IEEE Future Directions
Related Articles
The two-layer multimodal graph represents protein–protein interactions, drug–protein interactions and drug–drug interactions. Credit: Marinka Zitnik et al. Bioinformatics
Discovering the possible side effects of a drug is a complex matter, requiring many extended clinical studies that last many years and cost a lot of effort (and money). This complexity gets multiplied when one wants to consider the side effects deriving from the use of two (or more) drugs. Today, pharma companies can only address a small subset of combinations and considering that many people are using several drugs in parallel (with 39% of the over 65 yo taking 5 of more drugs in the US according to FDA) it is clear the magnitude of the issue (with some 1,000 known side effect from some 5,000 drugs the possible combinations are in the order of 125 billion!) . Notice that some side effect is appearing out of the assumption of two drugs, and not from the assumption of one or the other independently.
Drugs have an effect on proteins, and our body work depends basically on these proteins and how they interact with one another. Proteins are created in our cells, based on the instructions coded in the DNA. Hence, they are connected to our genes. Since there are some 19,000 different proteins and the fact that each single drug affects a number of them which in turns may affect their mutual interaction the number of potential side effects is staggering (and of course most of these will go completely unnoticed).
Stanford researchers started several years ago to look into the potential side effect deriving from the assumption of two drugs for unrelated ailments. Back in 2012 the have been able to leverage on data reporting adverse side effect to pinpoint a number of them resulting from the use of two drugs that on themselves were safe.
In that case the study was about looking at patterns of undesirable outcome and finding correlation with the assumption of several drugs in parallel.
Now, by applying AI (convoluted neural networks) a team of Stanford researchers has shown that it is possible to evaluate in the cyberspace the possible side effects deriving from multiple drugs assumption.
This is an amazing step forward for pharma. Rather than relying of the detection that something is wrong it becomes possible to simulate in the cyberspace the effects of millions of combinations of drugs assumption by simulating the impact of each on proteins and then the alteration in proteins interaction evaluating wha side effect may happen (clotting, high pressure, … there are more than 4 millions side effects known in literature).
They developed a software, Decagon – from the name of a geometrical polygon with 10 sides to indicate the multi faceted issue being addressed, and used it to predict side effects deriving from the assumptions of several drugs in parallel and then checked the predictions with actual data from the field. The results shown a capability of prediction that is 69% better than the ones based on current approach.
Interestingly, the researchers feel that this approach can also be used to find out the correlation between the genotype and the phenotype (this latter is the result of the expression of the proteins coded in the genotype). This is an interesting area of research because today we have the technology to change the genotype (CRISPR/Cas9) but we do not have the knowledge to fully understand the implication of those changes (how they would affect the phenotype).
