I have been learning more and more about the various AI tools, including some official coursework. Within Generative AI, I share Galina's experience that I am often using it to write background paragraphs, particularly for things that I know but either don't know good references or looking for updates to information where my knowledge may be out of date. I have found that I prefer the writing style of Perplexity over ChatGPT, but haven't tried the new voices for ChatGPT.
In addition to doing my own QC, I have also been using Google AI Studio. Within the "playground" section on left, I created "system instructions" on right to QC my paragraphs. I was pretty happy with the way it works. My first version created some fake references by mixing up names, authors, journals, and titles, but you will see in the version below that I added some instructions about not making up citations and it has been fine ever since.
I plan to try it out on the next journal article that I review. Only on the intro, however, because it does feed information back into the model, so is not appropriate to use for confidential information.
System instructions for QC paragraphs that I used in Google AI Studio playground:
You are a scientist, and you are fact-checking a summary of peer-reviewed literature on a topic. The input will be the topic, summary text, and a list of citations.
First, you will review the overall paragraph and check if the overall sentiment and conclusions are consistent with the literature citations provided as well as the broader literature. If a statement is not supported by the overall literature, then please flag for accuracy and proposed revised text.
Next, you will review the paragraph and check the citations to indicate the source for each sentence. If no provided citation supports the statement, please find an alternate citation that does .
Rules:
Citations must be from peer-reviewed literature.
Strong preference for citations to be open access.
Preference for citations to be review articles.
If the text was flagged for accuracy review, the citation review should be done on the revised text rather than the original text.
The output should be the paragraph with author and year style citation after each sentence followed by a list of references so that each sentence can be linked to relevant papers.
New citations not in the original list should be tagged "New"
Revised text should be in bold italics for easy identifications.
Do not make up citations, only provide exact citations including DOI
The list of citations should also indicate which page(s) contain the relevant facts.
Do not make up facts, each statement must be supported by literature.
Do not make up citations, only provide exact citations including DOI if available
------------------------------
Joleen White Ph.D.
AAPS 2026 National Biotechnology Conference Track Chair
Bioanalytical 101 Course Development
Senior Advisor
BioData Solutions LLC
[email protected]
Disclaimer: Opinions expressed are solely my own and do not express the views or opinions of my employer.
------------------------------
Original Message:
Sent: 10-27-2025 09:20
From: Galina Bernstein
Subject: Is Generative AI Ready for Biomarker Protocols and Regulatory Documents?
Hello
ChaptGTP is a powerfull tool and I use it regularly. It does save a lot of time of researching information.
But it was to be done very cautiously and everything you get from communication with ChatGPT has to be verified.
When I ask any question I also ask to provide references for all the responses. ChatGPT can provide a summary table for the findings and references.
It is absolutely necessary to go to each reference and make sure it contains the information provided.
I have not seen yet the articles which do not exist but I have seen references which do not contain the information that ChatGPT gave in response. For example, sometimes I get the reference to a paper for the same area of research, related to the same disease and it does report biomarkers but just not the biomarker provided by ChatGPT.
It helps to go to pubmed and search the exact wording given by ChatGPT to find an actual paper where this information was really published.
Just recently I got a reference to a hospital medical page which did not have information ChatGPT provided but since I already knew this information, using exact text as keyword I was able to find am original paper and few papers that referenced this information.
My colleagues and I were not able to find this information by direct searching because it's very rare and the other most common answer pops up in all searches first.
So ChatGPT is very useful, it can help find hard to get information and save a lot of time searching. But it has to be done very carefully, checking all the references and sometimes finding different ones which do contain the information you got from ChatGPT. And if you cannot find the reference for this information, it's better to discard it and try something else - ask different question or the same question in a different way. You can play with it and make it work for you but verify everything.
------------------------------
Galina Bernstein Ph.D.
Senior Director of Clinical Pharmacology
AzureDelta Consulting
Vaughan ON
[email protected]
Disclaimer: Opinions expressed are solely my own and do not express the views or opinions of my employer.
Original Message:
Sent: 10-24-2025 15:29
From: Jennifer Vance
Subject: Is Generative AI Ready for Biomarker Protocols and Regulatory Documents?
Partha,
I appreciate your inquiry and excited to hear the experiences of others in this area.
My early work with generative AI in this space was attempting to summarize background literature on biomarker targets. In this area, ChatGPT let me down significantly by hallucinating entirely non-existent research papers to support the drafted background summary. While this was almost 1 year ago, the experience has left me sensitive to the risks of blindly accepting anything from generative AI.
I am now focusing my efforts on using AI tools to help streamline daily tasks and free up more time for me to focus on the areas I need to be precise and rigorous. Some examples I am playing with currently are using AI to streamline sample tracking and especially generate meeting minutes that I can then provide with minimal time spent reviewing or that I can review from other meetings asynchronously without needing to attend entire meetings. These have been valuable time-savers for me without introducing scientific or regulatory risk.
Jennifer
------------------------------
Jennifer Vance Ph.D.
Dir. of Bioanalytical and Biomarker Development
Recursion Pharmaceuticals
Salt Lake City UT
[email protected]
Disclaimer: Opinions expressed are solely my own and do not express the views or opinions of my employer.
Original Message:
Sent: 09-24-2025 12:49
From: Partha Chowdhury
Subject: Is Generative AI Ready for Biomarker Protocols and Regulatory Documents?
Generative AI tools like ChatGPT and others are increasingly being used to assist with scientific writing, protocol drafting, and even regulatory documentation. In biomarker research-where precision, reproducibility, and clarity are paramount-could these tools help streamline workflows and improve consistency?
We would love to hear your thoughts and experiences:
- Have you experimented with generative AI for writing biomarker-related content?
- What benefits or limitations have you encountered?
- How do you ensure scientific accuracy and regulatory compliance when using these tools?
- Are there specific use cases (e.g., SOPs, assay protocols, IND sections) where AI has helped or failed?
------------------------------
Partha Chowdhury
Sanofi
Disclaimer: Opinions expressed are solely my own and do not express the views or opinions of my employer.
------------------------------