Märkte & Technologien ability to bind drugs. Information on the targets can be obtained experimentally by high-throughput screening, or new targets can be found in scientific publications. The latter case looks promising as it can cut the expenses by reducing laboratory experiments, however, it is quite challen- ging. If we consider that the number of publications in PubMed database is more than 18 million abstracts and every month about 60,000 new abstracts are added 8, it is scarcely possible for a human resear- cher to read so many articles in their whole life. At the same time, AI augmented data mining enables the extraction of information about novel targets from publicly available sources: publications, patents, and other sources by using Natural Language Proces- sing methods. As reported by BenevolentBio’s CEO Jackie Hunter, it is possible to gain a fourfold increase in speed of identifying targets, validation, and the R&D success rate with the adop tion of deep learning. 4 1) Barabási, A.-L. and Oltvai, Z. N. (2004) ‘Network biology: under- standing the cell’s functional organization’, Nature Reviews Gene- tics, 5(2), pp. 101–113. doi: 10.1038/nrg1272. 2) Cumming, J. G. et al. (2013) ‘Chemical predictive modelling to improve compound quality’, Nat Rev Drug Discov, 12(12), pp. 948– 962. doi: 10.1038/nrd4128. 3) DiMasi, J. A., Hansen, R. W. and Grabowski, H. G. (2003) ‘The price of innovation: New estimates of drug development costs’, Journal of Health Economics, 22(2), pp. 151–185. doi: 10.1016/ S0167-6296(02)00126-1. 4) Ernst & Young LLP (2017). Staying the course. Biotechnology re- port. Beyond borders. E&Y. 5) Kadurin, A. et al. (2016) ‘The cornucopia of meaningful leads: Applying deep adversarial autoencoders for new molecule deve- lopment in oncology’, Oncotarget, 8(7), pp. 10883–10890. doi: 10.18632/oncotarget.14073. 6) Schadt, E. E. (2009) ‘Molecular networks as sensors and drivers of common human diseases.’, Nature, 461(7261), pp. 218–223. doi: 10.1038/nature08454. 7) Schneider, G. (2010) ‘Virtual screening: an endless staircase?’, Nature Reviews Drug Discovery. Nature Publishing Group, 9(4), pp. 273–276. doi: 10.1038/nrd3139. 8) Yang, Y., Adelstein, S. J. and Kassis, A. I. (2009) ‘Target discove- ry from data mining approaches’, Drug Discovery Today, 14(3–4) After the targets are found, the search for new compounds – future drugs – begins. A large number of small molecules with a higher than average binding affinity are the candidates for the chosen target. AI can efficiently filter out all “non- interest ing” molecules from the list. An impressive example of such filtering is a deep neural network built by the Insilico Medicine start-up, where 69 molecules with a specific anticancer characteristics were identified from 72 million entities stored in PubChem.5 The classic Machine Learning (ML) methods like classification or clusteri za- tion can also be fruitful in searching for hits. As reported by Schneider, “a hit rate of 40% was achieved for finding novel inhi bitors of arylamine Nacetyltrans- ferase among more than 700 million conformers.” 7 Boost the discovery Furthermore, candidate filtering is driven by the optimization of two characteristics – “drug-likeness”, “lead-likeness”. This process goes through hits discovery and ends up with a number of leads. AI can boost the discovery of hits and then leads by finding patterns in molecules’ proper- ties and simultaneously optimizing the two characteristics. Even if new leads are discovered, their chances of becoming a novel drug can be drastically diminished by adverse effects during Phase I trials. Moreover, the proba- bility of a new drug successfully passing Phase II trials is only 34% (Cumming et al., 2013). AI methods, e.g. automated Quanti- tative relationships (QSAR) or Matched molecular pair analy- sis (MMPA), are used to run chemical predictive modeling to foresee the adver- se effects and assist a process chemical optimization of the leads. QSAR models apply classification structure-activity regression and into drug to find the meaningful approaches pattering in chemical databases. With the incorporation of chemical predictive modeling development, AstraZeneca increased the solubility of compounds, an important factor affecting the “drug-likeness”, from 5.9% to 39.7%.2 Only a third (see above) of drug candi- dates found during the previous stages pass the trials. The use of drugs that passed Phase II and III trials but did not get into the market has a high potential for drug repositioning approach. The idea of repositioning drugs originates from the fact that a single target can be involved in many pathologies.1 In order to identify the new targets for old drugs, researchers use molecular networks6, where the hidden relationships between the components could be revealed with the help of AI methods – the Bayesian networks theory. This network-based ideology forms a foundation of network medicine and, ultimately, precision medicine. Many com- panies invest in acquiring expertise in this field, such as DataArt that is developing a mathematical theory and numerical algorithms of different types of Bayesian networks. The results are relevant to a broad range of drug development problems. Conclusion As you can see, AI can boost every stage of the drug development process: from identification of new targets to optimiza- tion of leads. There is no doubt that AI adoption brings considerable benefits by decreasing the length of the process by increasing the chances of successful clinical trials as well as reducing the R&D expenditure by reusing tested leads. Moreover, today‘s AI goes even further and can help plan clinical trials, collabora- tive filtering, for example, helps to choose participants that best fit the specific needs of a trial. 04-2017 „Health“ ls 37