This article appears in Volume 62, Issue 2.
Photo by Steve A Johnson on Unsplash
Unless you’ve spent the last year on a wild collection trip with no outside connection you’ve likely heard of AI tools and chatbots taking over all the mundane tasks from organizing your inbox to solving the puzzles of protein folding.
As wild as the marketing for AI may seem, many of these breakthroughs are tangible improvements that can - and should - change the way we as a community should be approaching our research. The speed at which we can solve problems is no longer held back by large computational issues and instead are only seemingly limited to how we construct our hypothesis.
At the ASP annual meeting in Grand Rapids, we hosted a workshop on the introduction to coding, focused around getting natural product researchers into the world of tool-building by introducing them to Python and Application Program Interfaces (APIs). During the question/answer session, the topic of AI coding came up, and our advice was “it’s coming, but it’s not ready yet”. One year later, and those words may no longer be true.
But why spend time learning how to code when you can simply chat with one of a dozen models and have it built for you? Why spend weeks writing drafts of an article when you can use a copilot to get the findings across in less words?
The answer to both - perhaps not surprisingly - is review and oversight.
If you’ve used google in the past year, you’ve likely seen the introduction of a new tool that uses AI to parse through sources and simply provide a targeted, distilled, and often completely wrong answer. How can we expect AI to transform our research programs when we can’t rely on it to find us the correct ruleset for a game of 8-ball pool?
These AI hallucinations are not rare - and have a worrying effect on the public trust of scientific literature. Quantity, not quality, of an answer often persuades the AI overview towards a misconception. Worse, deliberate training in the wrong direction could have drastic consequences on public perception of a new therapeutic, the risk of vaccines, or which mushrooms are edible more than once.
Tool building too is a risky endeavor, if a coding agent doesn’t catch a fundamental flaw in comparing different types of data. You could spend months chasing a lead compound that was a noise peak, not filtered out because of a missed parameter, or found in a blank that was mislabeled. The risks are higher than ever, but so are the payoffs. Your expertise as scientists is invaluable.
As natural product chemists, we are often trained across the entire process from collection to characterization to commercialization. We keep a cautious eye on our cultures and double check every result along the way before we even begin to write the outline of a paper. We understand complex systems and the effects one change can have on an entire experiment.
Tools like coding agents can remove major hurdles in writing code that can do exactly what we envision at a pace that we never thought possible. The attention to detail we see reflected in our work should not be compromised by doing things faster. Instead, we need a frame shift. Writing the code is no longer an issue but understanding it will always be a challenge. Your experience in data analysis can help steer these agents towards success and away from pitfalls and false positives. So, put that to work.
In the post-code world, your metadata become invaluable, and adhering to FAIR principles is an absolute. Your data needs to be organized and laid out like you had on your first day in the laboratory- it is properly labeled, separated, and easy to access. Approach these AI tools with excitement and curiosity, but with a skeptical mind. Asking the agent “why did we do that” when you don’t understand a process can be a learning experience or a saving grace. It may cost tokens now, but it will save countless hours of frustration.
Importantly, cyber security can also be very dependent on your ability to catch a new dependency the AI agent decides to use. Openclaw – a popular “set it and forget it” kind of AI agent - became a security nightmare due to several instances of bad dependencies and automatically opening doors that should remain locked.
So, with all these cautions, why would you ever want to use an AI agent with your valuable data? The answer: if you don’t, you’re going to be left behind. We’re scientists – experiment! But do so after reading the “MSDS” and completing your “biosafety training.” Create a safe spot to experiment where your data cannot be deleted or manipulated.
These tools will undoubtedly improve, and we should spend the time now understanding how, and when, they work. Gaining experience now will have compounding effects, as new agents improve in both speed and functionality. We have always been at the forefront of new technologies, and I am excited to see what kinds of amazing utilities you will offer to the next generation of natural product scientists.
Joe’s 5 tips for tool building:
- Metadata is the key to success - The clearer your metadata are, the easier it will be to find trends you weren’t looking for.
- Check intermediate results against your expertise - if something looks too good to be true, be weary.
- Pay attention to licenses - lawyers walk amongst us, and AI doesn’t always adhere to copyright law.
- Organize your thoughts before starting - approach every session like a project proposal and you’ll be amazed at how refined your work can be.
- Start small, grow big (with datasets) - it’s more reasonable to process 10 data files and review those, rather than spend a week of computation time on junk results.
* This article was entirely written by a Human without the input of any AI.
