Beyond the Prompt: How to Actually Make GPT-5 a Game-Changer for Bioinformatics & Science
Z
Zack Saadioui
8/12/2025
Beyond the Prompt: How to Actually Make GPT-5 a Game-Changer for Bioinformatics & Science
What’s up, everyone. Let’s talk about the future. For anyone in bioinformatics, genomics, or just plain old scientific research, the buzz around AI is impossible to ignore. GPT-4 has already shown us some pretty cool tricks, but honestly, it’s still a bit of a clumsy lab assistant. It can write a decent email, but ask it to do some heavy-lifting with genomic data & you might get a confident-sounding, but totally wrong, answer.
But here’s the thing: GPT-5 is on the horizon. And the jump from GPT-3.5 to 4 was big enough that we can expect the next leap to be MASSIVE. The question isn't if it will be powerful, but how we can actually harness that power for real, complex scientific work. It's not going to be about just typing in a simple question anymore. Getting GPT-5 to work better for bioinformatics will be about how we prompt it, train it, and integrate it into our workflows.
So, let's get into the weeds. I’ve been digging into the current research, seeing what works & what doesn’t with today's models, & I've got some ideas on how to prepare for what's next. This is the insider's guide to getting ready for a truly useful AI lab partner.
The Prompt is Everything: Speaking the Right Language
First things first. The quality of your output depends almost ENTIRELY on the quality of your input. With a model as powerful as GPT-5 will be, giving it a vague prompt is like handing a Formula 1 car to someone who’s never driven & just saying "go." You need to be a skilled driver.
The research community has been quietly figuring this out. It’s not just "prompt engineering," it's about giving the model context, constraints, & a framework for "thinking."
Chain-of-Thought (CoT): Make it Think Step-by-Step
This is one of the most effective techniques, & it's surprisingly simple. Instead of asking for a final answer, you ask the model to outline its reasoning process. Turns out, forcing the AI to "show its work" dramatically improves its accuracy on complex tasks.
Bad Prompt: "Analyze this gene sequence for potential issues: [SEQUENCE]"
Good Prompt (using CoT): "You are a senior bioinformatician. I want you to analyze the following DNA sequence. First, describe the standard steps you would take to analyze a novel sequence. Then, execute those steps one by one. Identify the open reading frames, compare it against the BLAST database for homologous sequences, & finally, highlight any known mutations or variants associated with disease from the ClinVar database. Think step-by-step. Here is the sequence: [SEQUENCE]"
This approach forces the model to break down the problem, which reduces the chance of it jumping to a flawed conclusion. It's less of a black box & more of a transparent process that you can follow & verify.
Few-Shot Prompting: Give it Examples
Large Language Models (LLMs) learn from patterns. So, give them a pattern to follow. This is called "few-shot" or "one-shot" prompting. You provide one or more examples of what you want before you ask the real question.
Let's say you need to convert a bunch of gene descriptions into a specific JSON format for a database.
Good Prompt (using Few-Shot):
"I'm going to give you a gene description, & you will convert it into a JSON object.
Example 1:Input: 'TP53, tumor protein p53, is a gene that provides instructions for making a protein called tumor protein p53. This protein acts as a tumor suppressor.'
Output:
1
2
3
4
5
{
"gene_name": "TP53",
"full_name": "tumor protein p53",
"function": "Acts as a tumor suppressor"
}
Now, do the same for this description:
Input: [Your new gene description here]"
This is WAY more effective than just asking it to "summarize this gene info in JSON." You're giving it a template, which is critical for structured data tasks common in bioinformatics.
Beyond Prompts: The Game-Changing Power of Fine-Tuning
Okay, so prompting is crucial. But the REAL magic will come from fine-tuning. Out of the box, GPT-5 will be trained on the general internet. It's a brilliant generalist, but we need a specialist. Fine-tuning is the process of taking that pre-trained model & training it further on a specific, domain-relevant dataset.
Imagine a GPT that has not only read Wikipedia, but has also read every paper on PubMed, every entry in the Gene Ontology database, & every clinical trial result for a specific disease. That's the promise of fine-tuning. Studies are already showing that fine-tuning models on biomedical texts makes them significantly better at tasks like recognizing genes, diseases, & chemical names in research papers. It can even bring a general model up to the level of a purpose-built one like BioBERT.
The challenge? Fine-tuning is computationally expensive & requires a massive, clean, labeled dataset. This can be a huge barrier for smaller labs or companies. You need serious hardware & expertise.
Honestly, this is where things get interesting with new platforms. For many labs or biotech companies, building a full-on fine-tuning pipeline is overkill. This is where a solution like Arsturn comes into play. It’s a no-code platform designed to let businesses create custom AI chatbots trained on their own data. You could upload all your lab's research papers, protocols, & internal wikis, & in a few clicks, have an AI assistant that can answer questions with specific, verified knowledge from your own work. It’s a practical way to get the benefits of a specialized model without needing a team of AI developers.
Let The AI Write the Code, Not Do the Math
One of the biggest limitations of LLMs right now is their struggle with quantitative analysis. They are text prediction engines, not calculators. They can often give wrong answers to math or statistical problems, which is obviously a deal-breaker for science.
So, how do we get around this? We use the model for what it's good at: understanding & writing language... including programming languages.
The "Code Interpreter" feature (now a default part of GPT-4 & whatever comes next) is the key. You don't ask the AI to analyze the data; you ask the AI to write the Python or R code that will analyze the data.
This is a PARADIGM SHIFT. A biologist who doesn't code can now describe a complex analysis in plain English & get a functional script.
The Workflow:
Describe the Goal: "I have a CSV file named
1
expression_data.csv
with gene expression levels from two groups, 'control' & 'treatment'. I need to perform a t-test for each gene to see if the expression difference is statistically significant. Please write a Python script using the pandas & scipy libraries to do this. The output should be a new CSV file with the gene name, p-value, & adjusted p-value."
Generate the Code: The model provides the Python script.
Human Review: This is CRITICAL. The scientist—the human expert—reviews the code to ensure the logic is sound. You don't have to be a master coder to read a script & see if it's generally doing what you asked.
Execute & Analyze: You run the code & get your results, which you can then interpret.
This "human-in-the-loop" approach leverages the AI's strengths (coding, language) while keeping the human expert in control of the crucial validation step. It's about collaboration, not delegation.
You Are the Senior Researcher, GPT-5 is Your Intern
This brings us to the most important point of all. Even GPT-5 will make mistakes. It will "hallucinate" plausible-sounding nonsense. It won't understand the real-world implications of its output.
Therefore, the single most important skill for using this technology in science will be critical thinking & expert oversight. You must treat the AI as a brilliant but sometimes erratic intern.
It can draft your manuscript introduction, but YOU must verify every claim & citation.
It can summarize 20 research papers, but YOU must read the most relevant ones to catch the nuance it missed.
It can brainstorm hypotheses, but YOU must use your deep domain knowledge to decide which ones are actually worth pursuing in the lab.
The future of scientific research isn't a lab run by an AI. It's a lab where a human scientist is amplified by an AI, able to work faster, smarter, & more creatively than ever before.
Building a Custom Knowledge Hub for Your Lab
Think about all the specialized knowledge locked away in your lab's servers. SOPs, past experimental data, unpublished manuscripts, grant proposals, meeting notes. This data is priceless, but it's often hard to search & access quickly. An "off-the-shelf" GPT-5 won't know any of it.
This is another area where the idea of self-contained, custom AI becomes so powerful. You don't always need a globally-aware AI; sometimes you need an expert on your world.
This is the exact problem Arsturn is built to solve for businesses, & the application for a research or biotech setting is obvious. Imagine being able to build a no-code AI chatbot, trained securely on your lab's private data. New grad student needs to know the protocol for a Western Blot? They can just ask the chatbot. Need to find that one figure from a 2019 presentation? The chatbot can pull it up. It’s about building a conversational AI platform that can create meaningful connections with your team & your data, providing personalized, instant answers that boost productivity. This isn't about replacing a lab manager; it's about giving everyone a tool to access institutional knowledge instantly.
So, What Will GPT-5 Really Bring to the Table?
Okay, let's speculate a bit based on where the research is heading. GPT-5 won't just be a better version of GPT-4; it will likely have fundamentally new capabilities.
True Multimodality: We're already seeing the beginnings of this with GPT-4o, but imagine a model that can genuinely understand a microscopy image, a data chart, & your text-based prompt all at once. It could look at a gel electrophoresis photo & suggest what might have gone wrong based on your description of the protocol.
Better Fact-Checking: Future models are being designed with real-time data integration to check their own work against current information, which could reduce hallucinations.
Self-Improvement: Researchers are working on models that can generate their own synthetic training data to improve their performance on specific tasks. This could accelerate the development of specialized "expert" models.
But even with these advances, the core principles will stand. The most effective scientists will be the ones who master the art of prompting, understand the value of custom-trained models, use the AI as a coding partner, & NEVER turn off their own expert critical thinking.
Hope this was helpful. The next few years are going to be a wild ride, & the labs that figure out how to properly partner with AI are the ones that are going to be making the big discoveries. Let me know what you think.