4 ms·
ASK HN: Computational tools for predicting drugs effectiveness in diseases
Clinical trials are expensive, very difficult to set up in the case of rare diseases. If we look at the entire cycle from a university lab to a successful Phase III, a positive result is around 1 in 1,500.
So many people wonder if clinical trials could be replaced by computational trials. This happened at least in one case in USA. There are very sophisticated software in the area of PK/PD, but they are based on very simple assumptions such as storage and diffusion of drugs through the biological compartments.
A new article on Nature[0] tells that it is possible to predict side effects using a recommender system. Basically the scientists built a matrix between drugs and their side effects and inferred probabilities of side effects, including side effects that were still not observed. The researchers claim their tool makes it possible to correlate these probabilities with physiology. The code is available [1].
I guess it would be simple to extend this idea to predict how a drug would be successful for a disease.
This is actually the goal of any clinical trial. This could save billions of dollars in research in biotechs and academia while saving time and life of patients. It would also render animal testing obsolete.
We have databases such as Drugbank [2] linking drugs to diseases. It is quite trivial to make a matrix linking diseases to drugs. The questions are "how much trustworthy it is?" and "how it could demonstrate trustworthiness to non-engineers or non-mathematicians, specially to medical specialists".
For having any success, the issue of trust certainly need a lot of scrutiny.
Any thought?
[0] https://pubmed.ncbi.nlm.nih.gov/32917868/
[1] https://github.com/paccanarolab/Side-effect-Frequencies
[2] https://en.wikipedia.org/wiki/DrugBank
- gus_massa 6y agoThe predictions in the article are too generic. I don't think they have an accurate model to predict the exact probability of each side effect. Extending this model to predict if a drug will cure a disease is very difficult. Side effects are very generic like "vomiting" and a wide set of drugs cause "vomiting" for a wide set of reasons. If a drug is useful against a disease sometimes is very specific, for example if the drug makes enzyme X of one virus or bacteria use the drug instead of another substance Y, but it does not confuse the human version of X. Or if it blocks the enzyme X of one virus or bacteria, but it doesn't block the human version of X. Some drugs are very generic, like some antibiotics, and sometimes you can use broad classifications like Gram positive or negative bacteria. But in many drugs the effect is more specific. Note that the article was not published in "Nature" but in "Nature Communications". Nature has created a bunch of journals https://www.nature.com/siteindex#journals-N https://www.nature.com/siteindex#journals-N that have a less strict level than the original "Nature" journal and I have seen a few dubious articles published in these journals. I don't remember something specific about "Nature Communications", but a general personal warning about "Nature X".
- JPLeRouzic 6y agoThanks for this very detailed answer!
- gus_massa 6y agoYou may be interested in https://en.wikipedia.org/wiki/Quantitative_structure%E2%80%93activity_relationship https://en.wikipedia.org/wiki/Quantitative_structure%E2%80%9... It's a family of methods to try to predict the properties of molecules using computer, in particular from similar molecules. There are a thousand of variants that are useful in some cases. It can be used for an initial filtering or to get some ideas, but it can not replace [1] a clinical trial. [1] For now, and probably for a long time. Nobody know what will happen in a few decades/centuries.