Best Practices for AI Prompt Engineering in Life Sciences

A prompt is a question, task, or other input given to an LLM to get a specific response from the model. Prompt engineering is the crafting of a prompt. It is both an art and a science, requiring creativity and precision.

A prompt has four basic parts: Context, Data, Task, and Format. Depending on your use case, you may not need all the different parts in your prompt, nor do they have to be in any particular order. A key part of prompt engineering is figuring out which parts of a prompt are needed and crafting them to get the response you want.

Here is an example of a prompt with the different parts color-coded:

The task here is clearly stated and then formatting instructions are given so that the model will only respond with “Positive” or “Negative”. In this prompt an example is given to provide additional context, with the text for analysis following.

Note that providing an example to the LLM is a highly effective strategy for getting the model to perform the task correctly and to respond in the way that you want. This is called one-shot prompting and is a key tool for prompt engineers.

Prompt engineering can be used for a number of different tasks. The most common tasks are summarization, extraction, and question and answer. Each type of task requires a different type of prompt and often a different ordering of the basic prompt components described above.

The examples below are just two of the AI prompting techniques available.

LLMs can be used to extract data from text. Using Certara’s tools, you can extract data from scientific papers, clinical studies, and other documents.

For an extraction-type prompt, you need to tell the model what to extract and where to extract it from.

There are several ways to improve the prompt. For example, if you include an additional formatting instruction to only return the percent, then the model would return only 10%.

For Question and Answer-type tasks, you can either ask the LLM a question directly or you can ask it a question about some provided text or data. With the direct questioning, the LLM will try to answer based on the large set of data that it was trained on. This is good for general knowledge or questions about public data but is more prone to hallucination and responses should be verified.

For example, we can ask about the pharmaceutical industry relying on publicly available data:

If you’re just getting started with prompt engineering, here are a few simple things you can do to help craft an effective prompt.

You don’t need to craft the perfect prompt on your first try. Prompt engineering is naturally an iterative process, so it can help to start with a small prompt and then build on those results until you get the response you want.

It’s also important to consider if what you’re doing is a one-time query or if you will be doing the same thing multiple times. If you’ll be reusing the prompt over and over, then it is worth the extra effort to refine the prompt. If not, maybe your first or second try is just fine.

Using phrases like

Be precise.

and

Do not make things up if you don’t know. Say “I don’t know” instead.

can help limit the amount of hallucination.

Providing extra guidance can also help get you better responses. Instead of saying

Summarize this paper.

you may get a better response by being more specific, such as

You are a smart and friendly assistant that reads scientific papers on medicinal chemistry and clinical outcomes. Summarize the drug-drug interactions found in this paper. Put each drug-drug interaction found in a separate bullet point.

It can also be helpful to repeat an instruction to the LLM but phrase it differently. For example, instead of just saying

Respond with only one word.

You can say

Respond with only one word. Do not add any additional text.

We saw an example above how providing an example of the output you are looking for, called one-shot prompting, will help the LLM. In this example, we are making a statement, but it is unclear, even to a human, what we are looking for as a response. The LLM does its best and comes up with an answer:

By providing the LLM with an example of the output we are looking for, the LLM is able to come up with the correct response:

For more complex prompts, it may be useful to provide multiple examples, such as a True and False case. This is called few-shot prompting, but the idea is the same as one-shot prompting.

You can find additional prompt engineering guidance on Certara.AI’s developers pages.

Implementing and using AI has become a top priority for many companies. Some estimates say that there is between $5-7 billion worth of value to be unlocked by AI in the life sciences². Despite this investment, 74% of companies struggle to achieve their AI goals³.

Building AI tools since 2017, Certara can help you and your company achieve its AI goals and master Prompt Engineering. Certara.AI’s CoAuthor provides an interactive environment for you to test your prompts and have the LLM extract, summarize, and glean insight from documents and scientific data.

CoAuthor is a tool designed with Medical Writers in mind. It comes with a library of pre-designed prompts to help write reports for regulatory submissions, but you can also write your own prompts and build your own prompt library.