Exploration of GenAI in Life Sciences: Journey from Pharmaceutical Advancements to Nobel Awards Recognition
Harini Gopalakrishnan, Global Chief Technology Officer (CTO) in Life Sciences, at Snowflake.
In the realm of modern life sciences, generative AI (GenAI) has become a topic of significant interest.
I frequently hear concerns from my clients in the life sciences sector regarding the practical applications and the underlying technical concepts of GenAI. This article aims to provide an executive-level overview, addressing common inquiries.
What distinguishes conventional AI from GenAI?
AI stands for artificial intelligence, encompassing computer systems that replicate human intelligence to perform tasks such as learning and reasoning.
Traditional AI is primarily predictive, with its primary focus on examining data to foresee outcomes or categorize patterns.
On the other hand, GenAI is generative, concentrating on creating new content based on learned data, whether it's text, images, or molecular structures. All GenAI is grounded in deep learning, with variations in architecture. Deep learning models are based on neural networks, initially proposed in the 1940s.
In essence: AI > machine learning > neural networks > deep learning > GenAI.
Is all GenAI identical?
Underneath the surface, GenAI is categorized into four types based on their deep-learning architectural patterns. While this article offers a brief overview of each, you can delve deeper into their technical aspects in a LinkedIn post I authored.
The four types are:
1. Transformers were instrumental in initiating the GenAI revolution in 2017 with their groundbreaking paper, titled "Attention is all you need." They serve as the foundation for some of the most powerful AI models today, such as GPT-4, and are suited for processing large, sequential datasets like text. Transformers made a significant impact in life sciences, contributing to the development of DeepMind's AlphaFold, a revolutionary technique for predosing protein structures derived from amino acid sequences (non-textual data) in a matter of minutes, a task previously requiring scientists months using X-ray crystallography. The pioneers of AlphaFold were awarded the 2024 Nobel Prize in Chemistry.
2. Diffusion models (e.g., DALL-E 2 and Midjourney) are primarily used for generating images. In life sciences, diffusion models have sparked innovations like MIT’s DiffDock, which forecasts how drug molecules interact with proteins—a significant step in designing new therapies.
3. Generative adversarial networks (GANs) are utilized for synthetic data generation, including images. For instance, in clinical trials, GANs can generate a "synthetic control arm,” supplying a simulated patient group to replace or supplement real patient data.
4. Variational autoencoders are employed primarily for creating embeddings (encodings) of objects. Embeddings are essentially numerical representations of objects, such as words, images, or protein sequences, among others. Another LinkedIn post I wrote sheds light on its use in life sciences.
What makes GenAI distinct in life sciences?
While GenAI is employed for operational improvements in most enterprises, life sciences primarily focus on problem-solving with a unique perspective. GenAI’s contributions within life sciences can be categorized into two primary paradigms:
1. Accelerating Innovation in R&D via Domain-Specific Models: Transformers and diffusion-based models like AlphaFold and DiffDock accelerate drug discovery, leading to faster time-to-market and enhanced compliance, critical success factors in drug development.
2. Facilitating Optimizations and Content Generation: This involves utilizing pre-trained transformer models with minimal fine-tuning for tasks like sales summarization, marketing content creation, etc. GenAI copilots—typically built on architectures like GPT-4 and Claude—simplify complex tasks, reducing analyst involvement and boosting productivity. These advancements enhance operational efficiency throughout the life sciences value chain by automating repetitive processes and decreasing manual labor.
Is this a realistic or hyped-up trend?
From a financial standpoint, AI in drug discovery is projected to grow from $10.93 billion to $10.93 billion by 2031. Annual third-party investment in AI-driven drug discovery more than doubled for five consecutive years, increasing from $2.4 billion in 2020 to an astounding $5.2 billion in 2021.
GenAI is more than a mere fad in life sciences, with over 10 drug candidates in clinical trials integrating some form of AI in their development. Companies utilize a blend of the GenAI frameworks discussed above to drive these use cases across both paradigms, as highlighted below.
Accelerating Innovation
Generate:Biomedicines developed Chroma, a GenAI tool akin to "DALL-E 2 of biology,” enabling the production of custom protein designs based on characteristics like shape, size, or function. Chroma recently secured a billion-dollar partnership with global pharmaceutical giant Novartis.
Recursion Pharma employs its Lowe GenAI platform to collaborate with multiple AI models for intricate drug discovery, with candidates currently in clinical trials.
Terray Therapeutics utilizes diffusion models to create chemicals with optimized pharmacokinetic properties, improving efficiency and regulatory success.
Big Pharma companies like Eli Lilly, Sanofi, and Moderna are collaborating with OpenAI with the objectives of developing treatments for drug-resistant diseases and expediting pipeline development.
Driving Optimizations
Pfizer, in partnership with Publicis, utilizes the Charlie GenAI platform to facilitate more efficient brand content creation, editing, fact-checking, and legal reviews.
Novartis and Lilly have invested in Yesop, a company that focuses on streamlining medical writing processes by automating the creation of clinical study reports, patient narratives, etc.
The selection of GenAI models relies on particular business objectives and tasks.
Typically, pre-prepared transformer-based models cater well to most data-examination scenarios. However, for niche applications like AI-steered pharmaceutical discovery or picture interpretation, custom models or combinations of methods might be demanded.
Take, for instance, the recent advancements such as Open AI's SORA, which effectively links diffusion models with transformer structures referred to as DiT. This partnership has boosted the ability to produce top-notch videos from text commands with notable precision. Such advancements can spur developments in healthcare, for instance, in the domain of surgical training.
It's fundamental to recognise that not all tasks necessitate the development of models from scratch. Applications such as chatbots can commonly be developed through modest fine-tuning or, in real-time, by integrating context using techniques like retrieval-amplified generation (RAG).
Regardless of the employed model, responsible AI practices are indispensable. This embeds practices like monitoring the model's legitimacy, acknowledging the sources, and striking a balance between unusual generation and controlled outputs. For specialized assignments, further measures like conditional diffusion or GANs controlled by human intervention might be vital to ensure achievable and steady results that are anchored in reality, thereby minimizing hallucination.
To round up, a human figure always remains an integral part of this loop. AI might expedite the process, but, as of now, it cannot replace human expertise altogether.
Are you inviting me to join your exclusive group titled "Our Website Technology Council" for top-notch CIOs, CTOs, and technology executives?*
In the realm of using GenAI in life sciences for drug discovery, Harini Gopalakrishnan, the Global Chief Technology Officer (CTO) in Life Sciences at Snowflake, authored a LinkedIn post detailing the various types of GenAI and their applications.
In the medical writing process, Pfizer partners with Publicis to utilize the Charlie GenAI platform for efficient brand content creation, editing, fact-checking, and legal reviews, with the involvement of Harini Gopalakrishnan's expertise in technology.
