The Molecule Has to Survive the Body

A molecule can look magnificent on a screen and disappear uselessly in a liver.

That is the small, stubborn fact beneath the greatest promise of AI medicine. We are learning to search biological possibility at extraordinary speed. But a patient is not a benchmark, and chemistry does not award prizes for persuasive language.

I am optimistic because the evidence is becoming tangible. The ambition is no longer confined to algorithms ranking compounds in retrospect. AI-designed medicines have reached controlled human studies. Structure prediction has expanded into molecular interactions. Robotic laboratories can close portions of the loop between hypothesis and measurement. Taken together, these developments suggest a new industrial shape for discovery.

They do not yet establish an automatic cure factory.

Three breakthroughs that belong in the same picture

Start with molecular understanding. AlphaFold 3’s 2024 Nature paper describes a system for predicting structures involving proteins and other biomolecules, including small molecules. That matters because drugs act through interactions, not through isolated pictures of proteins. A useful structural hypothesis can help researchers decide which experiments to perform.

The prediction remains a hypothesis. Binding geometry is not clinical benefit. Living systems include concentration, metabolism, competing targets, tissue access, and feedback. A confident-looking image can hide uncertainty that only an experiment will expose.

Then consider translation. A 2025 randomized phase 2a study of rentosertib tested an AI-discovered TNIK inhibitor in 71 people with idiopathic pulmonary fibrosis over 12 weeks. The trial’s existence is an important milestone: the discovery story has crossed into prospective human evaluation. Its size and duration also define the limits of the claim. This is not equivalent to regulatory approval or proof of long-term benefit across populations.

Finally, consider the physical laboratory. In autonomous exploratory chemistry research, mobile robots connected synthesis and analytical equipment, with a decision process using different measurements and checking promising reactions for reproducibility. The achievement is specific and practical: automate a scientific workflow while keeping contact with measurements of the world.

These three strands fit together. Models generate better candidates. Experiments distinguish attractive fiction from useful chemistry. Clinical studies ask whether a candidate helps human beings.

The new scarcity is trustworthy feedback

When hypotheses become cheap, a laboratory’s value changes. Generating another plausible molecule may be easy. Knowing which result is reliable becomes precious.

This is why I would invest in assay quality as seriously as model quality. A noisy assay can train a discovery system to chase artifacts. A mislabeled sample can turn an elegant optimization loop into a machine for repeating a mistake. An unrecorded change in temperature, reagent batch, or protocol can make one promising result impossible to reproduce.

The practical frontier is therefore partly unglamorous: calibration, metadata, controls, chain of custody, independent replication, and experimental design. An advanced model connected to poor measurements is an advanced way to become confidently wrong.

A discovery organization should retain failed experiments. Negative results help define what the world refuses to do. If every unsuccessful compound vanishes from the record, the next model inherits a distorted landscape full of successes without their cost.

Intelligence becomes medicine only when the body is allowed to disagree.

What Africa should build

Africa’s opportunity is larger than supplying samples to foreign discovery engines. The valuable capability is the complete chain: choosing locally important questions, generating data, developing candidates, conducting rigorous studies, manufacturing reliably, and negotiating access.

A Nairobi team could begin with a narrow disease problem and a small number of validated assays. It could partner with clinicians who understand the real treatment gap and chemists who know how promising molecules fail. It could build a clean, reusable experimental record before attempting a vast automated campus.

The disease choice matters. A project should not choose a fashionable target merely because public datasets are abundant. Local clinical need, attainable validation, available expertise, and a plausible route to affordable treatment should influence the programme from the beginning.

African plant diversity and associated knowledge may offer starting points, but neither geography nor tradition guarantees therapeutic activity. An extract is a mixture. Its composition can vary. A compound that kills cells indiscriminately may look exciting in one assay and be useless as a medicine. Safety, consistency, and mechanism require careful work.

The Nagoya Protocol’s provisions on traditional knowledge also make the social architecture inseparable from the scientific architecture: access must respect relevant domestic requirements, community involvement, prior informed consent or approval, and mutually agreed terms. A discovery platform built on extraction without benefit sharing would reproduce an old industry in a new interface.

What AGI might change

If broadly capable systems arrive, I expect them to make the coordination of discovery far more effective: connecting literature, experimental design, chemistry, instrumentation, and analysis. That is a forecast. It does not remove physical timelines or ethical obligations. Some experiments must wait for biology to unfold. Some harms become visible only after longer follow-up.

The most exciting future is a laboratory that learns continuously while becoming more accountable with every cycle. Its models propose boldly; its instruments answer precisely; its clinicians judge relevance; its communities share authority and reward.

The molecule leaves the screen. It enters a vial, a cell, a study, a human life. Every boundary it crosses asks a harder question.

That journey, rather than the first beautiful rendering, is the breakthrough worth building.

Editorial note: This essay expresses the founder’s ideas, developed with AI-assisted research and writing and reviewed against the linked sources on 29 October 2026. Convictions, proposals, and forecasts are identified in the text; scenes framed as imagined are illustrative.

Cover photograph / illustration: Bennettscience. CC BY 2.0. Original source. Illustrative image: Homemade Protein Structural Model.jpg. Credit: Bennettscience; CC BY 2.0. This image does not document the research or proposed facility described.