Deep learning algorithms applied to medicine research

Deep learning algorithms applied to medicine research Servio Lima Reina Human IST institute University of Fribourg, Switzerland Abstract There are several applications for Deep Learning (DL) in medicine. Among the most popular are image recognition (e.g. used in radiology, MRI, fMRI, PET scans, etc.) and biochemical discovery. Current image recognition DL algorithms help to solve inaccuracy problems. They are 10% better in prediction than a human specialist. On the other hand, DL algorithms for biochemical discovery help to process a large quantity of databases [1] containing millions of molecules [2] that could be possible candidates for curing a disease such as Alzheimer, cancer or VIH. This paper summarizes the different DL algorithms used in modern medicine for the aforementioned applications. The study is not comprehensive but is a good starting point for understanding the state of the art of the technology and its potential. Introduction The science of choosing the right DL algorithm for solving a particular problem is not an easy task. The data scientist should have a deep understanding of mathematics, linear algebra, probability and calculus. The following is a selection of the most innovative and used DL algorithms applied in image recognition and biochemical discovery, two subfields that have the highest potential to change medicine as we know it. a. Image recognition Since the explosion of DL algorithms usage that started in 2012, image recognition has been improving year over year. Just a few years ago, DL algorithms beat the best radiologist in determining if an image contained a cancer tumor. Even the problem of scarce image data has been solved so far. Algorithms such as Generative Adversarial Networks (GANs) help to solve the lack of enough data to train DL algorithms by creating synthetic data. In [3], the authors created synthetic abnormal brain tumor MRI images that helped to improve the accuracy of predicting cancer. b. Biochemical discovery DL algorithms have also impacted the biochemical industry. By that, we mean the drug discovery and protein discovery industry. In regards to drug discovery, there are new companies such as Insilico [4] that are trying to cut research time and cost for delivering FDA approved drugs from decades to just one or two years. Several studies has been published so far: ChemGAN [5] is a model that uses GANs for drug discovery. DeepChem [6] is an open source library for drug discovery. In regards to protein discovery, at the recent 2018 competition held by CASP [7] (Critical Assessment of Techniques for Protein Structure Prediction), for the first time in human history, a deep learning algorithm outperformed the best scientist in predicting the shape of proteins, generated by amino acids. DeepMind enter into the competition and won against a total of 98 competitors. While the team in the second place only predicted 3 out of 43 proteins, DeepMind Alphafold [8] algorithm correctly predicted the structure of 25 proteins, out of the same set. There are twenty types of amino acids that could be arranged in different orders to create proteins. In fact, proteins fold in different shapes according to the interaction between the amino acids and the environment. These proteins fight viruses, transmit signals to coordinate biological processes, provide structural support for cells, transport molecules, and act as a storage devices, among other functions. The ability to predict the proteins shape allow us to better diagnose and treat diseases that are believed to be caused by misfolded proteins, such as Alzheimer, Parkinson, Mad cow and Transthyretin amyloidosis. Knowing how proteins operate opens up new potential within drug discovery. It will enable scientist to create much more effective and efficient cures for several diseases which will ultimately improve the quality of life for millions of persons around the world. On the other hand, any material with any set of desired properties could be constructed using proteins. We can build a material as hard as steel or soft as a jelly fish. We can build liquids that dissolve plastics or build the most flexible fibers known today. Proteins are like generalized materials: if we could understand the inverse function of material properties, meaning their proteins structures, it will change the world as we know it. Moreover, we could improve crop efficiency by creating insecticide proteins and we could regenerate human tissues through self-assembling proteins. In addition, a whole new breed of cosmetic and enhancement supplements for improving health could be introduced. As a summary, solving proteins folding does not only mean curing diseases. It means helping to solve world´s hunger, enhance our own physical capabilities and creating new bio materials. Thus far, scientists have been able to predict some proteins shapes using experimental techniques like cryo-electron microscopy, nuclear magnetic resonance and x ray crystallography. Each of these methods is a time intensive process and often takes years and considerable investments to infer a single protein structure. But the rapidly dropping cost of genetic sequencing has increased the amount of available genetic data to train on and this is the major reason why biologist are turning to deep learning as an alternative to traditional prediction processes for difficult proteins. Deepmind Alphafold uses Deep learning Residual Networks [9] (Resnets, Invented by Microsoft) to predict protein chain structures. It used known protein datasets to generate new protein structures from a given set of amino acids. c. Deep Learning algorithms explained There are several DL algorithms used for medicine such as Long Short Term Memory (LSTM) [10], Variational Autoencoders (VAE) [11] and Generational Adversarial Networks (GAN)[12]. In the following lines, we are going to describe succinctly the characteristics of each algorithm and use the graphical nomenclature depicted in figure 1. Figure 1: Nomenclature for describing DL algorithms [13] c.1 Long Short Term Memory (LSTM) LSTM are a special kind of Recurrent Neural Networks, capable of learning long term dependencies. They differ from RNN in that the latter are just capable of learning short term dependencies. The key to LSTM is the cell state. LSTM has the ability to remove or add information to the cell state, carefully regulated by structures called gates. The gates could be a sigmoid function where a value of 1 means let all the information pass and a value of 0 means the contrary. An LSTM has three of these gates, as depicted in figure 2. LSTM are particularly helpful for data in different timeframes. An example of this is health tracking, where data from a patient could arrive in real time (i.e. from a wearable device), could correspond to yearly data (i.e. after a through health examination) or from a span of several years (i.e. since childhood to adulthood). Figure 2: LSTM network [13] c.2 Variational Autoencoder (VAE) Autoencoders (AE) are DL algorithms where the input is equal to its output and there is a hidden layer that compress or represents the input data in a different way. Here the focus is to get the new representation of the data. Variational autoencoders (VAE) are similar to AE but they add Gaussian Noise to select decoding hidden layers and estimate the amount of information in that layer. AE and VAE are particularly suitable for anomaly detection. They could find anomaly patterns that are not easily found by doctors. This could be used for example for finding patterns of a new epidemic outbreak in a specific region of the world. (1) (2) Figure 3: Autoencoders (1) and Variational Autoencoder (2) [13] c.3 GANs (General Adversarial Networks) The basic idea of GANs is to take a collection of training examples and form a representation of their probability distribution. The typical method for this, is to infer a probability density function (PDF) directly. GANs are composed of two networks: A generator and a discriminator that play a game (i.e. minimax game) against each other. The objective of the generator is to produce and object (e.g. a picture or person) that would look like a real one. The goal of the discriminator is to be able to tell the difference between generated and real images. In this game, the generator becomes really good and this fact can be used for generation tumor images that are scarce in the medicine world. Figure 4: General Adversarial Network [13] d. Commercial solutions Deep learning in medicine is already a reality. In [14], the company helps doctors to alert about the potential of a brain stroke. In [15], the research is about helping oncologist to process a big body of research in their field and provides customized care for patients. [16] is focused in computational pathology for cancer treatment. Bibliography [1] Wishart DS, Feunang YD, Guo AC, Lo EJ, Marcu A, Grant JR, Sajed T, Johnson D, Li C, Sayeeda Z, Assempour N, Iynkkaran I, Liu Y, Maciejewski A, Gale N, Wilson A, Chin L, Cummings R, Le D, Pon A, Knox C, Wilson M. DrugBank 5.0: a major update to the DrugBank database for 2018. Nucleic Acids Res. 2017 Nov 8. doi: 10.1093/nar/gkx1037. [2] DiMasi, J. A.;Grabowski, H. G.; and Hansen, R. W. 2016. Innovation in the pharmaceutical industry: new estimates of r&d costs. Journal of health economics 47:20–33. [3] Hoo-Chang Shin, Neil A Tenenholtz, Jameson K Rogers, Christopher G Schwarz, Matthew L Senjem, Jeffrey L Gunter, Katherine Andriole, and Mark Michalski. Medical Image Synthesis for Data Augmentation and Anonymization using Generative Adversarial Networks. Workshop on Simulation and Synthesis in Medical Imaging - SASHIMI2018. [4] Insilico: http://insilico.com/ [5] Mostapha Benhenda. ChemGAN challenge for drug discovery: can AI reproduce natural chemical diversity?. Archiv preprint: 1708.08227v3 [6] Ramsundar, B. deepchem.io. https://github.com/deepchem/ deepchem, 2016. [7] CASP13: http://predictioncenter.org/casp13/ [8] DeepMind Alphafold: https://deepmind.com/blog/alphafold/ [9] Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun. Deep Residual Learning for Image Recognition. Computer Vision and Pattern Recognition. 2015 [10] Hochreiter, S., and Schmidhuber, J. 1997. Long short-term memory. Neural computation 9(8):1735–1780. [11] G´omez-Bombarelli, R.;Duvenaud, D.; Hern´andez-Lobato, J. M.; Aguilera-Iparraguirre, J.; Hirzel, T. D.; Adams, R. P.; and Aspuru-Guzik, A. 2016. Automatic chemical design using a data-driven continuous representation of molecules. arXiv preprint arXiv:1610.02415. [12] Ian J. Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, Yoshua Bengio. Generative Adversarial Networks. arXiv:1406.2661v1. 2014 [13] Complete chart of Neural networks: http://www.asimovinstitute.org/neural-network-zoo/ [14] SVIN: https://svin.viz.ai/ [15] VIECURE: https://viecure.com/ [16] PAIGE: https://www.paigeai.com/