From AI to Impact: How GPT Applications are Accelerating Drug Discovery [Webinar]
by Elizabeth Bolitho
Causaly cofounders Yiannis Kiachopoulos and Artur Saudabayev hosted a webinar, “From AI to Impact”, in April 2023, exploring the transformative power of AI and GPT applications in early drug discovery. During this captivating session, they demonstrated how Causaly combines visual evidence with high precision knowledge graphs to answer complex research questions in mechanistic biology.
The Transformative Power of AI
We are witnessing an incredible acceleration of generative AI and machine-learning innovation rates,¹ which is transforming the drug discovery landscape. A remarkable qualitative leap in AI was achieved in November 2022 when ChatGPT was released – a chatbot with contextual understanding and the ability to hold a coherent conversation.² GPT, which is a type of a Large Language Model (LLM), uses stored parameters to generate responses, providing context in a human-readable form. At the same time, knowledge graphs have been gaining increased attention in the drug development sphere for a while, gaining momentum recently. These structured databases store and retrieve data to provide factual answers and enabling visualization of relationships.
A major advantage of GPT over knowledge graphs is that it provides detailed contextualized responses that are easily interpreted by humans, even when a user’s query is not a perfect match. In this way, GPT is highly adaptable. On the contrary, knowledge graphs, being reliant on structured data and relationships, may struggle to provide coherent answers when information is fragmented. A notable shortcoming of GPT, however, is the lack of transparency. Users are unable to determine or validate the sources used to generate responses, thereby introducing a potential for selection bias. This is problematic as it limits opportunity for the user to exercise scientific judgement, leading to potential misinformation and wrong claims. Knowledge graphs on the other hand, are typically generated through a comprehensive structured information extraction process. They can surface all evidence on a given topic and visualize it in a helpful way, preventing bias from the system.
Drug Discovery: GPTs vs. Knowledge Graphs
AI-enabled solutions, such as GPT and knowledge graphs, are vital for the acceleration of drug discovery, whether this is enhancing disease understanding or identifying and validating promising drug targets. The crucial question is: “Which technology should I use?”. The choice between GPT and knowledge graphs depends on several factors, and we have identified three primary considerations.
- 1. Nature of the research question:
- The choice of technology depends on the type of question: focused or explorative.
- GPT is suitable for focused, specific research questions, as it can provide narrow and precise answers. For example, “How is secukinumab absorbed, distributed, metabolized, and eliminated by the body?” can be answered by GPT with some precision.
- Knowledge graphs may be more suited for open-ended, explorative questions, such as “What are the potential targets of lupus?” as it allows visualization of complex data with narrowing capabilities whilst maintaining user judgement.
- 2. Exercise of scientific judgement:
- The second factor to consider with AI in the early stages of R&D depends on the desired level of scientific judgement.
- Although contextual, the user cannot determine the source or validity of statements made by GPT, thus, introducing a selection bias.
- 3. Complexity of Data:
- Both GPT and knowledge graphs are adept at handling complex data, with knowledge graphs having the advantage of network analytics and complex reasoning.
- Asking the question: “What are the side effects of drug X on liver function?”, a knowledge graph will provide a structured and concise response but may lack contextual understanding; GPT provides detailed response with context explaining the side effects, their implications and additional factors to consider.
In conclusion, both LLM and knowledge have their own set of advantages and disadvantages, which complement each other. The future belongs to those who understand how to harness the potential of both technologies. Please contact us if you would like to discuss how you can leverage this opportunity for your organization.
 Sevilla et al., 2022, Compute Trends Across Three Eras of Machine Learning, International Joint Conference of Neural Networks (IJCNN), 2022, p1-8.
 Temsah et al., Overview of Early ChatGPT’s Presence in Medical Literature: Insights From a Hybrid Literature Review by ChatGPT and Human Experts. Cureus, 2023; 15(4), e37281.