Last week, OpenAI officially launched the next iteration of its generative AI-based natural language model, GPT-4, the processing model that powers ChatGPT. Presently, GPT-4 is only available via ChatGPT Plus (open AI’s subscription version of ChatGPT) and it is expected to become more widely accessible.
What is new about GPT-4?
Among a raft of new capabilities, GPT-4 accepts a far greater amount of text input, helping to increase the quality and reliability of the output. The model also shows a significantly improved capability on reasoning tasks than earlier iterations based on the GPT-3 and 3.5 models. In addition, GPT-4 is capable of taking images as well as text as input. Since its release, users have been quick to test its possible applications. To date, it has been used to:
- turn a sketch on a napkin into a fully functioning webpage;
- generate social media captions;
- identify a plant by variety; and
- read a map.
Use of GPT-4 in drug discovery
With the release of GPT-4, OpenAI published its full Technical Report, outlining the developments to the technology as compared to its predecessor, GPT-3.5, as well as its potential use cases. In the Technical Report, OpenAI described a number of potential use cases which can assist in drug discovery, including:
- finding similar compounds to those that researchers are studying;
- proposing re-engineered compounds and identifying mutations that alter pathogenicity; and
- determining whether the compounds are patented.
These use cases all demonstrate the improved reasoning capabilities in the new system. Its improved capabilities are further underlined by GPT-4’s performance in various examinations, including scoring in the top 1% of test takers in the Biology Olympiad (improving on GPT-3.5’s previous score being in the lower 31%).
By way of example of GPT-4’s utility in pharmaceutical fields, the GPT-4 Technical Report provides a GPT-4-generated output (Example of Chemical Compound Similarity and Purchase Tool Use) which demonstrates the system:
- identifying compounds with similar properties to a specified medicine;
- modifying a compound to make a novel compound;
- performing a patent search to check if that compound is novel; and
- purchasing the compound from a supplier.
It is clear that generative AI will have a significant impact on the research and development of new medicines, due to the ability of the algorithms to search billions of data points in seconds. This ability to identify connections much more rapidly than the human brain should result in the far more rapid development of new treatments and will almost certainly revolutionise medicine and healthcare. This reasoning capability will lead to rapid advances, but a note of caution should be sounded. Generative pre-trained transformers remain prone to possible ‘hallucination’, producing seemingly convincing but incorrect output. Therefore, any use of GPT-4 or any similar system in life sciences fields needs to remain subject to rigorous checking (and double checking) of the output.