Deep Learning Analytics for Predicting Prognosis of Acute Myeloid Leukemia with Cytogenetics, Age, and Mutations Andy Nguyen, M.D., M.S. Medical Director, Hematopathology, Hematology and Coagulation Laboratory, Memorial Hermann Laboratory Professor of Pathology and Laboratory Medicine, University of Texas-Houston, Medical School GeneMed Feb 2018
Outline of talk Define application of Deep Learning method, a technological breakthrough, to big-data analytics Describe our study using Deep Learning algorithm to predict prognosis for acute myeloid leukemia (AML) using cytogenetics, age, and mutations Financial Disclosures: No relevant financial relationships with commercial interests to disclose
Deep learning and genomic medicine Big companies are analyzing large volumes of data for business analysis and decisions, using Deep Learning technology (Google s search engine, Google Photo, automobile companies: self-driving cars, IBM s Watson). The application of deep learning to genomic medicine has a promising start; it could impact personalized diagnostics, and treatment. The genotype-phenotype divide, our inability to connect genetics to disease phenotypes, is preventing genomics from advancing medicine to its potential. Deep learning can bridge the genotype-phenotype divide, by incorporating an exponentially growing amount of data, and accounting for the multiple layers of complex biological processes that relate the genotype to the phenotype. This gap necessitates the application of Deep learning, a more recent type of machine learning
Machine Learning Machine learning: explores the construction of algorithms that can learn from and make predictions on data - i.e. gives software the ability to learn without being explicitly programmed Numerous machine learning methods: Decision tree, Cluster analysis, Support vector machine, Random forest, Bayesian, Regression analysis, Neural network. Neural network (inspired by biological neural networks): artificial nodes ("neurons ) are connected together to form a network for prediction/classification tasks Fire/not Fire
Early Generation of Neural Networks with Supervised Training (model is trained with known or labeled outcomes) pos neg Calculating connection weights 3 types of data: -Training -Validation -Testing -New (for new cases)
Disadvantages of Early Networks They relies only on labeled data (known outcomes) for training. However, labeled data is often limited, and thus for many problems it is difficult to get enough examples to fit the parameters of a complex model. Given the high degree of expressive power of deep networks, training on insufficient data would also result in overfitting. Training a network with multiple hidden layers using supervised learning: (1) Parameters often do not converge; i.e. being stuck in local minim; (2) Model not scaling well (diffusion of gradients causing poor learning in earlier hidden layers) Iterations: Convergence to minimize error Local minima: No convergence
Deep Learning (3 rd Gen Neural Network) A major breakthrough in 2006: Hinton (U of Toronto) won a contest held by Merck to identify molecules that could lead to new drugs. The group used deep learning to zero in on the molecules most likely to bind to their targets. Deep Learning algorithms: (1) Unsupervised learning ->allows a network to be fed with raw data (no known outcomes) and to automatically discover the representations needed for detection or classification (2) Extract high-level & complex data representations through multiple layers; avoid problems of last-gen networks (previous slide) Supporting hardware: multiple graphics processing units (GPU)
A Deep Learning Neural Network to Detect Image: Extracting higher-level Features With Unsupervised Learning Feature extraction: -Each hidden layer applies a nonlinear transformation on its input to transform the input to higher level of representation in its output. -Multiple levels of abstraction of the image: from pixels to complex shapes and objects defining a human face -Deep learning process works similarly for non-visual objects
Our Study Objective: AML Prognosis The risk stratification of acute myeloid leukemia (AML) based on recurrent chromosome abnormalities has been well established. Similarly, some mutations in AML cases with no chromosome abnormalities are known to play a role in risk stratification. Multivariate statistic analysis becomes a challenge with addition of numerous input variables (mutations from next-gen sequencing). Risk classification is difficult to assess for a patient with a particular profile N ENGL J MED 366;12 Mar 22, 2012
Our Hypothesis Hypothesis for this study: Deep Learning can be utilized to accurately predict prognosis of AML using combined data from several sources. Specifically we attempt to determine the correlation between prognosis and cytogenetics, age, and mutations in acute myeloid leukemia. Our Study Materials 94 AML cases from TCGA (The Cancer Genome Atlas) database. Data include cytogenetics, age, mutations, prognosis PX (days to death, DTD). Cytogenetics (10 common abnormalities): t(8;21), inv(16), t(15;17), t(9:11), t(9;22), trisomy 8, del (7), del (5), del (20), complex chr abnls Mutations (23 most common): FLT3, NPM1, DNMT3A, IDH2, IDH1, TET2,RUNX1, TP53, NRAS, CEBPA, WT1, PTPN11, KIT, U2AF1, KRAS, SMC1A, SMC3, PHF6, STAG2, RAD21, FAM5C, EZH2, HNRNPK
Method: Deep Learning- Programming Platform We design Deep Learning neural networks with stacked (multi-layered) autoencoder in R language. R is a programming language for statistical computing and graphics supported by the R Foundation for Statistical Computing, currently used extensively in machine learning In this study, we use functions obtained from R packages which are available from the Comprehensive R Archive Network Stacked autoencoder network: Pre-training with unlabeled data (unsupervised) -> Stacked Autoencoder Algorithm <- Fine-tuning with labeled data (supervised)
AML Prognosis-Results Our Deep Learning model which incorporates unsupervised feature training find excellent correlation between prognosis PX (DTD) with 10 cytogenetics, abnormalities age, 23 most common mutations Median DTD is 730 days Good PX for DTD >730 days; Poor PX for DTD<= 730 days Ten-fold validation: exclude 10% of the cases at a time to train the network and use the resultant network to test these excluded cases Mean Accuracy of 81%, Cross Validation Data Sets Sensitivity of 74%, and Specificity of 86% 1 90% Mean accuracy if some input is excluded: Exclusion of cytogenetics -> 67% Exclusion of mutations -> 74% Exclusion of age -> 61% This indicate critical contribution of all input categories 2 80% 3 80% 4 90% 5 100% 6 80% 7 70% 8 70% 9 80% 10 75% Mean Accuracy: 81%
The Ranking of all Input Data used in Training (The ranking of input data is based on the sum of the absolute weights of the connections from the input node to all the nodes in the first hidden layer) Using the 14 top-ranked attributes (out of 34) -7 chrom abls: tri8, del5, del7, Complex, t(8;21), inv(16), t(15;17) -Age -6 mutations: FLT3, NPM1,TP53, DNMT3, KIT, CEBPA Accuracy= 83% (slightly better than 81% using 34 original attributes, likely due to data redundancy)
SUMMARY Deep Learning method, a disruptive technology, is predicted to be an integrated part in future practice in molecular diagnosis & prognosis prediction using nextgen sequencing data. Our preliminary study demonstrated a practical application in this area Limitations of our preliminary study: - The relatively small size of cohorts (94 cases) due to limited data in TCGA database - This study nevertheless provides excellent preliminary results for future studies that include many more cases, more mutation data, and other clinical data such as co-morbidity index. With more data, the expected accuracy would be higher than that of this preliminary study (> 83%) The successful validation of such deep learning software would be of tremendous value to personalized treatment of AML patients, i.e. stratifying treatment for each patient based on predicted prognosis The software s database can be continually kept up-to-date by adding new patients data (with new tests, etc.) to preserve its predicting ability Using input ranking techniques, critical parameters which impact prognosis can be detected -> helps to identify sets of important data to predict prognosis (novel patterns)