Comments
If AI Accelerates Scientific Discovery, Will African Research Infrastructure Become the Bottleneck? Lessons From Running a Biotechnology Centre in Nigeria
4 min read
The advancement of frontier models is beginning to change our understanding of how we understand science and scientific discoveries. Scientists are now able to solve complex problems using AI. AI models have now been used to identify novel drug targets, develop novel drugs. For example, AI model was used to design rentosertib, a drug candidate for idiopathic pulmonary fibrosis that has progressed into human clinical trials. In another example, researchers used AI model to discover abaucin, a new antibacterial compound targeting multidrug-resistant Acinetobacter baumannii, which was subsequently validated experimentally and shown to reduce bacterial burden in an animal infection model.These examples suggest that AI is beginning to compress parts of the journey from scientific question to candidate intervention
As frontier models become very intelligent, we will continue to see increasing scientific discoveries, however every scientific claim needs validation in the lab, this therefore exposes a critical bottleneck to scientific prosperity. Irrespective of whatever you find Insilco, You may still need clinical isolates, reliable freezers, sequencing reagents, ultracentrifugation, animal facilities to validated your claims assays. And sometimes you simply need a replacement electronic component for a laboratory instrument that cannot easily be sourced locally.
When one component of a production system becomes dramatically cheaper, the value of its complementary inputs increases. As AI makes some components of scientific discovery dramatically cheaper, we are now experiencing a significant advancement in scientific understanding, where a scientist in Jos, Plateau State can use AI to solve a complex problem that could have taken years to training . The scarce complementary inputs—laboratory infrastructure, experimental validation, manufacturing capacity and clinical-trial systems—may become relatively more valuable. As AI makes some components of scientific discovery substantially cheaper, we may therefore be entering a period in which the principal constraints on scientific progress begin to shift.
The traditional pathway to medical innovation had been:
Scientific ideas + expertise + laboratory infrastructure + capital + regulatory/clinical capacity → biomedical innovation
What we have seen in recent times is that AI has substantially lowered the cost of some forms of expertise and idea generation and the need for some kind of expertise. Historically, scientific expertise itself was one of the scarcest inputs. Advanced knowledge was concentrated in a relatively small number of individuals and institutions, and acquiring that expertise could require years or decades of training.
AI is beginning to alter that equation.
What we are seeing now is
AI ↓ cost of scientific cognition + experimental infrastructure + capital + regulatory/clinical capacity → biomedical innovation
While AI does not eliminate the need for scientists or specialised expertise- I hope we do not get to that, with the advancement in development of humanoids, but AI models can substantially lower the cost of accessing certain forms of scientific cognition: literature synthesis, computational analysis, hypothesis generation, experimental planning and increasingly aspects of molecular and drug design.
This creates what I call a bottleneck shift. AI may accelerate the generation of hypotheses, while simultaneously increasing demand for the infrastructure required to verify those hypotheses.
This creates a new potential bottleneck:
Hypothesis generation ↑ → demand for experimental verification ↑
If the capacity for experimental verification does not increase at a similar rate, scientific progress may increasingly become constrained not by our ability to generate good ideas, but by our ability to test them.
This may have particularly important implications for scientists in low- and middle-income countries. This creates an AI democratization paradox: AI may democratise access to scientific cognition while simultaneously making disparities in physical scientific infrastructure more consequential.
Research funding is increasingly competitive, and the constraint is particularly acute in resource-limited settings. Much of the global-health funding available to LMIC researchers is understandably directed toward implementing, evaluating, or scaling interventions against established health problems. Far less attention is often given to building the scientific capacity required to generate the next generation of solutions.
For example, a funder may be more willing to support a study evaluating the delivery of an existing malaria intervention than to invest in the laboratories, experimental platforms, medicinal chemistry, genomic infrastructure, or clinical-research capacity needed to develop new malaria diagnostics, drugs, vaccines, or other technologies.
If AI is shifting the bottleneck from scientific cognition toward experimental validation and translation, then the conventional distinction between “research infrastructure” and “high-impact research” may itself need reconsideration
Few years ago, I had the opportunity to visit the University of Waterloo, Canada. The purpose of the visit was primarily to learn the techniques of vaccine development using phages. I believed that with the advancement of AI, one could develop practical solutions for endemic diseases and prepare against catastrophic risks. We have used AI models to validate some ideas such as development of typhoid fever vaccines, Lassa fever vaccines and other treatments models using phages.
The important point is not that AI has independently produced validated vaccines or treatments. It has not. Experimental validation remains indispensable. Rather, AI has substantially reduced the cognitive barrier between having an idea and developing it into a scientifically testable hypothesis.
This experience has made the bottleneck shift very tangible to us. We can now ask more sophisticated questions, explore unfamiliar scientific territory more rapidly, analyse possible approaches and develop experimental concepts at a speed that would have been difficult for a small research centre only a few years ago. But eventually every computational insight reaches the same boundary: the physical experiment.
A proposed vaccine must still be constructed and characterised. A candidate therapeutic must still be tested against biological systems. Safety and efficacy must still be demonstrated. Promising results must still progress through preclinical development, manufacturing, regulatory review and clinical trials.
In other words, AI can increasingly help us answer the question, “What might work?”
Our limiting question is increasingly becoming, “Do we have the infrastructure to find out?”
In conclusion, as AI systems become increasingly capable and scientific idea generation accelerates, the bottleneck in research is likely to shift from generating hypotheses to experimentally testing, validating, manufacturing, and translating them into real-world solutions. If this happens, laboratory infrastructure, capital, skilled personnel, regulatory capacity, and clinical-trial systems will become even more consequential.
We therefore need to think seriously about how to expand these complementary capabilities, particularly in resource-limited settings. Otherwise, AI may democratise access to scientific intelligence without democratising the ability to turn that intelligence into discovery and innovation. Addressing this bottleneck will be essential if the coming advances in AI are to translate into shared scientific prosperity.
