LAUSR

Abstract Psychiatric disorders are complex and heterogenic brain related diseases which are still not fully understood. Indeed, while other fields of medicine have seen significant advances in diagnosis and prognosis over the past decade, the field of psychiatry has been relatively stagnant. In contrast to this lack of understanding, a variety of medications are available today for people who suffer from psychiatric disorders. For example, there are dozens of currently available medications for depression, which according to the World Health Organization is the leading cause of disability at the western world. The fact that pharmacological reagents, such as antidepressant medications, affect behavior suggests that the phenotypic behavior is the end result of biological processes which occur within the brain and body of the organism. These aspects are still not fully elucidated due to fact that many genetic alterations and environmental factors are probably involved in depression and impact the efficacy of antidepressants. Considering all possible factors and variants is very challenging. Furthermore, while some patients respond to specific medications with minor adverse effects, others either do not respond or experience adverse effects, suggesting that there are fundamental differences between patients diagnosed with the same disorder. Therefore, in order to predict efficacy and adverse effects of an antidepressant treatment, analysis of many genetic factors is needed. Additional elements should be considered as well; for example, the patient's childhood development and current life status (demographic factors), and medical history (clinical factors). Innovations in computer science now allow for the collection and sharing of large amount of data, and the advancement of next generation sequencing technologies facilitates accumulation of genetic information. Taking advantage of these advances, we combine machine learning algorithms with clinical knowledge to design successful prediction algorithms for efficacy and adverse effects of common antidepressant medications. These prediction algorithms are validated using large databases obtained from the NIH. It is important to note however, that machine learning algorithms are not designed for the analysis of genetic data, as the number of tested subjects using these algorithms has to be several scales higher than the number of tested features. Since the human’s genome is comprised of over 3 billion nucleotides, each considered as a genetic feature, even if whole genome sequence of all the world’s population (7-8 billion people) were available, it still wouldn’t be enough to apply machine learning algorithms. At Taliaz we developed a pipe-line of algorithms and procedures which allows the use of machine learning algorithms on genetic data combined with clinical and demographic information. Using this pipeline, we have succeeded in designing prediction algorithms for the efficacy and adverse effects of current antidepressants, with up to 77% accuracy (76% specificity and 79% sensitivity). In my talk I will describe the features pre-scoring pipeline designed by Taliaz, and provide examples of prediction algorithms that we have designed using this pipeline.