Part three: pragmatic guidelines to getting the best out of LLMs

There have been a slew of announcements over the past few months from AI-led biotechs around the potential of Large Language Models (LLM) in early drug discovery. In the third of a three-part series, Dr Raminderpal Singh presents some pragmatic guidelines for scientists in accessing and obtaining value from LLMs.

In our previous article, published Monday 15 July, we presented a simple case example to download and practise with ChatGPT,¹ or other accessible LLM systems. In this article, we address the challenging topic of guidelines that will lead to useful scientific insights. Thank you to Nina Truter²for her support.

The guidelines are separated into two sections: those specific to using ChatGPT, and those to use when deciding on which LLM system to use. The former section allows the reader to begin now, with the recommendation not to share any confidential information on the tool. The latter section allows the reader to think ahead about an affordable, private and trustworthy system for the next couple of years.

Pragmatic guidelines when using ChatGPT

Note that many of these recommendations reflect the need to upload your own .pdf and .csv files, when using ChatGPT.

Where measurement data is used in your uploaded documents, label them as variables in your queries.
In your queries, do not embed statements within statements – bring everything out in simple single purpose statements.
Where ChatGPT stops because of long answers, break up the prompts and ask it to export to .csv or .pdf instead of printing on the screen. If it still stops part way through, type “Please continue from where you left off and finish the answer.”
When the results are inaccurate and you need to re-run the query, try switching to a new ChatGPT window/chat. Another trick is to clear short-term memory within a prompt workflow, using this command “From now on, assume [new context] without considering our previous conversation.” For example, “From now on, assume we are talking about human-only clinical trials without considering our previous conversation.”

Pragmatic guidelines when deciding what LLM system to use

LLM systems are made up of LLM models (for example, OpenAI³and Claude⁴) and software frameworks (for example, Autogen Studio⁵and Open Web UI⁶). Both these technology types are evolving rapidly, with multiple offerings. Any system you select needs to be adaptable to technology changes, for example if an effective drug discovery LLM model is offered as open-source next year.
In addition to a rapidly evolving technology landscape, it is important to understand the changing balance between commercial and open-source offerings. As with other industries, these words of wisdom stand (relatively) true: “Wait a year, and someone will offer it for free!” It is important to be agile and respond to changes in this balance.
Your LLM system may need to support both research-intensive activities, such as extracting new insights from large amounts of (.pdf) research publications, and workflow-driven tasks, such as designing a lab experiment. These are different problems that need to be solved with different approaches. A useful tip here is to design your system with a human performing the proposed LLM functions, and then swap the LLM technology in for the human once the design is ready.

The above guidelines are the tip of the iceberg and will hopefully help you get started. Please reach out to the author if you have any questions.