Jesse G Meyer PhD

It was peer reviewed. Figure 2 directly compares the vibe coded platform to a local reference pipeline using simulated data with known ground truth. That validation step was improved during review.

Thanks for sharing this. I’ve found that hands on use clarifies both the strengths and the limits. In experienced hands these tools can accelerate iteration, but they still require judgment and validation like anything else in science. I try to emphasize that balance in the paper too.

The debate around vibe coding in science is exactly what should be happening. New automation always raises real concerns. The path forward is not blind adoption or blanket rejection, but shared standards

Grateful you shared this. The intent was to spark exactly this discussion: how we can use AI to speed up tool building while keeping validation, reproducibility, and rigor front and center. I very much welcome the conversation.

Models absolutely inherit bias. That is a real concern. But this paper is not training AI to infer biology from literature. It uses an LLM as a coding assistant to scaffold a standard proteomics workflow, then validates the outputs against a reference pipeline with known ground truth.

I am a fan of judicious vibe-coding, but it requires training in a methodology to evaluate results. E.g., I pointed a wet lab PhD student to a LLM to code an excel formula for decoding mass spec composition strings. I think this is OK where you have an orthogonal method to validate your results.

What specifically don’t you like about the paper? It explains the approach to generating code, tests that the code produces correct results for a dataset with known properties, and provides the prompts, code, and data. The discussion includes caveats and areas needing further work.

in one month i vibe coded, trained, and eval'd a new family of deep learning models for de novo peptide sequencing. I applied ideas from ML preprints from Oct and Dec 2025 and achieved comparable performance to Casanovo on a single consumer-grade GPU

I have had several related experiences. Building in days huge deep learning ideas, testing many variants, and finding quickly that it's not a worthwhile. Imagine that a PhD student had instead done that manually over years and then had nothing for their dissertation

If you have tried vibe coding, you know it can be a super power for experienced coders 👇

On the "vibe coding omics analysis is here" demo paper, and some responses (run for the hills!), a thread for myself:
- we know that LLM-assisted or even driven coding is here. if you haven't tried it even in the last 3 months, you are behind
- yes it is powerful and enabling

1/7

Thanks for reading it and for your perspective

Most genius ideas seem obvious in hindsight

We are all software engineers now

No apologies needed! Just wait until you try the coding interfaces like Claude Code or antigravity! Antigravity will write thousands of lines for you if you give it clear long term goals and test definitions

We got the NOA for an MPI R01 from the NIA yesterday.

Greatful for all the collaborators, facilitators, mentors, and trainees in my group who made this possible.

Research Scoop - AI-Powered Research Paper Discovery Stay ahead of research with AI-powered paper discovery. Get weekly digests of the latest papers from PubMed, arXiv, bioRxiv, and medRxiv tailored to your interests.

It's hard to stay up to date with new literature. So many papers each week, which ones should I read?

That is why I built ReScoop.xyz - try it for free!

If you find it useful, less than the price of one Starbucks per month

From Articles to Code: On-Demand Generation of Core Algorithms from Scientific Publications AbstractMotivation. Scientific software packages impose persistent maintenance costs due to dependency churn, version incompatibilities, and bug triage, ev

Please check out our recent work published in Bioinformatics:

From Articles to Code: On-Demand Generation of Core Algorithms from Scientific Publications

url: academic.oup.com/bioinformati...

I haven't tried Claude Code personally but I have watched people in my group using it and it looks awesome. In my experience complex ideas are possible in antigravity if you define good test cases and go step by step

Research Scoop - AI-Powered Research Paper Discovery Stay ahead of research with AI-powered paper discovery. Get weekly digests of the latest papers from PubMed, arXiv, bioRxiv, and medRxiv tailored to your interests.

It's hard to stay up to date with new literature. So many papers each week, which ones should I read?

That is why I built ReScoop.xyz - try it for free!

If you find it useful, less than the price of one Starbucks per month

I suggest the free trial of antigravity, it's much better than copilot for software dev in my opinion

Thanks for your comment.

Have you tried Claude Opus 4.5, or when did you last try? Things got a lot better with the new models and agent systems like antigravity.

In my experience even much, much more complex tasks are not only possible, but nearly 100% successful

Vibe Coding Omics Data Analysis Applications Building custom data analysis platforms has traditionally required extensive software engineering expertise, limiting access for many researchers. Here, I demonstrate that modern large language models (LLMs) and autonomous coding agents can dramatically lower this barrier through a process called “vibe coding”, an iterative, conversational style of software creation where users describe goals in natural language and AI agents generate, test, and refine executable code in real time. Importantly, the goal here is not to introduce a new analysis platform. Instead, the example application illustrates that, in minutes, LLMs can now perform work that would normally require at least days of manual programming effort, lowering the cost and time investment by orders of magnitude. As a proof of concept, I used vibe coding to create a fully functional proteomics data analysis platform capable of performing standard tasks, including data normalization, differential expression testing, and volcano plot visualization. The entire application, including user interface, backend logic, and data upload pipeline, was developed in less than 10 min using only four natural language prompts, without writing any additional code by hand, at a model usage cost of under $2, not including hosting or personnel time. Previous works in this area have typically required substantial investment of personnel time from highly trained programmers, often amounting to tens of thousands of dollars in total research effort. I detail the step-by-step generation process and evaluate the resulting code’s functionality. This demonstration highlights how vibe coding enables domain experts to rapidly prototype sophisticated analytical tools, transforming the pace and accessibility of computational biology software development.

Vibe Coding Omics Data Analysis Applications pubs.acs.org/doi/10.1021/... #coding #proteomics #bioinformatics

Penicillin–Streptomycin Treatment Rewires Core Metabolic and Ribosomal Programs in HepG2 Cells Antibiotics are routinely added to mammalian cell culture media to prevent bacterial growth. However, the use of antibiotics in a cell culture can confound downstream experimental results. While genomic and transcriptomic differences between cell cultures treated with and without antibiotics are well-documented, far fewer, if any, comprehensive proteomic comparisons on the use of antibiotics in cell culture have been performed. Here, we present a study on the proteome-wide differences of culturing HepG2 cells in antibiotic (i.e., penicillin/streptomycin) and nonantibiotic-containing media. Using a longitudinal and crossover treatment study design, we analyzed 119 samples across nine passages and four conditions. On average, 9,374 proteins were detected per sample, and we identified 383 proteins that were differentially abundant between conditions. These changes included ribosomal and mitochondrial proteins, demonstrating that off-target effects of antibiotics on mammalian cells occur at the protein level. Linear mixed-effect modeling suggested that the proteomic impact of antibiotic treatment is strongest in the first passage after treatment and stabilizes after approximately three passages. Furthermore, initiating antibiotic treatment induced a greater number of differentially abundant proteins than discontinuing treatment. Lastly, we compared our results to existing literature on the use of common antibiotics in mammalian cell culture. We identified proteins and pathways conserved across studies, omics layers, and cell types. We hope that this detailed proteomic survey of the ubiquitous pencillin–streptyomcin-treated HepG2 in vitro model will aid researchers in comparing cross-study or cross-condition results from antibiotic-treated mammalian cells and inform appropriate experimental designs for the use of antibiotics in cell culture.

A wake up call for cell culture. Antibiotics have very specific effects on the proteome. Read our latest work here 🧪

Penicillin–Streptomycin Treatment Rewires Core Metabolic and Ribosomal Programs in HepG2 Cells pubs.acs.org/doi/full/10....

Vibe Coding Omics Data Analysis Applications Building custom data analysis platforms has traditionally required extensive software engineering expertise, limiting access for many researchers. Here, I demonstrate that modern large language models...

We are now in a revolution on the magnitude of the printing press. Code production will soon be 100% automated. This paper attempts to document and discuss that shift. What do you think?

Vibe Coding Omics Data Analysis Applications 🧪 pubs.acs.org/doi/full/10....

​Was it worth documenting the printing press in scientific literature, or just the first book it printed? Grateful for the dialogue—it's how we move from boilerplate to actual discovery! 🚀

However, the platform is just the proof of concept for a foundational shift in code production. The real story isn't the specific choices the AI made in this draft, but the fact that you can now simply put your own expert preferences into the prompt and have a custom, functional pipeline in minutes.

I really appreciate the engagement and the technical pushback—it's vital for refining the future of these tools. If this paper were just about the specific platform built in 10 minutes, I’d agree that debating z-scoring or imputation methods would be the core issue

Grateful for the pushback—it’s how we refine the future! 🚀

It’s a bit like debating whether the first page off a printing press was 'good' enough to publish. The real story wasn't the page itself, but the fact that the barrier to producing it had just collapsed forever. Marking that shift in the literature is exactly how we start building the new norms.