Bessemer’s Morgan Cheatham: Generative AI Will Unlock $1T in Value for Healthcare and Life Sciences
ABSTRACT
KEY POINTS FROM MORGAN CHEATHAM'S POV
We’ve been pontificating about AI in healthcare and life sciences for a long time. What is different now?
- With the explosion of biological and clinical data, the opportunity for disruption remains large. Along with an uptick in data, sequencing costs have decreased 10-fold since 2015, and regulations like the HITECH Act have digitized more than 90% of clinico-omic data. Companies are also focused on developing generative AI frameworks, like Deepmind’s AlphaFold, which performs predictions of protein structure.
What are the applications or use cases that might be attached to this category?
- “Generative models will tackle challenges across the healthcare and life sciences value-chain, spanning drug discovery to back-office automation,” says Cheatham. Areas include:
- Drug discovery: Generative AI enables methods for exploring the chemical and biological space more efficiently, and in some cases, may yield results in a much quicker fashion. (See chart, below.)
- Prognostics: Leveraging clinical data can improve predictions for disease progression and unlock increased options for care and prevention
- Synthetic data: Generating new data based on an underlying distributions will help maintain privacy and improve regulatory compliance
- Workflow automation: Numerous tools can improve provider efficiency by increasing output and improving data record quality. This includes copywriting or AI-powered writing assistants, and video content generation for medical or patient education.
- Workforce enablement
What are some of the potential roadblocks?
- Strict regulatory requirements and liability concerns weigh on go-to-market models for the majority of use cases. “For computational diagnostics, the regulatory pathway to enter the clinic remains a long, arduous, and costly road for most startup companies,” says Cheatham.
- Providers will be hesitant to trust and implement complex models that lack established success. “For generative AI-powered workflow automation, establishing efficient distribution methods and building provider trust with models that may not be fully explainable will pose challenges.”
VISUAL: AI-FIRST DRUG DISCOVERY GENERATING TANGIBLE RESULTS
IN THE INVESTOR’S OWN WORDS
Generative models represent a recent breakthrough in machine learning research whereby algorithms can utilize existing data to create new, plausible data — a departure from canonical AI approaches that produce analyses of existing information (e.g., prediction, classification). We now have high-quality, cheap, fast AI models for generating text, images, videos, software code, music, voice, 3D models, and more.
Big tech has played a critical role in developing generative model frameworks including OpenAI’s GPT-3, a deep learning model that can generate human-like text, and DALL-E, which can create realistic images from a natural language description, as well as life sciences-specific tools like Deepmind’s AlphaFold, which performs predictions of protein structure. Burgeoning open-source communities are forming around many of these tools — along with foundation and task-specific model landscapes — as access to compute increases, even as costs subside.
Coinciding with the rapid advancement of generative models is the explosion of biological and clinical data and increasing clinico-omic data liquidity. The cost of sequencing has decreased by more than 10x since 2015 and chemical and biological data has become digitized. In parallel, regulations such as the HITECH Act have digitized more than 90% of clinical data, and more recently, the 21st Century Cures Act has mandated and incentivized increasing healthcare data interoperability across stakeholders.
MORE Q&A
Q: What do other market participants or observers misunderstand about these categories?
A: It’s easy to underestimate how essential quality data will be in order to create equitable outcomes and insights. As generative AI methods find healthcare and life sciences applications, it’s essential that we partner with leaders in AI ethics and biosecurity to ensure appropriate and equitable use.
For example, we have well-established documentation on the underrepresentation of underserved and minority populations in clinical trials. Though increasing diversity in clinical trials has become a key focus area for many biopharma companies, generative AI models trained on incomplete datasets could perpetuate bias or underperform. Similarly, like any breakthrough technology, these models could be used to generate biological or chemical products that pose risk to society and therefore require collaboration with the biosecurity field.
WHAT ELSE TO WATCH FOR
Rather than selling into larger organizations, players opting for direct-to-clinician, product-led growth will place AI workflow solutions directly in front of increasingly tech-hungry end users. “I'm compelled by direct-to-consumer models for scientists, clinicians, and other specialists in the healthcare and life sciences industries,” says Cheatham. “We’re on the cusp of a major shift in how healthcare and life sciences software is built and sold. Increasingly, individual workers such as clinicians and scientists are adopting consumer-grade technology through bottoms-up motions. Healthcare and life sciences tech companies are leveraging product-led growth strategies first pioneered by cloud giants.”