First published in The Pharma Letter on July 4, 2024
Getting the most out of AI foundation models: tips for pharmaceutical companies
By Brendan Frey, Founder and Chief Innovation Officer of Deep Genomics
We’ve all heard of and used ChatGPT. But what if we could apply this to pharmaceutical R&D? To effectively use AI for business and patient outcomes, a fundamental technological, cultural and organizational shift is required, says Deep Genomics Founder and Chief Innovation Officer, Brendan Frey.
Most previous efforts to use artificial intelligence (AI) for drug discovery have fallen flat, but in the past two years there have been radical advancements in AI that could transform the pharmaceutical industry. ‘AI foundation models’ have demonstrated potential for accelerating drug discovery, optimizing clinical trials, and even predicting possible adverse effects.
Several AI-driven drug candidates – in areas ranging from cancer to dermatology – have failed to make it into commercial production. Some experts are skeptical, saying that this reality has cast a shadow over the true capabilities of AI in drug discovery. What the headlines fail to capture, however, is a significant change occurring behind the scenes – a fundamental shift toward a new kind of AI.
Most previous uses of AI are narrowly focused on a single, isolated task. Now, companies are embracing a new paradigm: AI foundation models. These versatile models are trained on broad datasets containing billions or even trillions of datapoints, and are capable of tackling a wide range of complex challenges simultaneously.
To leverage AI foundation models for positive and impactful change, here are three things pharmaceutical companies need to know:
Promises and pitfalls
As the pharmaceutical industry grapples with AI promises and pitfalls, understanding the power and potential of AI foundation models is crucial. The first lesson? This is no Chat-GPT plug-in. It cannot be overstated how truly revolutionary the shift will be from the way we used to utilize AI to the way we see foundation models being used, just ahead on the horizon.
Imagine searching for gold, but you’re only allowed to dig at one location, in one direction and five feet deep. This is how the industry started with AI, focusing AI models on narrow questions. Now imagine being able to search an entire region for gold, at any location, in any direction and at any depth. This is what moving to an AI foundation model feels like.
At Deep Genomics, our focus is on genome biology. Shifting to an AI foundation model meant moving from 40 separate, specialized AI models that answered narrow questions, to one comprehensive model that understands a broad swath of genome biology, including molecular and cellular biology. This foundation model realizes synergies between the narrower tasks and as a result becomes much more powerful at all of them. Also, whereas it was not possible to scale up the 40 separate models to tackle bigger datasets and task-specific datasets, such as patient data, scaling up the foundation model is relatively straightforward.
Another example of foundation models is those that have been deployed for predicting protein folding, where the industry has unleashed a ton of innovation.
Genome biology is very complex and compared to other areas, succeeding will require more compute, more data, and more experienced people. This introduces an important lesson: your people don’t need to speak the same language, but they do need to understand each other.
The success in collaboration: Multilingualism
Effectively leveraging AI foundation models to reap the benefits requires more than just implementing the technology itself. Truly disruptive technologies demand a foundational shift in an organization’s culture, weaving cross-disciplinary collaboration into the fabric of the business.
In the past, both pharmaceutical companies attempting to leverage AI and AI companies attempting to apply their models to drug design have faced significant challenges. These setbacks are rooted in a failure to bridge the cultural and language barriers between the disparate disciplines involved. Adopting AI foundation models requires a fundamental mindset shift that fosters seamless collaboration across domains. At the core of this transformation is the need for "multilingualism" – the ability to facilitate effective dialogue between experimental biologists and AI researchers.
Multilingualism has to be a core value, recognizing that true breakthroughs in AI-driven drug discovery can only be achieved when researchers transcend the silos of their respective fields. This new generation of "amphibious" scientists, able to operate fluidly in both the wet labs where tissue samples are tested and the dry labs where AI researchers work, is pivotal for building a successful company in this area.
For example, whereas the tech sector has a rich history of collaboration and open sourcing, this ethos is relatively new to the pharmaceutical industry. By open sourcing GenomeKit, a set of tools for accessing and manipulating genomes, we are participating in a broader cultural shift towards openness that we believe can raise the tide for the industry. Companies adopting disruptive AI technologies like foundation models must be willing to open up, collaborate across disciplines, and crowdsource solutions – ultimately bringing vital new medicines to patients faster.
A new era of TechBios
Merely bolting AI capabilities onto existing pharmaceutical processes is insufficient. Generating the wide and general outputs that AI foundation models promise requires a level of deep technological integration and true collaboration that cannot be achieved through traditional approaches.
Strategic partnerships between traditional pharmaceutical companies and a new, emerging class of "TechBios” will steer drug discovery toward a better future for pharma. Built from the ground up with AI as the driving force, these TechBios enable seamless integration and collaboration between biologists, chemists, and AI researchers, based on a complex software technology stack, often involving many dozens of different databases and pipelines that ensure correctness, reliability, reproducibility and scalability.
This approach calls for a unique kind of partnership that brings together the AI capabilities of technology-driven organizations and the domain expertise of pharmaceutical companies.
‘Artificial intelligence’ is a buzz word that has almost become white noise in the backdrop of every job, sector, and industry. However, it’s worth taking the time to understand how applying this new type of AI, the foundation model, to drug discovery points to a true paradigm shift that will impact pharmaceutical R&D more profoundly than any AI advancement that has come before it.