For decades, the study and development of AI was reserved to high level academics. Today, much to the amusement of statisticians and data scientists, anybody with a little bit of time on their hand can train themselves into becoming a “prompt engineer” and call themselves an expert within weeks (hours?) of their learning journey. This short article aims to raise awareness about how impactful AI can be for the pharmaceutical industry and especially for assay scientists.

Let’s start with a (summarized) timeline of the raise of AI:

• 1956: The term “Artificial Intelligence” (AI) was coined by computer scientist John McArthy.
• 1964: The first AI chatbot was developed. Funny enough for assay scientists, it was called ELIZA.
• 1980s: Skepticism won over and the fundings got cut, research slowed down.
• 1990s-2000s: With the rise of computing power, advanced machine learning models started seeing the light of day, AI started getting some spotlight again. This was very much helped when an AI bot defeated the world champion in a game of chess.
• 2010s: AI became a buzzword to describe anything with applied machine learning involved.
• 2020s: Enters ChatGPT, and the snowball effect it had on the use of Generative AI.

Since the rise of ChatGPT in late 2022, it seems as if a new AI startup is created every hour. Everyone is an expert. From those who say that “ChatGPT isn’t true AI because it doesn’t resonate” to those who believe that they can create AI agents because they can link their app to GPT, everybody and their brothers believe that they understand how it works (or doesn’t work).

Don’t’ get me wrong: Overall, Generative AI is a tremendous tool that can help us all in our daily tasks. Specifically for Bioassay. But it comes with buts:

• Machine learning models predict active compounds based on assay results and molecular features, thus reducing the need for physical testing.
  • BUT: This is not new! Bayesian statisticians have been able to do that for a while, and they are the only ones you should trust to combine all the information you have in order to generate trust-worthy simulations.
• Neural networks may analyze bioassay outcomes to predict organ-specific toxicity or adverse drug reactions.
  • BUT: AI predictions may fail to capture long-term toxic effects. Your AI is only as knowledgeable as your data!
• Automated Dose-Response Analysis
  • BUT: statisticians with bioassay knowledge are rare for a reason: it takes more than computing power to understand what is going on! I strongly suggest avoiding this one unless a statistician is leading the effort.
• Automated analysis of image-based assays
  • BUT: AI doesn’t have the life or death responsibility to decide weather a drug can enter the market. AI model validation is not something one can take lightly.
• AI integrates genomics, proteomics, and metabolomics data with bioassay results, identifying biomarkers for precision medicine.
  • BUT: Crap In = Crap Out! The job of a data scientist is not simply to run a model on your data. Omics data tends to be very messy, and AI may overlook many issues.
• Quality Control and Data Validation, AI can detect anomalies or inconsistencies in bioassay datasets to ensure reliability.
  • BUT: Many data anomalies are due to something too complex for a rule-based algorithm. Over-reliance can lead to missing subtle but important data patterns.
• AI optimizes and automates experimental workflows using robotic systems.
  • BUT: Yay automation! Unless something goes wrong, and the inaccuracies are propagated throughout the whole process.

None of these BUTs aim to discourage the use of Artificial Intelligence. The goal is simply for people to understand that there is no such thing as a free lunch. Vendors (and sometimes internal data scientists) tend to overestimate the benefits and underestimate the risks. It is natural to want to use any tool that will save time and resources. Just make sure you understand that AI is not who will have to face congress if your company sells a dangerous medicine.  

It is to be noted that the list above is not exhaustive. This article is being written in the first quarter of 2025 and considering how fast the field is evolving, it may be completely irrelevant in the near future. Stay tuned!