Machine Learning and Drug Discovery

 

(Figure 1). Image extracted from source 1: Uma Introdução à tecnologia Deep Learning para Produção. (n.d.). Retrieved September 29, 2020, from https://www.cognex.com/pt-br/blogs/deep-learning/what-is-deep-learning?page=404

    Have you ever wondered what Machine Learning really is? This is a term that gets thrown around a lot, but how does that differ from conventional programming? Well, machine learning is the branch of computer science that studies and develops learning algorithms for machines. In conventional programming the goal is to give the computer specific instructions -algorithm - to perform the desired task.  On the other hand, in machine learning the goal is to create a generalized set of instructions which allows the machine to learn from data/experience and generate its own algorithms to perform the desired task ². One example I like to use to depict this difference is chess. Imagine you had never played chess before and I gave you two manuals. The first manual consists of a over a billion positions in the chess board and the exact move you should play when faced with each position. However, the second manual consists of a couple pages explaining how each piece is allowed to move, the goal of the game, and the few exceptions to the rules ³. In this case the first manual would be comparable to regular programming while the second manual resembles machine learning. With the first manual you would not need to know anything about chess and you would still be able to play very effectively, whereas with the second manual you would need to play/watch many games to achieve a similar level of play. Just like humans, machine learning systems require extensive training to become proficient at a task. Because of this kind of similarity to us it is considered a form of artificial intelligence (AI). 

  

(Figure 2) “Regular Programming Vs Machine Learning”. Created by Author using Adobe Illustrator.

   Aside from serving as inspiration for great dystopian movie franchises such as The Terminator and The Matrix, AI is actually becoming a very useful tool in modern society. One of the most promising fields for the application of AI is pharmacology. Here machine learning has been used to predict interactions between drugs and their targets, generating new useful molecules, and predicting absorption, distribution, metabolism, excretion, and toxicity ⁴.  The vast amount of data generated by experiments and clinical trials over the years is a rich tool for the development of systems that can improve the process of drug discovery ⁵. These improvements would not only help bring about more effective drugs but could also address a major barrier to drug development, the extensive time and money that is required for this process. Currently it takes about 1-2 billion dollars and 12-15 years of research to develop a drug from scratch ⁶.

    But how are these machine learning algorithms able to extract information from complex chemical     structures? One way this can be accomplished is through a technique called one-hot encoding. This is     how it works, first, we must take a graphical chemical structure, translate it into the SMILES chemical language, and then transform it into a binary table as shown in the figure below ⁷. This step makes it easier for machine learning algorithms to work with the data. Once chemical structures are written in this format, they can be integrated with other information about properties of the molecules such as target interactions, toxicity, distribution, etc. Data on a large number of drugs allows AI systems to generate algorithms capable of inferring structures for new molecules with the desired properties. Even-though this process is still in the early experimental stages, it has already shown promise by aiding in the discovery of new dopamine receptor ligands ⁸.

(Figure 3) “One-hot Encoding Workglow”. Image extracted from source 7: Blaschke, T., Olivecrona, M., Engkvist, O., Bajorath, J., & Chen, H. (2018). Application of Generative Autoencoder in De Novo Molecular Design. Molecular Informatics, 37(1–2), 1700123.

    We are just starting to scratch the surface of what AI aided drug development can accomplish. With further improvements in computational power, learning algorithms, and the increasing amount of data available these techniques should become more efficient and widespread throughout the years ⁹. Another interesting area for development will be the integration of drug discovery and pharmacogenomics through machine learning. One day this technology may assist us in developing drugs catered to people with variants of a gene that render conventional treatments less effective or dangerous. With that being said, there will still be plenty room for humans to work in drug development. The goal is not to substitute people but rather to integrate machine and human intelligence so that both sides can work where they are best suited ¹⁰. In this manner, machines will be able to provide us with insight and improve the quality of our work. I believe that AI has great potential in many areas of human endeavor including drug development. But as any other powerful new tool we must be wise in going forward to augment its benefits while deterring its pitfalls.

 

    

By: Bernardo Aguzzoli Heberle

References:

1.      Uma Introdução à tecnologia Deep Learning para Produção. (n.d.). Retrieved September 29, 2020, from https://www.cognex.com/pt-br/blogs/deep-learning/what-is-deep-learning?page=404

 2.      Michie, D., Spiegelhalter, D. J., & Taylor, C. C. (1994). Machine learning. Neural and Statistical Classification, 13(1994), 1-298.

3.      Silver, D., Hubert, T., Schrittwieser, J., Antonoglou, I., Lai, M., Guez, A., Lanctot, M., Sifre, L., Kumaran, D., Graepel, T., Lillicrap, T., Simonyan, K., & Hassabis, D. (2018). A general reinforcement learning algorithm that masters chess, shogi, and Go through self-play. https://doi.org/10.1126/science.aar6404

 4.      Hessler G, Baringhaus KH. Artificial Intelligence in Drug Design. Molecules. 2018;23(10):2520. Published 2018 Oct 2. doi:10.3390/molecules23102520

 5.      Jing Y, Bian Y, Hu Z, Wang L, Xie XQ. Deep Learning for Drug Design: an Artificial Intelligence Paradigm for Drug Discovery in the Big Data Era [published correction appears in AAPS J. 2018 Jun 25;20(4):79. Xie XS [corrected to Xie XQ]]. AAPS J. 2018;20(3):58. Published 2018 Mar 30. doi:10.1208/s12248-018-0210-0

 6.   DiMasi, J.A.; Hansen, R.W.; Grabowski, H.G. The price of innovation: New estimates of drug development.

 7.      Blaschke, T., Olivecrona, M., Engkvist, O., Bajorath, J., & Chen, H. (2018). Application of Generative Autoencoder in De Novo Molecular Design. Molecular Informatics, 37(1–2), 1700123. https://doi.org/10.1002/minf.201700123

 8.      Madapa S, Gadhiya S, Kurtzman T, Alberts IL, Ramsey S, Reith M, Harding WW. Synthesis and evaluation of C9 alkoxy analogues of (-)-stepholidine as dopamine receptor ligands. Eur J Med Chem. 2017 Jan 5;125:255-268. doi: 10.1016/j.ejmech.2016.09.036. Epub 2016 Sep 14. PMID: 27688181; PMCID: PMC5148686.

 9.      Dana D, Gadhiya SV, St Surin LG, et al. Deep Learning in Drug Discovery and Medicine; Scratching the Surface. Molecules. 2018;23(9):2384. Published 2018 Sep 18. doi:10.3390/molecules23092384

 10.    Topol, E.J. High-performance medicine: the convergence of human and artificial intelligence. Nat Med 25, 44–56 (2019). https://doi.org/10.1038/s41591-018-0300-7