Not Just Another ChatGPT: Why Life Sciences Needs Domain-Specific AI

Deep learning/artificial intelligence concepts have been around for decades. Significant advancements were made in the late ‘80s and throughout the ‘90s as scientists experimented with recurring neural networks (RNNs).

As our concept and expectations of deep learning evolved over the years, data quality and the computing power needed to train and deploy new models often hindered advancements. Graphic processing units (GPUs) have helped overcome this computational challenge, which has led to many achievements in recent years.

Google introduced transformers, or large language models (LLMs) in 2017. This technology is helping us solve some of the biggest data and analytics challenges. It’s been especially useful for analyzing complex unstructured data.

Instead of being seen as just an innovative tool for science and research, user-friendly ChatGPT brought AI into the mainstream. In fact, it’s the fastest app to reach 1 million users, which it achieved in 5 days. For context, that’s 70 days faster than Instagram and 145 days faster than Spotify.

Beyond being an incredibly powerful tool, the accessibility of ChatGPT has changed expectations for how we use AI. In many cases, people interacted with AI via complex coding languages or use-case specific apps. However, the chatbot functionality and interaction of ChatGPT made AI approachable for non-programmers. It also reinvigorated interest in using AI across industries, including life sciences.

This evolution has created a cascade of new AI models entering the market. And with each model deployed, the size of the model and the data it’s trained on grows exponentially. These changes enable the models to comprehend even greater amounts of information. For example, Llama 2 from Meta is trained on 70B parameters, and Falcon just released an 180B parameter model. These models have been trained on even greater amounts of data. Thus, they’re even more powerful than previous releases.

As these models get larger, they learn incredibly quickly and are quickly surpassing human benchmarks. It’s becoming not a matter of if, but when, using AI can enhance your work.

The growth of GPTs is incredibly promising. However, the life sciences industry continues to face challenges, with nearly 73% of companies struggling to adopt appropriate AI technologies.

One key barrier to adoption is that GPTs by nature have flaws when operating independently. For example, most generalized GPTs- including ChatGPT- are trained on broad information (in many cases Wikipedia). This broad training makes them great at suggesting attractions to visit on a vacation.

But try to do drug discovery with ChatGPT? It’ll fall short when attempting to answer complex life science questions. This is due to a data accessibility and training issue where the model has “too general” knowledge, much of which may be outdated. As a result, these features impede the successful implementation of GPTs in life sciences, particularly drug R&D.

So, to answer the question asked at the beginning of this blog, “No, ChatGPT is likely not an ideal technology for your life sciences organization.”

At Certara, we believe that combining the appropriate data architecture with specialized GPTs will solve this issue. Our solution, Certara.AI, combines GPT models pre-trained on life sciences data with a flexible data connectivity platform, enabling GPT access to organizational data in real-time.

As a result, life science organizations can securely deploy specialized AI applications that meet their rigorous standards, without having to reconfigure their existing data environment. This approach not only enables rapid, trusted AI deployments, but it also provides flexible GPT deployment across various use cases in the organization.

Certara.AI contains exciting integrations that meet specific use cases across the drug development life cycle. For example:

• Drug Discovery: Understanding a molecule’s viability for further development requires analyzing complex chemical, biological, and computational data. Certara.AI can comprehend complex data types, so drug discovery scientists spend less time running mundane computational analyses and research tasks. For example, Certara.AI’s library of property prediction models enables discovery scientists to more quickly prioritize the most viable chemical structures for their R&D tasks, allowing them to spend less time on time-consuming calculations.
• Clinical Trial Benchmarking: The success rate of clinical trials has not improved in decades, which is a huge waste of time and resources. Designing a trial with the best probability of success requires fully understanding the existing data in the field. However, 80% of the data clinical trial teams need is unstructured content that isn’t amenable to conventional analysis. When trained in biomedical literature, Certara.AI reduces manual research tasks and accelerates collecting the data teams need to design and inform successful trial execution.
• Medical Writing: Medical writers spend tremendous time manually developing document drafts for regulatory submissions. AI-generated content helps medical writers streamline the early drafting process. This allows them to spend more time on fine-tuning content and quality control, enabling more productive development of regulatory documents.

While the era of GPTs is still in its infancy, addressing its shortcomings and accelerating early deployment will be critical to long-term successful adoption. Certara.AI is a flexible solution developed by industry experts. It comprises the foundation that life science organizations need for successful AI and GPT implementation.