes to roughly 80 classrooms, impacting 480 kindergartners.

The AI Screener system will analyze video and audio streams of children’s natural interactions, capturing speech, facial expressions, and other gestures. The system will use the data it collects to produce weekly summaries of these interactions highlighting vocabulary, pronunciation, video snippets, and more. Teachers can then use these summaries to monitor students’ speech and language processing abilities and, if required, suggest a more formal evaluation with a speech-language pathologist.

The AI Orchestrator system is an app that recommends personalized content tailored to students’ needs. It will continuously monitor students’ progress and adjust lesson plans based on collected data.

Historical limitation

To protect patient privacy in medical and healthcare research, researchers are required to de-identify patient data in electronic health records, such as name, address and birthdate. This process slows important medical research due to its time-consuming and labor-intensive nature. Additionally, even after all personal information has been removed from patient records there is still a chance that the patient could be identified.

AI innovation

University of Florida Health and Nvidia partnered to develop SynGatorTron, a 5 billion-parameter language model that generates synthetic patient profiles based on health records it’s learned from. Synthetic patient data mimics real populations and can be used for medical research without the risk or privacy concerns that come with real patient data.

The model was trained on a decade of data representing more than 2 million patients. This data allows the model to create synthetic profiles that mimic real health records and can be used to augment limited data sets, collaborate across institutions, produce more data for underrepresented populations, and serve as control groups in clinical trials.