Artificial Intelligence (AI) has become the latest buzz word in drug R&D, and many of the world’s biggest pharma companies claim to be using it to drive forward the process of scientific discovery.
The concept is that using AI’s capabilities, the previously unpredictable process of drug development can become much more predictable, and as a result less expensive.
While it is still early days, AI in its many guises can be used at many points during the drug research process.
It’s far too costly to develop drugs, which are becoming more and more difficult to find as scientists attempt to tackle diseases that are poorly understood.
It is also tough to find drugs that are significant improvements over existing medicines in more common diseases, and those which already have well-established treatments.
Companies are turning to AI from the very beginning of the R&D process, when they are looking for that needle in a haystack that is the ideal drug candidate for a particular disease.
Already this year there have been two major developments in early drug R&D. In September, Hong Kong’s Insilico Medicine found a way to use AI and deep learning techniques to design, synthesise and validate a novel drug in 46 days – 15 times faster than the best pharma companies.
And later that month, AI therapeutics firm Deep Genomics claimed a world first after using artificial intelligence to identify a therapeutic drug candidate.
These developments were hailed as pivotal moments for AI in drug R&D and suggest that there is much more to come as computers become more powerful, and the way that AI is used becomes more nuanced.
Fooling the computer
Insilico pioneered the use of cutting edge techniques such as Generative Adversarial Networks (GANs) and Reinforcement Learning for drug discovery and biomarker development.
GANs are based around the concept of two neural networks arguing with each other to create an ever-more accurate depiction of reality.
A GAN based system has already created artwork that is eerily similar to that drawn by humans. Last year Christie’s sold a fictional portrait generated by a GAN-based AI algorithm “artist” called min/G max/D E_x [log(Dx) ) E_z [log(1-D(G(z)))] for $432,500.
The portrait, called Edmond de Belamy, from La Famille de Belamy, is incredibly realistic, with only the bizarre artist’s signature and the slightly off-centre position of the “painting” providing the clue about its real author.
A GAN-based development model with this degree of realism could lead to new medicines that the pharma industry needs to create the next generation of drugs.
GANs work using two networks – Generator and a Discriminator – where the former generates data and the latter evaluates the data for authenticity.
The Discriminator is first fed real-world and fake data, then the Generator attempts to fool the Discriminator network into thinking that its images are real.
After many thousands of attempts, and the Discriminator is fooled, the image is created – and it’s hoped that GAN technology such as that created by Insilico could create an ideal drug instead of a fake painting.
And while the sum raised by the Sotheby’s sale was impressive, and nearly 45 times more than expected, a drug to treat an incurable condition such as Alzheimer’s is worth billions to the pharmaceutical industry.
September became an undeniably pivotal month for pharma AI when Deep Genomics followed Insilico’s achievement by using the technology to identify a therapeutic drug candidate.
Many companies are using AI as a guide – but the Canadian firm’s CEO Brendan Frey revealed at the Elevate festival that this was the first time an AI platform had confirmed the multiple genetic variants that cause Wilson’s disease, and found a drug candidate that matches the target.
While Insilico used AI for virtual screening, Deep Genomics’ ‘AI Workbench’ system did all the work, including suggesting the disease target.
Wilson disease is a rare condition causing an excess of copper to build up in organs, and has been targeted by specialist pharmas such as Alexion.
Deep Genomics will develop the candidate DG12P1 to treat the disease, which is caused by a genetic mutation that impairs the body’s ability to excrete copper.
The scale of the task was immense: Deep Genomics’ AI scanned over 2,400 diseases and more than 100,000 disease-causing mutations while searching for good drug development opportunities.
After all this number-crunching, the system was able to predict the precise disease-causing mechanism of the Met645Arg mutations that cause Wilson disease.
The AI was able to pin the disease to loss of function in the ATP7B copper-binding protein, and identify 12 lead candidates from thousands of potential compounds, accounting for factors such as in vitro efficacy and toxicity.
The result was DG12P1, which works by correcting the exon skipping mechanism of Met645Arg, which Deep Genomics selected on the basis of tolerability experiments.
Insilico and Deep Genomics have yet to prove their technology works in clinical trials, but have provided a glimpse into a future world of drug development.
They are just two companies amongst a whole range of different players that aim to use AI to change drug R&D.
Companies like Atomwise are striking deals with big pharma companies that are prepared to pay big bucks for successful drugs – Eli Lilly is prepared to pay the San Francisco-based firm a million dollars for each AI discovered drug, in a deal worth $550 million in total.
There are hundreds of other startups looking into this field of research, and the rigorous process of natural selection that the pharma industry relies on to produce new medicines will decide which of these will succeed and which will fail.
Surely, the question is not whether AI can be used in drug development, but how deeply the technology will become integrated into the global pharma pipeline.
Deep Genomics and Insilico’s achievements suggest that AI could outperform humans during the R&D stage and there are other uses of the technology further down the line.
AI is already used to interact with patients and could be a vital tool further along the process, helping with clinical trial design, patient recruitment, and monitoring to boost adherence and cut dropout rates.
The technology could also be used in helping doctors with diagnosis, and helping to pinpoint relevant patient populations for trials by analysing data such as handwritten forms and digital medical imagery.
There are also potential drawbacks though: data privacy, security and accessibility are issues in all health technology, including AI.
And technology such as GANs are only as accurate as the information they are fed, and there are concerns that the human inputs that help to shape the initial system could lead to biases when it comes to issues such as diagnoses.
Recently an AI doctor app from Babylon Health came under fire for giving out differing diagnoses for the same set of symptoms, depending on whether a patient is a man or a woman.
The Times reported that doctors found its algorithm gave advice to a 60-year old female smoker reporting a sudden onset of chest pain, saying the probable cause was a panic attack or pain caused by inflammation.
However a 60 year-old male smoker with the same symptoms was told that he may be having a heart attack, and was advised to go to A&E.
Meanwhile the female patient was told to contact her GP within six hours if the symptoms persisted.
Babylon defended the differing diagnoses, saying that it uses clinical data about most likely causes of symptoms, rather than the process of elimination that a GP would use to rule out serious conditions like a heart attack.
While there could be a degree of self-interest in GPs’ concerns about Babylon – the system could be seen as a competitor to traditional GP practices – the episode does highlight that AI does not always “think” in the same way as a human might.
It also shows how the human creators of AI systems can influence it from the start, and the results may not always be well received.
There is also the famous “black box” problem with AI in that it is not always clear how complex systems have arrived at a decision or recommendation, an issue that could slow implementation in the highly conservative world of healthcare.
A better future?
But even with these caveats the likelihood is that AI could have a pivotal role to play in the pharma and healthcare industries.
Optimists will hope that the work of Insilico and Deep Genomics have opened a doorway to potential therapies for previously untreatable diseases.
AI could help transform the fortunes of a pharma company, by helping to find drugs to halt or reverse the progress of Alzheimer’s, something that has eluded human scientists.
This leads to the question about who should receive the recognition for any AI-related success, especially something as transformative as a disease-modifying drug for Alzheimer’s.
In the art world, there is a debate about whether the Edmond de Belamy painting should be attributed to the AI, or Obvious, the group of artist-computer programmers that helped to create it.
In the world of pharma this could have implications for who owns intellectual property, and who lays claims to scientific achievements.
If the technology is as successful as some hope it could be, it creates the tantalising possibility the Nobel prize could one day be awarded to something with a name like min/G max/D E_x [log (Dx) ) E_z [log(1-D(G(z)))].
By Richard Staines
Source: Pharma Phorum
Thermo Fisher Scientific has acquired Novasep’s viral vector manufacturing business in Belgium, Henogen, for about €725m ($874.5m) in cash. Henogen offers biotechnology firms, as well as biopharma customers contract manufacturing services for vaccines and therapies.
Research and development group of Eli Lilly and Company, Loxo Oncology at Lilly, and clinical-stage oncology company Merus have announced a research collaboration and exclusive license agreement to develop T-Cell re-directing bispecific antibodies.
Chinese cancer biotech Adagene has filed to raise up to $125 million in a Nasdaq IPO. The listing will give Adagene the means to run early-phase clinical trials of antibodies against CD137 and CTLA-4.