A Nigerian scientist's AI could free insulin from the fridge

The lab is quiet, but the machines are not. Somewhere in Scotland, a server hums through another simulation run. It is not rendering graphics or training a chatbot. It is calculating the forces between a protein and a synthetic polymer at the molecular level. The protein is insulin. The polymer does not exist yet. But the AI running the simulation might find it.

That is the central tension of Martins Otun's work. He is not trying to build a better app. He is trying to make the most widely used diabetes drug in the world survive outside a fridge. And he is doing it with a tool most people still think of as a writing assistant.

"Cold-chain storage limits access to insulin for hundreds of millions of people; a thermally protective patch polymer could help, but the design space is too large for exhaustive experiment," Otun writes in his arXiv preprint.

The "design space" he mentions is not a nice turn of phrase. It is the number of possible polymer molecules — chemically plausible candidates that could wrap around insulin and keep it from degrading in heat. That number is astronomical. No wet lab on earth could brute-force its way through it. But software can.

The physics comes first

Here is the problem in plain terms. Insulin must be kept between roughly 2°C and 8°C. In most of sub-Saharan Africa, that is a hard promise to keep. Electricity is unreliable. Rural clinics may not have functioning fridges. The result is a cruel irony: a cheap, life-saving medicine that becomes useless precisely in the places where it is needed most. Research cites more than 500 million people living with diabetes globally. Cold-chain failure is not a footnote. It is the story.

Otun's answer is not another refrigeration gadget. It is a polymer patch that sits against the skin and keeps insulin stable at higher temperatures, allowing it to be delivered transdermally. That is the "fridge-free" ambition. But to find the right polymer, he needed a way to search a space too vast for ordinary experiments.

So he gave an LLM orchestral control of physics tools. The model proposes candidate polymers, encoded as PSMILES strings, and the system simulates how each candidate interacts with insulin using OpenMM and Packmol. Those are not toy tools. OpenMM is a molecular simulation engine used in serious biophysics. Packmol packs molecules into boxes for simulation. The LLM calls them through the Model Context Protocol, orchestrating the loop.

"Starting from that problem, we narrow to an agentic workflow: a large language model (LLM) calls physics-based tools through the Model Context Protocol (MCP), searching the discrete PSMILES space under a budget of OpenMM Packmol-matrix evaluations," Otun explains.

The "budget" matters. You cannot simulate a billion polymers. You have to choose where to look next. That is where the AI's judgment comes in.

A discovery world with a memory

Otun describes the LLM as "an implicit acquisition function conditioned on a persistent 'discovery world'." That world contains the model's hypotheses, the claims it has read in literature, and the simulation results it has already collected. Every loop, that world updates. The model is not just guessing fresh each time. It is learning from its own history.

"The LLM acts as an implicit acquisition function conditioned on a persistent 'discovery world': hypotheses, literature claims, and simulation outcomes updated each iteration," Otun writes.

The results, according to the preprint, are striking. The best autonomous campaign found candidates with an insulin–polymer interaction energy around -2263 kJ/mol. That number measures how strongly the polymer holds onto the insulin. More negative means stronger, more stabilizing contact. Otun reports the system outperformed reinforcement learning baselines by roughly 68% and beat Bayesian optimization by about 19% in that metric.

Three independent campaigns converged on a similar structural pattern: a high density of hydrogen-bond donors and acceptors per repeat unit. In other words, the AI kept landing on the same family of chemical features. Physics-based checks filtered out candidates that could not be packed correctly or whose names did not match their structures. The system was not just proposing molecules — it was verifying its own proposals.

The CPU-bound surprise

One detail deserves attention: the workflow is CPU-bound and runs on commodity hardware. No GPU cluster required. That is a practical signal, not just a technical note. It means a research team in Lagos or Nairobi, without access to a wall of Nvidia chips, could run this same kind of hunt. The codebase, called FRIDGEFREENET, underlines the intent.

That accessibility idea resonates with where this work could eventually land. In global reports, 50 to 60 AI-enabled biologics are in discovery, preclinical, or clinical stages, and 13 AI-derived biologics have reached clinical trials. But none of those specific candidates are attributed to Biologix or Algonix. This work is earlier-stage — a computational research project, not yet a filing with health authorities. That is an honest boundary worth drawing.

Ongoing physics-informed approaches in diabetes are often aimed at glucose control modeling, as documented in recent conference papers and control journals. Otun's focus on polymer stability is a different lane — formulation and delivery, not only prediction and dosing. The two could eventually meet: a stable, fridge-free patch plus an AI-driven closed-loop system would be a much more complete answer.

Why this matters for African tech

Techpoint Africa's tweet — @TechpointAfrica — drew modest engagement: two likes, no retweets, no replies. That silence is not a sign of irrelevance. It is a sign of how early this story is. The work is a preprint. The company, Algonix AI Ltd., is a UK-registered entity. Biologix appears to be a research project under that umbrella, not yet a heavily profiled startup with visible VC funding rounds.

But the narrative weight is real. African talent contributing to physics-grounded AI for biologics is rare. The continent's startup ecosystem remains tilted toward fintech and commerce. This is a different kind of signal: a Nigerian-affiliated researcher building software that could change how a life-essential drug is delivered in hot places with bad roads and unreliable electricity.

The regulatory pathway is not yet defined. No NAFDAC involvement is visible. No clinical trials are listed. Any eventual product would face scrutiny as a novel drug delivery system or medical device. That is a future conversation, not current news. But the precondition for all of it — a credible computational discovery engine — is what is on the table today.

Otun's work does not promise a fridge-free insulin patch by next year. It makes a more precise, more humble claim: that a physics-grounded AI can search polymer space faster and smarter than prior methods. If that claim holds up, the fridge-free ambition stops being a far-off dream and starts becoming an engineering timeline. The server in Scotland keeps humming. Somewhere in that noise, a molecular match waits to be found.