What a Drug-Discovery Lab Can Teach Designers About Building With AI

César de la Fuente's lab is using Codex and ChatGPT to scan vast biological databases for molecules that could fight drug-resistant microbes, compressing a search that once took years into a matter of hours. That is a long way from designing a habit-tracker app. But the way this team actually uses AI tools day-to-day maps almost perfectly onto what we keep seeing work for designers who are just starting to build.

The real story is not the science, it is the workflow

De la Fuente runs what he calls a "transdisciplinary" lab: biologists, chemists, computer scientists, and engineers all in one room, all working on the same problem, all with different skill gaps. Sound familiar? That is basically any product team where a designer is now expected to ship code.

Here is what matters: Codex and ChatGPT are not replacing any of these researchers. They are acting as a bridge between people who know biology but not programming, and people who know programming but not biology. A biologist can describe what they need, and Codex helps them build the script. A programmer can ask ChatGPT to explain a protein-folding concept so they can write better analysis code.

For designers, swap the terms. You know layout, hierarchy, user flows, and how people feel when software gets in their way. You may not know how to wire up a database or structure a fetch request. Codex is the bridge. You bring the domain knowledge; the tool handles the translation.

What "brainstorming partner" actually looks like in practice

De la Fuente uses ChatGPT specifically as a hypothesis-shaping tool before going deep on any experiment. He feeds it good ideas and bad ideas, treats it as a sounding board, and uses it to pull together thinking from across different fields.

This is a concrete, repeatable move for any designer building an app:

Before you open your vibe-coding tool, open ChatGPT first. Describe the problem you are trying to solve, the kind of user you are designing for, and the feature you are considering. Ask it to poke holes in your logic, suggest alternative approaches, or explain a technical concept you keep running into. Then take that clarified thinking into Cursor, Replit, or whatever environment you are building in.

De la Fuente's team also uses ChatGPT to let members work in their native languages, which lowers friction and keeps people focused on the actual thinking rather than on translating. If English is not your first language, or if technical jargon keeps slowing you down, this is a legitimate and underused feature. You can ask ChatGPT to explain a Codex error in plain language, in whatever language you think best in.

Using Codex for the "download and organize" layer

One of the more mundane but genuinely useful things the lab does with Codex is downloading, organizing, and pre-processing large datasets. This is not glamorous, but it is exactly where non-engineers tend to get stuck.

In a design-to-build workflow, your equivalent might be: pulling data from a public API, reformatting a CSV so it works with your front-end, or setting up a folder structure for a project. These are the tasks that feel like they need a "real developer" but absolutely do not. Describe the task to Codex in plain terms, review what it produces, and run it in small steps so you can catch anything that looks off.

The key phrase from the lab's approach is "write and refine." They are not asking Codex to produce a finished script in one shot. They are iterating, the same way you would iterate on a design file.

Ground-truth testing is not optional

This is where the story gets honest in a way that matters. De la Fuente is direct: a promising candidate molecule is not a drug. It still has to be tested, optimized, checked for toxicity, and cleared through trials. AI accelerates the search, but it does not replace verification.

The same logic holds when you are building a product. A vibe-coded feature that works in your local environment is not a shipped feature. You still need to click through it as a real user would, test it on a phone, and watch someone else try to use it without your help. "You have to always double-check for accuracy" de la Fuente says about AI output. That applies whether the output is a molecule or a modal window.

The takeaway

What this researcher's workflow shows is that AI is most powerful when it lowers the barrier between what you already know and what you need to learn next. It does not turn a biologist into a programmer overnight, and it will not turn a designer into a full-stack engineer overnight either. But it makes the gap crossable.

Start with the thinking, use AI to bridge the skill gap, iterate in small steps, and test everything yourself before you trust it. That is not a revolutionary method. It is just good practice, and it works at every scale from drug discovery to your first deployed app.