Artificial intelligence, or AI, will shape our world in the coming years. It will drive our cars, plan our groceries, and annoyingly, beat us in games. People already speculate how AI can superhumanly solve any human problem. Such as healthcare – such as the development of novel therapeutics for unmet clinical needs?
Yes, AI is a big topic in healthcare. Yes, AI will help improve diagnosis and treatments, and AI will keep us healthier. Especially when it figures out how to motivate us to lose weight, quit smoking, and drink less.
And no. AI will not solve some of our biggest health challenges any time soon. Like the one-trillion-dollar Alzheimer-bomb mentioned in an earlier blogging. Why not?
Intelligence means the ability to learn, based on the available information, to solve complex problems. AI does that through computerized algorithms. For instance, the AI algorithms of Aiforia can quickly and reliably identify and count the cell of interest in a massive amount of microscopy images. The learning part comes from a human ‘teacher’. The plain algorithm is helpless like a new-born baby. However, after the teacher provides a sample of images and marks the cells of interest, the algorithm learns to do the same.
In a nutshell: A clever algorithm digests huge amounts of imaging data (in theory, just any photos). A neat user interface that allows the user to teach what to look for (e.g. all cars). And there’s your AI that can quickly count how many cars there are in your thousands of photos.
How could you train an AI to develop an efficacious therapy to solve the one-trillion-dollar Alzheimer-challenge?
You could start with real-life patient data. In a country like Finland where demographic and healthcare records are digital, and the FinnGen project adds a large amount of genetic data on top of that, there’s a huge amount of relevant data for teaching AI algorithms. Such approaches will help us identify risk groups, learn how to delay symptoms, improve preventive actions, maybe develop better diagnostics. All that will be immensely valuable. But it will not provide a cure of Alzheimer’s – because there is no pattern of a cure within the data.
How about applying Natural Language Processing? We could teach an AI with all relevant scientific articles. That’s a massive amount of data based on which the AI could suggest new therapeutic approaches. It would master all knowledge on therapeutics, human brains, and Alzheimer’s. Unfortunately, with the best algorithms, this approach would be limited by the quality and boundaries of the ‘teaching’ data. Not all scientific articles are of high quality; any ‘promising’ results yet to be shown incorrect will mislead the algorithm. And since there are no disease-modifying treatments in Alzheimer’s there is no data to guide toward one. The AI might propose interesting approaches worth studying further; but again, that’s a long way from a cure. (The AI might also conclude that a solution does not exist since all promising approaches fail in the clinic!)
How to push AI beyond what is already known? You could have it learn a complete, perfect model of the human and the development of Alzheimer’s disease in that model. Alas: with some 100 billion neurons and their connections and continuous interactions, the complexity of the human brain is still way beyond AI models. Not to mention a detailed cell-level simulation of its changes over decades of ageing, stress, illnesses, and manners of living.
AI provides great tools for scientific research and drug development. Aiforia, mentioned earlier, is a great example: It can help scientists analyze large amounts of data. What was previously tedious manual work is becoming quick and reliable. Similarly, based on large amounts of data from thousands of research studies, algorithms can predict the toxicity and other attributes of novel drug molecule candidates.
Let’s just not demand too much from AI yet. Preclinical and clinical testing will remain the most important part of drug development for a long time. Laboratory results will surprise us again and again over computed predictions. And most importantly: Drug development is about human safety. Only very well justified risks are accepted. Therefore, it will take decades until the first regulatory authority in the world will dare say: “Here’s your new drug market approval, based only on the computed data you have provided. No clinical testing is necessary.”
The author is a biotech entrepreneur and CEO who started his career as an AI research scientist. Herantis Pharma uses computerized approaches to support its drug development.