Cossed Wires: There is a nobler side to the AI hype

The drumbeat of bad-news AI stories over the past few weeks (a few of them by this writer) has been draining – a cacophony of scare stories has sucked all the oxygen out of the room. It has left many hyperventilating so miserably over hypothetical dystopian futures that there is a danger of missing the other side of the story. A rebalancing of the narrative is due.

I recently heard about a company called Insilico Medicine, listed on the Hong Kong stock exchange. So extraordinary is their good news story that I dashed off to see if I could buy a few shares, but I was unable to find my way through the thicket of exchange bureaucracy to open an account there.

It sits at the centre of an embryonic new industry, that of creating new and unprecedented treatments for all that ails us, using AI as their primary R&D partner to create new molecules to hunt down and kill biological nasties that have stubbornly resisted previous treatment.

We’ll get back to Insilico and its ilk presently, but it is worth a brief diversion into how we treat disease, at least in traditional medicine (alternative medicine is a subject for another day). Almost all research done to develop cures and palliatives and mitigators is undertaken by companies operating in the pharmaceutical industry (although academia and others also contribute).

At its simplest, drug discovery undertaken by these companies is a search for the right molecule to interfere with the biology that is making us sick.

How have they done this, pre-AI? Researchers found these molecules in nature (occasionally using the signposts of traditional medicines) – plants, fungi, bacteria and other organisms – or created them through synthetic chemistry, eventually building libraries containing millions of compounds.

Once scientists identified a biological target associated with a disease, they could use high-throughput “screening” to test huge numbers of these compounds against it in the lab, then have medicinal chemists repeatedly modify promising “hits” to improve their effectiveness, safety and ability to survive inside the body.

Sometimes the process worked in reverse – researchers exposed diseased cells or organisms to many compounds in an educated scattershot, discovered that something unexpectedly worked, and only afterwards figured out why.

But the process is glacial and attrition-heavy. The traditional timeline to bring a new molecule from discovery to pharmacy shelves spans 10 to 15 years, routinely burning through more than two billion dollars per approved drug. It is a system built on informed brute force – the screening happens in physical labs, by scientists in white coats operating fancy equipment, hunting for a rare hit, only to watch roughly 90% of candidates fail when they finally reach human clinical trials.

This high cost of failure has also starved smaller patient populations – particularly children, who suffer from thousands of rare genetic disorders. There is little profit in rare disease, and salaries have to be paid, as do investors.

AI crunches the obstacles

AI changes the fundamental economics of this process. Instead of merely searching an enormous warehouse of existing molecular “keys” for one that fits a biological “lock”, generative AI can increasingly help identify the lock and design a new key specifically for it.

It does this by shifting drug discovery from what was essentially an empirical guessing game to a predictive computational science. By training deep neural networks on vast datasets of multi-omics (a fancy collective word meaning combining information about our genes, gene activity, proteins and cellular chemistry) and biological and chemical processes, machine learning models can now predict how complex proteins fold, map how cellular targets interact, and design custom molecules from scratch.

The AI then says – here are the instructions for the molecule you need to make, have a nice day. Any deeper explanation than that is above my pay grade, but the takeaway is that what used to take half a decade of laborious lab work can now be executed in weeks. AI and drug discovery are a perfect match.

So how close are we to healing the world? A long way still, but some Rubicons have been crossed.

Let’s start with the company I mentioned earlier – Insilico Medicine, probably the best example of a pure AI drug play to date. On 10 September, a patient in China became the first to be dosed in a Phase 3 trial of a drug called rentosertib, one of 320 people with a nasty disease called idiopathic pulmonary fibrosis, who will take it once a day for a year across 47 centres. The condition is degenerative and fatal – lung capacity gets smaller and smaller, basically asphyxiating the patient.

So how good were the Phase 2 results for rentosertib? In a 71-patient trial designed to test safety rather than efficacy, patients on the highest dose saw their lung capacity rise by an average of 98.4ml over 12 weeks, while those on placebo lost ground. That matters because pulmonary fibrosis is a disease that only goes one way. Around five million people have it, median survival is three to four years, and existing drugs slow the decline without reversing it. A lung that gets slightly better is close to unheard of with this condition.

This should have been front-page news – it is the first AI-designed drug to reach Phase 3 (if Phase 3 is successful, the drug goes to market). But alas, it was not.

Not making the headlines

Why does news like this sink without trace? Part of the answer is that damned attention economy. AI harms, real or imagined, are vivid (deepfakes, hacks, hallucinations, bioweapons, and, of course, the end of our species), while AI benefits in drug discovery are statistical, slow and invisible.

An AI drug that works shows up years later as a child who didn’t get sick. That is not a headline. And clinical science moves at the pace of a 52-week trial, while the scare cycle refreshes hourly. As we all know by now, bad news outshouts good.

Rentosertib is not alone, and the spread of diseases is the point. Generate:Biomedicines has taken GB-0895, an antibody engineered with AI to block a driver of airway inflammation, into two global Phase 3 trials enrolling about 1,600 adults and adolescents whose severe asthma current treatments cannot control.

In January, Insilico (again) dosed its first patient with garutadustat, a gut-restricted molecule for inflammatory bowel disease that went from programme start to preclinical candidate in 12 months, after roughly 115 compounds were synthesised rather than the thousands the old arithmetic demanded.

Recursion’s REC-4881 addresses familial adenomatous polyposis, an inherited condition that carpets the colon with polyps and carries a near-certain lifetime risk of colorectal cancer, and for which there is currently no approved drug and no option but surveillance and surgery.

In a small trial, polyp burden fell by a median 43% after 12 weeks, and by 53% at week 25, after patients had been off the drug for three months. And Takeda’s zasocitinib, for psoriasis, beat an approved rival head-to-head in Phase 3 in June and is heading for FDA submission this fiscal year.

Lungs, gut, colon, skin and more. In researching this, I found that more than 170 AI-originated molecules are advancing through various stages of clinical trials, with early data showing that AI-selected molecules are entering Phase I trials with significantly higher success rates than the historical industry baseline (these are described in detail here, for those who like foraging in the weeds).

McKinsey published this paper on 21 September, putting a number on how much faster AI drug discovery gets to human trials – it is between 15% and 80%, depending on the programme. Good news indeed, but McKinsey’s consultants, not a constituency given to underselling technology, provided this caution: “No amount of raw compute or added GPUs can speed up cell growth.”

Biology still runs on biological time, and no server farm has yet persuaded a cell to divide faster. Yes, of course – but even so, this is where AI is paying its proudest dues.

In any event, when some of these drugs finally end up saving lives a few years from now, after the mandatory trial track, there will be no headlines, just fewer people dying in unpleasant ways. And that is a PR problem which needs solving.

The anti-AI backlash needs challenging before the entire industry is hobbled by alarm-mongering. Our children’s health depends on it. DM

Steven Boykey Sidley is a professor of practice (ex-JBS, University of Johannesburg), a partner at Bridge Capital and a columnist-at-large at Daily Maverick, where he writes the weekly Crossed Wires column. His new book, It’s Mine: How the Crypto Industry is Redefining Ownership, is published by Maverick 451 in South Africa and the Legend Times Group in the UK/EU, available now.