Harini Gopalakrishnan, Global CTO, Lifesciences, Snowflake.
In recent years, generative AI (GenAI) has become a prominent topic of discussion in life sciences.
However, I often hear from my life science clients about the need for more clarity around its actual applications and the technical paradigms. This article offers executives a high-level overview, addressing some of the most common questions I encounter.
What’s the difference between traditional AI and GenAI?
AI refers to computer systems that mimic human intelligence to perform tasks such as learning and reasoning.
Traditional AI can be described as predictive, as it primarily analyzes data to forecast outcomes or classify patterns.
GenAI is generative, focusing on creating new content based on learned data, whether it’s text, images or molecular structures. All GenAI is grounded in deep learning with certain architectural variations, and all deep learning models are based on neural networks, which was originally proposed in the 1940s.
In summary: AI > machine learning> neural networks > deep learning > GenAI.
Is all GenAI the same?
Under the hood, there are four modalities of GenAI based on their deep-learning architectural patterns. While this article provides a short discussion of each, you can dive into more details on their tech concepts in a blog I wrote on Linkedin.
The four modalities are:
1. Transformers started the GenAI revolution in 2017 with this seminal paper, “Attention is all you need.” They are the foundation of some of today’s most powerful AI models like GPT-4 and are designed for processing large, sequential datasets like text. Transformers made a huge impact in life sciences with Deep Mind’s AlphaFold, a groundbreaking way to predict protein structures based on amino acid sequences (something not text) in minutes, a task that previously took scientists months in the labs via X-ray crystallography. For this project, AlphaFold’s founders won the 2024 Nobel Prize in Chemistry.
2. Diffusion models (e.g., DALL-E 2 and Midjourney) are employed primarily for image-generation tasks. In life sciences, diffusion models have inspired innovations like MIT’s DiffDock, which predicts how drug molecules will bind to proteins—a critical step in designing new therapies.
3. Generative adversarial networks (GANs) are leveraged for synthetic data generation, including images. For instance, in clinical trials, GANs can help generate a “synthetic control arm,” providing a simulated patient group to replace or supplement real patient data.
4. Variational autoencoders are primarily used for creating embeddings (encodings) of objects. Embeddings, simply put, are numerical representations of any object, like words, images, protein sequences, etc. Another blog I wrote on LinkedIn sheds more light on its use in life sciences.
How is GenAI in life sciences unique?
While GenAI is leveraged for improving operational efficiency in most enterprises, life sciences focus more on differentiated problem-solving. The two paradigms that highlight GenAI’s usage in life sciences can be categorized as:
1. Powering Innovation In R&D Via Domain-Specific Models: Transformers and diffusion-based models like AlphaFold and DiffDock accelerate drug discovery, enabling de-novo compound generation. These applications target faster time-to-market and enhanced compliance, critical KPIs that define success in drug development.
2. Driving Optimizations And Content Generation: This relies on pre-trained transformer models with minimal fine-tuning for tasks like sales summarization, marketing content authoring, etc. GenAI copilots—generally built on architectures like GPT-4 and Claude—simplify complex tasks, reducing analyst involvement and improving productivity. These innovations enhance operational efficiency across the life sciences value chain by automating repetitive processes and reducing manual effort.
Is this real or hype?
From a financial standpoint, AI in drug discovery is expected to grow $10.93 billion by 2031. Third-party investment in AI-enabled drug discovery more than doubled annually for five years, increasing from $2.4 billion in 2020 to $5.2 billion in 2021.
AI is definitely more than hype in life sciences, with over 10 drug candidates in clinical trials that integrate some form of AI in their development. Companies leverage a combination of the GenAI frameworks mentioned above to drive these use cases across the two paradigms and some examples are below.
Powering Innovation
• Generate:Biomedicines developed Chroma, the “DALL-E 2 of biology,” to generate novel protein designs tailored by shape, size or function, with a recent billion-dollar partnership win with pharma giant Novartis.
• Recursion Pharma uses its Lowe GenAI platform to interact with multiple AI models for complex drug discovery with candidates already in clinical trials.
• Terray Therapeutics employs diffusion models to design chemicals with optimized pharmacokinetic properties, improving efficiency and regulatory success.
• Big Pharma companies like Eli Lilly, Sanofi and Moderna are collaborating with OpenAI with goals like developing treatments for drug-resistant diseases and accelerating pipeline development.
Driving Optimizations
• Pfizer, in partnership with Publicis, utilizes the Charlie GenAI platform to streamline brand content creation, editing, fact-checking and legal reviews.
• Novartis and Lilly have invested in Yesop, a company that focuses on driving efficiency in medical writing processes by automating the creation of clinical study reports, patient narratives, etc.
Which model is right for my organization?
The choice of the GenAI model depends on specific business needs and use cases.
For most data-exploration scenarios, pre-built transformer-based models are sufficient. However, specialized applications—such as AI-driven drug discovery or image analysis—may require custom models or combinations of techniques.
For example, recent innovations like Open AI’s SORA have successfully merged diffusion models with transformer architectures called DiT, enhancing capabilities in generating high-quality videos from text prompts with high precision. These can drive innovations in health like surgical training.
It’s important to note that not all use cases require building models from scratch. Applications like chatbots can often be implemented through minimal fine-tuning or by adding context at runtime using techniques like retrieval-augmented generation (RAG).
Regardless of the model, responsible AI practices are crucial. This includes ensuring model observability, referencing sources and balancing novel generation with controlled outputs. For specialized tasks, additional steps such as conditional diffusion or GANs with human oversight may be necessary to ensure practical and stable results that are grounded in reality to reduce hallucination.
Finally, there is always a human in the loop. AI can accelerate the process but, for now, can never replace a human.
Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?
