Depending on who you ask, artificial intelligence (AI) is either a revolutionary force for good — or a threat to our very existence.
Defined by the Oxford English Dictionary as the capacity for computers or machines ‘to exhibit or simulate intelligent behaviour’, AI was recently described by academics and bosses at major tech companies including Google as being as much of a threat to human survival as nuclear war.
The fear is that AI could spread misinformation on such a scale that it leads to complete social breakdown.
In sharp contrast, barely a week passes without news of AI potentially transforming our health.
For instance, researchers at Lund University in Sweden recently revealed the technology had increased the detection of breast tumours found through screening by 20 per cent.
Experts believe AI will also shave years off the time it takes to get new medicines on to the market (stock photo)
The study compared the results of thousands of breast scans visually checked by two radiologists on a screen (the current NHS screening method) with the results when the same scans were checked by a computer program — and then confirmed the results with biopsies.
The results, in The Lancet Oncology, showed conventional checks found 203 cases of cancer — but AI detected a further 41.
AI is thought to improve detection because it eliminates the risk of tired, over-worked radiologists missing vital signs of a tumour.
Experts believe AI will also shave years off the time it takes to get new medicines on to the market.
Currently this can take a decade and cost at least £1 billion; even then, 90 per cent of new drugs fail safety and effectiveness tests.
But using AI to sift through mountains of data about the structure or DNA profile of millions of molecules to see if they might be suitable for a new drug, could mean new treatments would become available to patients in less than five years, say experts.
‘Artificial intelligence can analyse hundreds of millions — even billions — [of molecules] and in less time than existing software,’ says Dr Philip Crilly, a senior lecturer in digital pharmacy and public health at Kingston University in Surrey.
‘It’s about the volume of data and also the speed at which it gets through it. Lots of screening tools used at the moment cannot read through big data that’s really complex because they are not sophisticated enough.’
Earlier this year, pharmaceutical firm Insilico Medicine announced phase 2 trials (to check effectiveness on a small number of patients) of a drug it developed from scratch using AI in just over two years — less than half the time it would normally take to get a brand new medicine to this stage.
A breast scan using a technology called thermography employed at the Chiron clinic in Harley Street
The drug is a possible new treatment for idiopathic pulmonary fibrosis (IPF) — an incurable lung condition affecting around 32,000 people in the UK, where the tissue in the lungs becomes thick and inflamed. Currently most IPF patients survive for only three to five years after diagnosis.
Apart from sheer computing power, AI’s other ‘skill’ is that it learns from previous decisions or interactions — as with music streaming, where your app remembers what artists or tunes you have searched for and ‘learns’ to recommend similar ones you might like. This ability is proving especially useful when it comes to drug firms identifying whether a new medicine is going to be safe.
‘It can analyse millions of compounds for ones with similar features to the potential new medicine a company may be interested in,’ says Dr Crilly.
By pooling data from similar molecules and their known effects, AI software is able to predict whether the new drug is likely to be harmful to patients.
An estimated 30 per cent of experimental medicines fail during initial trials, because they are found to be potentially toxic.
‘It means drug companies can identify much earlier in the process if there is likely to be a problem [with the safety of a new drug] and walk away from it sooner, which will potentially save them hundreds of millions of pounds,’ says Dr Crilly.
In February, researchers at Sheffield University announced they had developed an AI package — called DrugBAN — with scientists at AstraZeneca, which can predict if a candidate drug will, as hoped, interact with its target organs or tissues.
The computer program is trained to learn the structure of millions of proteins in the human body — and also the structure of the potential new drug compounds. It then ‘learns’ how these structures interact with each other and predicts whether a drug is likely to have its desired effect, or whether it will be too toxic to use, long before it has entered the testing stage.
Yet there is an underlying problem with AI in drug development, warns Penny Ward, visiting professor in pharmaceutical medicine at King’s College London: much of the information on which AI relies to make decisions about drug development comes from research that has already been published.
But a long-standing issue — positive publication bias — means negative research findings are less likely to be published than positive ones, skewing the evidence.
‘This is either because editors want to print only positive stuff, or investigators are reluctant to have negative findings on their academic record,’ says Professor Ward.
‘This is a weakness because AI is only ever as good as the data that it is being fed.’