11 September 2024

AI won’t solve all of medicine’s problems

Technology is invaluable to the future of healthcare – but we must also be realistic.

The Latin terms in vitro and in vivo began appearing in the medical literature around the turn of the 20th century, as the scientific revolution in medicine gathered pace. Directly translated as “in glass” and “in life”, they distinguish biological processes conducted in the laboratory from those occurring within a living organism. An antibiotic, for example, might kill bacteria in the Petri dish yet fail to cure infections in human patients. A drug might interrupt degenerative changes in nerve cells grown in tissue culture but cause unacceptable side effects if administered as a treatment for dementia. In vitro findings are the engine of research, but it is in vivo outcomes that ultimately matter.

In vitro fertilisation is the instance most people will be familiar with. These days IVF seems commonplace, a routine if expensive option for couples unable to conceive naturally. Few now remember the collective bated breath that accompanied its first use by the fertility pioneers Patrick Steptoe and Robert Edwards in 1978. In vitro research had proved the process of life could be started in the lab, and animal models offered assurance that uneventful pregnancy, birth and normal child development should follow once embryos had been implanted in the womb. But until Louise Brown, the first “test-tube baby”, proved IVF was both possible and safe in humans, no one could be entirely certain the in vivo outcome would match the experimental promise.

Discrepancies between the two contexts have many causes, but chiefly reflect the complexity of biological systems compared with the relative simplicity of experimental conditions. Just getting a drug into a patient can be challenging. Many medicines can be given by mouth, but some will be destroyed in the digestive tract, or subject to poor absorption, so have to be administered by injection, inhalation, or across the skin. Once inside the body, these alien molecules will be broken down and excreted, and some may not reach the target tissue or organ in meaningful concentrations. They might also exert unanticipated effects on other systems within the body, causing deleterious effects unanticipated in the laboratory – one reason rigorous testing in animal models precedes human trials.

In recent decades, other Latinate terms have entered the literature. In silico – “in silicon” – was coined in 1987 to denote the computer modelling of biological processes. Most of us are now familiar with one field of application: who can forget the predictions about the havoc Covid was anticipated to wreak on the NHS and wider society? One lesson learned during the post-pandemic inquiries was how persuasively authoritative these in silico models appeared, leading politicians to impose draconian countermeasures. Yet, just as living organisms are immeasurably more complex than in vitro approximations, so are societies compared with their in silico counterparts. Many Covid models turned out to be unduly cataclysmic. The term in populo was proposed by a group of American epidemiologists in 2010 to express the inherent limitations of applying research findings to whole populations – but it never caught on. Perhaps, post-Covid, it might.

The current explosion of machine learning and artificial intelligence suggests we are on the cusp of an in silico revolution in medicine every bit as seismic as that created by the application of laboratory science in the last century. AI is revealing secrets of three-dimensional protein structure – key to all biological processes – generating in a matter of hours insights that would take scientists years of painstaking experiment. Databases and computer models are identifying, or even designing, novel molecules that could be the basis for new antibiotics or drugs. Machine learning is discerning hitherto unappreciated features in pathology specimens that might improve the treatment of cancer.

There is heady excitement surrounding these new technologies, but while they hold extraordinary promise, we should be realistic about them. Just as with the cross-application from other forms of research, the proof of this in silico revolution will ultimately have to be made, case by careful case, in vivo.