Artificial intelligence utility for drug development: ChatGPT and beyond

Abstract

Generative pretrained transformer (GPT) tools, most notably ChatGPT, are making headlines as the next revolution in artificial intelligence (AI). It affects diverse research fields, from biology and medicine to exact sciences, economics, engineering, and other knowledge-based and technology-driven fields. Novel AI applications, along with challenges, are foreseen in the fields of biomedical research and healthcare (Aubignat & Diab, 2023; Cahan & Treutlein, 2023; Cascella et al., 2023; Corsello & Santangelo, 2023; Liebrenz et al., 2023; Ruksakulpiwat et al., 2023; Sallam, 2023). Among key challenges facing biomedical researchers interested in using GPT AI tools is filtering out false or irrelevant information which limits their utility (Owens, 2023; Sanderson, 2023). Among the worrying faults of GPT-based tools is their reliance on web sources, which sometimes include fabricated or heavily-biased information that reduces the value of their generated feedback. A unique worrying fault is their “hallucinations”—a tendency to occasionally present made-up information as factual, particularly when failing to find robust data (Goddard, 2023). To improve their value for biology-inspired drug development research, GPT-based AI tools such as ChatGPT should include measures for filtering-out false or fabricated information and for prioritizing the most reliable biomedical data resources. In the context of drug development, a research field where AI tools are already applied, a potential route for overcoming such concerns is to prioritize the rich and reliable datasets curated by the National Center for Biotechnology Information (NCBI), a center operated by the National Institutes of Health's National Library of Medicine (NIH/NLM) and freely available for the global research community. NCBI tools include PubMed (https://pubmed.ncbi.nlm.nih.gov), which includes over 35 million abstracts of biomedical literature, PubMed Central (https://www.ncbi.nlm.nih.gov/pmc), which allows free access to full texts of over 8 million biomedical articles, and NCBI Gene (https://www.ncbi.nlm.nih.gov/gene/) which curates sequenced genomes from a wide range of species. All these NCBI tools are openly available to the public. Additionally, open molecular biology resources are available from the European Molecular Biology Laboratory (EMBL; https://www.embl.org/) (Thakur et al., 2023). These NCBI and EMBL tools are highly reliable compared with general media on the web and are comprehensively annotated, making them highly valuable for drug development (Morales et al., 2022; Quiñones et al., 2020; Rangwala et al., 2021). PubChem (https://pubchem.ncbi.nlm.nih.gov/) is another NCBI database highly relevant for drug development: it includes information for over 304 million substances, 35 million scientific articles, and 42 million patents, as well as cellular pathway and target protein information (Kim et al., 2022). Another useful tool is NCBI's Gene Expression Omnibus (NCBI GEO; https://www.ncbi.nlm.nih.gov/geo/) that curates many thousands of RNA expression datasets deposited by researchers (Hadley et al., 2017). However, in our chats with ChatGPT (Supporting Information: File), it mentioned mostly PubMed and PubMed Central when replying to questions on genomics, while not citing NCBI Gene, PubChem, NCBI GEO, or additional NCBI resources. Here, we examined the capacity, efficiency, and value of ChatGPT to provide information relevant to drug development. We chose to ask mostly questions requesting information relevant to the development of novel antidepressant drugs, including tentative nucleotide-based therapeutics, a topic of wide interest (Jordheim et al., 2013; Kamel et al., 2023; Li et al., 2022; Müller et al., 2017); as well as microRNA therapeutics (Carrella et al., 2022; Duan et al., 2023; Jia et al., 2023). We are aware that the examples of our questions and the ChatGPT replies shown in the Supporting Information: File represent a tiny fraction of the multifaceted aspects of drug development, or the IT capacities of ChatGPT. The examples in our Supporting Information: File may provide developers of GPT-based AI tools with insights for improving such tools for the drug development research community. The utility of ChatGPT for drug development projects awaits a true-life demonstration. Its potential for drug development is evident from some of its highly informative replies to user questions, as demonstrated by several chats carried out between March 26 and April 9, 2023 (Supporting Information: File). These chats used version GPT-3.5 of ChatGPT (see reply in chat #6 of the Supporting Information: File). The strengths and limitations observed in these chats are summarized in Table 1. As evident from this summary table, GPT-3.5 users may encounter shortcomings that limit its value for providing information useful for drug development. Of note, ChatGPT feedback depends on the phrasing of questions: it could therefore be that such limitations represent inadequate phrasing of questions, and that differently phrased questions on similar topics may produce more informative replies. It could also be that GPT-4 (which, at the time of writing this commentary, requires payment) may provide more informative and relevant information in replies to the same questions (Sanderson, 2023). According to OpenAI, “GPT-4 can solve difficult problems with greater accuracy, thanks to its broader general knowledge and problem-solving abilities” (https://openai.com/product/gpt-4; accessed on July 10, 2023). However, asking GPT-4 some of the same questions did not result in more insightful replies. Rather, it allowed longer replies, giving more examples of published research articles in reply to requests to list genes (including miRNAs) implicated in the mode of action of antidepressant drugs (not shown). Our examples of chats with GPT-3.5 about the molecular biology of depression suggest that information it selects to include it its replies comes mostly from the NCBI PubMed database. It appears to be capable of providing valuable and solid insights of relevance for drug development. However, it seems that ChatGPT can be improved by using further NCBI and EMBL resources. The quality of scientific literature curated by PubMed and additional NCBI resources is variable. In particular, in-vivo animal model studies are far more relevant for drug development projects compared with in-vitro cell culture studies, and studies reporting on clinical trials are definitely of the topmost importance. From our chats it appears that ChatGPT capacity to report and evaluate clinical trial data requires further development (see Supporting Information: File, chat #2). Developers of future GPT-based AI tools are therefore encouraged to include measures for grading the value of individual scientific articles, such as their citation rates, the publication of independent confirmation studies, and clinical trial findings. This information, according to our chats with ChatGPT, is not yet available at the time of writing this commentary. Another obvious weakness is that, as reported in reply to a question about the NCBI resources used by ChatGPT (Supporting Information: File, Question #3.1), “ChatGPT's access to these resources is limited to the information that was available in 2021, which is the cutoff of its knowledge.” Efforts are required to keep the information gained from NCBI resources as updated as possible. In conclusion, GPT-based AI tools have great potential to facilitate the complex, expensive, and time-consuming drug development process by identifying the most promising drug targets and potential lead compounds (Tiwari et al., 2023; Turon et al., 2023). Improved GPT-based tools will be coming, and the pharma R&D sector will need to identify and prioritize the optimal modes for using them. Noam Shomron and David Gurwitz were supported by the Israel Ministry of Science grant number 3-17928 in the framework of the ERA PerMed project “Artificial intelligence for personalized medicine in depression (ArtiPro).” The authors thank Gal Alon (Center for Innovation in Teaching and Learning, Tel Aviv University, Tel Aviv, Israel) for his generous help with ChatGPT-4 chats. The authors declare no conflict of interest. The data that support the findings of this study are available from the corresponding author upon reasonable request. Please note: The publisher is not responsible for the content or functionality of any supporting information supplied by the authors. Any queries (other than missing content) should be directed to the corresponding author for the article.

Ähnliche Arbeiten