Can AI Shorten the Road to a Medicine? What Has Been Achieved and What Has Not

Ask a chip engineer about his industry and he will say that power doubles while cost falls. In drug-making the situation is reversed, so much so that the authors of a study published in 2012 in Nature Reviews Drug Discovery named the phenomenon "Eroom's law", that is, Moore's law spelled backwards. These researchers, led by Jack Scannell, estimated that the number of new drugs approved per billion dollars of research spending has roughly halved every nine years since 1950, after adjusting for inflation.

What does a drug really cost?

The figures vary with whoever does the counting. A study published in JAMA in March 2020, led by Wouters (O. J. Wouters), relied on public financial reports and put the median cost of a drug approved in the United States between 2009 and 2018 at about $985 million, with a mean of $1.3 billion, whereas older estimates from industry itself spoke of a mean as high as $2.8 billion. The researchers acknowledged the limits of their work: the selectivity of what companies disclose, and the over-representation of some diseases in the sample. No single figure should be taken as absolute truth, but the direction is clear: the cost is high, and much of it is the price of failures loaded onto the drugs that succeed.

What did AlphaFold actually achieve?

A protein's shape determines its function, and knowing the shape makes it possible to imagine how it might be disabled or activated. Predicting the shape from the sequence of amino acids remained a hard problem for decades. On 9 October 2024 the Royal Swedish Academy of Sciences awarded the Nobel Prize in Chemistry in two halves: one to David Baker for computational protein design, the other to Demis Hassabis and John Jumper for protein structure prediction.

Reports indicate that the AlphaFold team predicted structures for nearly all of some 200 million known proteins, with accuracy approaching X-ray crystallography, the most precise laboratory method. The AlphaFold DB database, run by Google DeepMind in partnership with the European Bioinformatics Institute (EMBL-EBI), makes these structures available to researchers. It is an achievement that spares laboratories years of work. But a protein structure is not a drug: between it and the patient lie a molecule that must be designed, a safety record that must be established, and an efficacy that must be measured.

The first "machine-designed" drugs put to the test of patients

The clearest test so far is rentosertib, for idiopathic pulmonary fibrosis, developed by Insilico Medicine. Nature Medicine published the Phase 2a results on 3 June 2025, possibly the first peer-reviewed result for a molecule whose target and structure were both generated by AI. The trial enrolled 71 patients, randomised and double-blind, for 12 weeks. Among those taking 60 mg daily, lung function (forced vital capacity) rose by a mean of 98.4 ml, against a fall of 20.3 ml on placebo.

The figures need careful reading, though: the high-dose arm had only 18 patients, follow-up was short, the participants were all from China and of similar background, and seven patients stopped treatment because of liver injury or dysfunction. The authors described the results as encouraging for further study, not as proof of success. That is the sound position: a first piece of good news, not a final verdict.

Who is represented in the data?

An algorithm knows only what it has been fed. A review in Nature Medicine (2022) showed that people of European ancestry made up 86.3 per cent of genomic study samples as of June 2021, against 1.1 per cent for Africans and 0.8 per cent for South Asians; it gave no figures for Arabs. The worry is that models will predict well for those whose data are plentiful and fail the rest, widening health disparities instead of narrowing them.

From this angle the Saudi Genome Program, one of the Vision's projects, takes on particular importance; it aims to build the first genetic map of Saudi society and to develop personalised medicine and premarital and newborn screening programmes. As for the number of genomes completed, I found no figure in the official source, so I do not state one. Genetic data are highly sensitive, so the safeguards that protect them deserve review (see the article "Privacy in the Age of AI").

What to do

• Clean, reliable local data, collected with informed consent and protected.
• Skills that combine biology and computer science, since a model cannot replace a scientist who understands what it reads.
• Explainable decisions: it is not enough for a system to say "this molecule is toxic"; it must show why.
• Partnerships between universities, companies and regulators that safeguard patients.

Taking the means is part of true reliance on God; medicine is one of the means, and the One who heals is God. If the algorithm shortens the road to treatment for those who need it, it is a blessing that calls for gratitude and good use, and what matters is that medicine reaches everyone, not that it arrives quickly only for those whose data are plentiful.