Artificial intelligence, physiological genomics, and precision medicine

Abstract

Big data are a major driver in the development of precision medicine. Efficient analysis methods are needed to transform big data into clinically-actionable knowledge. To accomplish this, many researchers are turning toward machine learning (ML), an approach of artificial intelligence (AI) that utilizes modern algorithms to give computers the ability to learn. Much of the effort to advance ML for precision medicine has been focused on the development and implementation of algorithms and the generation of ever larger quantities of genomic sequence data and electronic health records. However, relevance and accuracy of the data are as important as quantity of data in the advancement of ML for precision medicine. For common diseases, physiological genomic readouts in disease-applicable tissues may be an effective surrogate to measure the effect of genetic and environmental factors and their interactions that underlie disease development and progression. Disease-applicable tissue may be difficult to obtain, but there are important exceptions such as kidney needle biopsy specimens. As AI continues to advance, new analytical approaches, including those that go beyond data correlation, need to be developed and ethical issues of AI need to be addressed. Physiological genomic readouts in disease-relevant tissues, combined with advanced AI, can be a powerful approach for precision medicine for common diseases.

INTRODUCTION

Big data are a major driver of precision medicine.¹ In 2011, the US National Research Council developed a comprehensive plan for how to advance precision medicine, specifying the development of a “New Taxonomy” for human diseases, to integrate molecular, environmental, and phenotypic data (41a). Big data alone are limited without analysis methods to transform the data into clinically actionable knowledge. Thanks to the advances in parallel computing, many researchers are turning to artificial intelligence (AI) to develop robust data analysis approaches (5, 42). In this article, we will first give a broad introduction to machine learning (ML), which is a major approach of AI, and then discuss how physiological genomics, combined with ML, can be a critical and powerful approach to advancing precision medicine for common diseases.

MACHINE LEARNING

ML is an approach to achieving AI that utilizes modern computer and mathematic algorithms to build models based on available data sets, in which the model itself can improve with experience. ML can recognize complex combinations of variables that reliably predict an outcome (42). ML is useful for analyzing large, heterogeneous data sets, such as genomic or epigenomic data (34). Based on how the machine “learns” and depending on the question the “learning” is trying to answer, ML can be categorized into supervised, unsupervised, and reinforcement learning. The major application of ML in precision medicine comes from supervised learning where the patient outcome is known (labeled). Unsupervised learning utilizes uncategorized data sets to find hidden patterns or grouping in data. Reinforcement learning utilizes simulated data and is not generally used in medicine.

Traditionally, supervised learning is conducted in five phases (Fig. 1). First, an algorithm is selected with the goal of maximum learning of the training set. Several examples of ML software libraries used with Python and R are shown in Table 1. Second, preprocess treatment of the data, or data extraction, transformation, and loading (ETL), is performed. The data are formatted and cleaned to remove incomplete data and address missing information. Then the data may be transformed; common transformations include scaling (taking attributes with a mixture of scales and putting them on the same scale such as between 0 and 1), decomposition (splitting attributes into constituent parts), or aggregation (combining related attributes into a single feature). Third, the data are split into training and testing sets, with a majority of data assigned to the training set. The training set is then used to develop a model or “train” the AI. The parameters in the model are tuned to optimize the final model built. Fourth, the model is used to make predictions on the testing set, and the predictability of the model is assessed. K-fold cross-validation can be used to assess the stability of the model. In the cross-validation, the data are divided into k subsets. Now the built model is repeated k times, such that each time, one of the k subsets is used as the test set and the other k − 1 subsets are put together to form a training set. The prediction error rate is averaged over all k trials to obtain total effectiveness of the built model. Other statistical methods may be utilized to evaluate the predictability of the model, including sensitivity, accuracy, and receiver operating characteristic (ROC) curves. Area under the curve (AUC) may be calculated from the ROC curve, though major limitations of AUC utilization have been identified (21, 54). AUC is a measure of how accurate the model is in predicting the actual patient outcome, with an AUC of 1 representing perfect accuracy. Steps 1 through 4 may be conducted for many different algorithms to select an algorithm with optimum predictability. Lastly, the model is deployed to make predictions on other clinical data sets (1, 34).

Fig. 1.

Schematic of machine learning (ML) steps for supervised learning. ML is conducted in five phases. 1) An algorithm is selected. 2) Data extraction, transformation, and loading (ETL) are performed. The data are formatted and then cleaned to remove incomplete data and address missing data. The data may also be transformed by transformations such as scaling, decomposition, or aggregation. 3) The data are split into training and testing sets. The training set is then used to develop a model or “train” the artificial intelligence (AI). The parameters in the model may be tuned to optimize the final model built. 4) The model is used to make predictions on the testing set, and the predictability of the model is assessed. Cross-validation can be performed to compare the predicted outcome to the actual patient outcome. Sensitivity, accuracy, and receiver operating characteristic (ROC) curves may also be used to evaluate the predictability of the model. 5) The model is deployed to make prediction for other data sets.

Table 1.

Examples of software libraries used with Python and R to perform various ML tasks

Task	Example Software Libraries Used with Python	Example Software Libraries Used with R
Data preprocessing	NumPy, SciPy, pandas, Matplotlib, seaborn	R(data.frame), dplyr, ggplot2
Training testing data split	scikit-learn	Caret, mtcars, caTools, dplyr
Linear regression	scikit-learn	Stats, H2O
Classification_K-NN	scikit-learn	Class
Classification_tree	scikit-learn	C50
Classification_SVM	scikit-learn	e1017
Classification_naive bayes	scikit-learn	e1017
Classfication_random forest	scikit-learn	randomForest, H2O
Clustering PCA/K-Means	scikit-learn	Stats, H2O
Association analysis	scikit-learn	arules
Dimensional reduction	scikit-learn	Stats, H2O
Deep learning_ANN	Keras, Tensorflow, Theano	H2O, neuralnet
Deep learning_CNN	Keras, Tensorflow, Theano	KerasR, MXNetR
Deep learning_RNN	Keras, Tensorflow, Theano	rnn

ML, machine learning.

Games have frequently served as milestones for large advancements in ML (41). Early accomplishments were in backgammon (50), checkers (44), chess (6), Jeopardy! (14), and Atari (40). These games are perfect-information games, meaning that every player has identical, complete information about the current game status and can see all possible scenarios for every move. Once simple perfect-information games were mastered, programmers turned their focus toward the game Go, which is considered the most complex perfect-information game.

Go is an ancient Chinese game with simple rules, but it requires highly complex strategy because of the 19×19 grid board, which contains 10¹⁷⁰ possible configurations (18). Google researchers developed an algorithm named AlphaGo. AlphaGo was first trained with supervised learning using 30 million iterations from professional games played by humans. Then AlphaGo played against itself using reinforcement learning, simulating thousands of random games using Monte Carlo tree search programs. AlphaGo’s success was due to its ability to overcome the complexity of the game by developing prioritization methods that reduced the number of possible moves the algorithm had to consider for any one given move. In 2015 and 2016, AlphaGo successfully defeated European champion Fan Hui and 18-time world champion Lee Sedol (46). More recently, Google developed AlphaGo Zero, an AI that utilizes a novel reinforcement learning algorithm that initiates training from completely random play against itself rather than through observation of human games. AlphaGo Zero was clearly stronger than previous versions of AlphaGo (47).

Imperfect-information games like poker, which involve chance or incomplete information, are more challenging to master with ML because they require more complex reasoning. All possible cards are known in poker, but players have uncertainty in the cards their opponents have and asymmetric knowledge about the game status due to each player’s private hand. DeepStack is an AI trained in heads-up no-limit Texas hold’em, a two-person poker game. DeepStack used reinforcement learning to play thousands of simulated poker hands against itself. The key to DeepStack’s success was incorporating recursive reasoning methods that effectively addressed issues with incomplete and asymmetric information. Although DeepStack has beaten several professional poker players, the technology is still ineffective in the standard 10-person no-limit Texas hold ’em game because of increased incomplete information, since the number of cards in opponent hands are 2/52 in two-person games and 18/52 in 10-person games (41).

APPLICATION OF AI IN MEDICINE

Many of the techniques and algorithms developed for games have been extrapolated to develop AI to assist in medicine. Watson, an AI developed by IBM originally intended to play Jeopardy!, is a cognitive natural-language processing system, meaning it has the ability to read and interpret words. In a recent study, the degree of treatment recommendation agreement was compared between Watson for Oncology and a 15-member multidisciplinary tumor board of physicians for 638 breast cancer patients. Treatment recommendations made by Watson for Oncology were highly concordant (93%) with the multidisciplinary tumor board, except for in cases of stage IV triple negative breast cancer (P < 0.05) (48).

Some other notable advances have been made using image recognition strategies for disciplines that utilize image diagnostic tools such as ophthalmology, pathology, and radiology. In a few recent studies that utilize image recognition, ML was compared with human doctors performing the standard of care to assess our progress in ML. Deep neural networks were used to train AI to make diagnoses on images of skin cancer (15) and diabetic retinopathy (19). When compared with professional dermatologists and ophthalmologists, the AI performed just as well and in some cases slightly better than their human counterparts.

In another study, multiple types of ML algorithms were compared with the most widely accepted guidelines in clinical practice, the American College of Cardiology/American Heart Association (ACC/AHA) guidelines, to predict the first cardiovascular event over a 10 yr period (52). Compared with ACC/AHA guidelines (AUC 0.728), the four ML algorithms improved predictions slightly: random forest (AUC 0.745), logistic regression (AUC 0.760), gradient boosting (AUC 0.761), and neural networks (AUC 0.764), with neural networks preforming the best. In-depth reviews of advances made in genomics (34), epigenomics (23), and proteomics (31) using ML can be found elsewhere.

KEY TO THE SUCCESS OF AI FOR PRECISION MEDICINE: DATA QUALITY AND RELEVANCE

While the performance of AI was impressive in the medical studies discussed above, one must ask why AI is not performing even better. For instance, in the study of cardiovascular risk prediction, the AUC was only around 0.75 for each of the AI algorithms even though the AI was provided with practically all available data. It is certainly possible that AI needs even more data or better algorithms to improve its performance. The quantity of data used for training sets is essential for maximizing the performance of ML algorithms. ML algorithms are extremely data-hungry, frequently requiring millions of data points to perform optimally. Additionally, programmers have found that using more data with simple algorithms produces more effective models than using less data with complex algorithms (20). Indeed, much of the current effort in precision medicine is focused on obtaining ever larger sample sizes and ever more DNA sequence data and electronic health records including mobile health records.

However, it is just as important to ensure that the data obtained are accurate and relevant. Data accuracy is important because it significantly affects the predictability of the model. The common “garbage in, garbage out” principle applies to ML, meaning that the quality of the predictions made by the model will never exceed the quality of the training set used to build the model. While conventional statistical methods and ML methods can be employed during ETL to check data quality and remove outliers, these methods have limited effectiveness due to the size of the data sets. Most of the time, ML itself does not judge the data quality and instead assesses all of the data equally to build the model. Thus, inaccuracy of data points will greatly impact the integrity of the final model built. Incomplete or even wrong data points will introduce biases. For example, one study that assessed the impact of data accuracy on model predictability found that using a training set with 10% inaccurate data increased the incorrect edges in a Bayesian network by 325% (27).

It is also important to collect the types of data that are relevant to the disease of interest from disease-applicable tissues. Many recent attempts to use ML to answer precision medicine questions have relied heavily on DNA sequence data (10), even though the vast majority of common diseases are widely believed to result from both (epi)genetic and environmental factors as well as their interactions. Unfortunately, causative environmental and lifestyle factors for many of these complex diseases are difficult to quantify, and the specific environmental exposures driving disease progression for a particular patient and their effect size are unknown. For these reasons, functional genomics are an appropriate surrogate, since they measure cumulative effects of environmental factors on the genome and provide biological links between DNA sequences, environmental factors, and the disease. Functional genomic readouts can take the forms of epigenetic modifications of the DNA, profiles of RNAs (mRNA, microRNA, lncRNA, etc.), proteins, and small metabolites, and physiological function (32, 37).

The logistic bottleneck for utilizing functional genomic readouts for precision medicine in common diseases is the difficulty in obtaining patient tissues directly relevant to these diseases. Because of this limitation, researchers have devoted significant effort toward identifying surrogate genomic readouts in body fluids or other readily obtainable materials. While this method may be useful, disease-applicable tissues are more likely to possess clinically actionable molecular information (51). There are important exceptions to the difficulty of obtaining disease-applicable tissues. One example is the kidney needle biopsy. Although it is an invasive procedure, it is done at a high enough frequency that it provides nephrology a significant advantage in driving forward functional genomics-based precision medicine. Importantly, it is possible to isolate and analyze specific tissue types or substructures in kidney biopsy samples, such as glomeruli, which would substantially mitigate concerns about tissue complexity. Another example is resistance arterioles, which are important determinants of cardiovascular physiology and can be readily obtained through biopsy (53). Physicians and scientists need to work closely together to collect tissues relevant to common diseases for testing the functional genomic approach to precision medicine. Importantly, the tissues need to be collected, processed, and stored in a manner appropriate for the intended analysis. While it is challenging to do so, the payoff is potentially very high.

EXAMPLES OF ML APPLICATION IN FUNCTIONAL GENOMICS STUDIES USING DISEASE-APPLICABLE TISSUES

Functional genomic data have been used to advance the precision medicine of cancer (25, 26). Recent progress has been made in utilizing functional genomics to inform precision medicine of common noncancer diseases including kidney disease. The following examples illustrate how functional genomics data obtained from disease-applicable tissues can be analyzed with ML in addition to other statistical methods to work toward the development of precision medicine for common noncancer diseases. Reeve et al. (43) collected 1,208 kidney transplant biopsies to assess rejection-related disease by using archetypal analysis of mRNA phenotypes combined with principal component analysis (PCA), a dimensionality reduction method. Archetypal analysis is a type of unsupervised learning that identifies extreme or pure phenotypes within a training set. They calculated bootstrap corrected C-statistics, which are analogous to AUCs, and found that their archetypal analysis scores (C-statistics 0.73) were more predictive of allograft loss than histological diagnosis using the Banff classification (C-statistics 0.60) (P = 3 × 10⁻⁶).

In another study, Liu et al. (35) examined cell type-specific differential mRNA expression in glomeruli isolated from IgA nephropathy patients. They used “in silico nanodissection,” an ML algorithm that predicts cell type-specific transcripts based on large data sets, to identify mesangial and podocyte cell-specific genes that were differentially expressed between IgA nephropathy and healthy control patients. They concluded that the mesangial cell genes were highly correlated to serum creatinine (P = 0.024) and estimated glomerular filtration rate (eGFR) (P = 0.025) progression for up to 4 yr following biopsy. Mass spectrometry was used to examine differentially expressed pathways in these samples as well as human mesangial cells that were incubated with serum from IgA nephropathy patients. They found several inflammatory pathways that were differentially expressed both in vivo and in vitro, along with pathways related to inducible nitric oxide synthase and endothelial nitric oxide synthase.

In a recent study from our laboratory, we used human formalin-fixed paraffin-embedded renal cortex samples to examine differentially expressed microRNAs among four different types of chronic kidney disease (2). MicroRNA are small noncoding RNA molecules that regulate gene expression through targeting specific mRNA for translational repression or degradation, thus leading to change in protein expression that can greatly affect the structure and function of a cell (3, 16, 55). Laser capture microdissection was used to collect glomeruli and proximal tubules, allowing tissue-specific whole genome microRNA libraries to be generated. This approach could elucidate diagnostic biomarkers within a particular nephron segment affected by disease pathogenesis that would not have otherwise been detected had the RNA been extracted from whole kidney cortex tissue.

We identified a number of microRNAs that were differentially expressed in glomeruli and proximal tubule samples between the control group vs. each disease, and in glomeruli between any two disease cohorts. Interestingly, we didn’t detect any differentially expressed microRNAs between any two disease cohorts in the proximal tubule samples. This may suggest that since the diseases studied were all glomerular diseases, the disease-specific effects on microRNA occur in the glomeruli and the pathological changes in the proximal tubules represent changes that are common across all proteinuric renal diseases.

ML also comes with its own limitations. Overfitting occurs when the model makes predictions on spurious correlations rather than causal correlations. Classification and regression trees are particularly at risk for overfitting, since these methods can continue to split data sets until a perfect correlation is observed (17). Another limitation with ML is the “black box” problem. With many types of ML, the data and model specifications are entered into the ML software, then the algorithm makes predictions without elucidating the algorithm’s decision-making process. Without this information, researchers can miss warning signs of defective models.

Effective validation strategies are needed to mitigate risks associated with overfitting and black box analysis. One method is to evaluate the predictability of models by using multiple independent patient cohorts to assess the generalizability of study findings (13). Another approach that we have utilized in our laboratory is to verify study results in in vitro or in vivo model systems. In a recent study of human salt sensitivity, we identified five urinary metabolites significantly associated with systolic blood pressure and one urinary metabolite significantly associated with diastolic blood pressure by using random forest and generalized linear mixed model analysis. To validate our results, we performed a proof-of-principle experiment to examine the effects of β-aminoisobutyric acid, a metabolite that was found to be significantly correlated to systolic blood pressure, on high-salt-induced hypertension in Dahl salt-sensitive (SS) rats. Treatment with β-aminoisobutyric acid significantly attenuated high-salt-induced hypertension in SS rats, supporting the relevance of the inverse relationship between β-aminoisobutyric acid and blood pressure we observed in humans (9).

FURTHER DEVELOPMENT OF ANALYTICAL APPROACHES INCLUDING THOSE BEYOND DATA CORRELATION

As we improve the data that we use in ML, there is also a great need to develop new analytical approaches that take into account the challenges and limitations of the study design and the data used for the training set, similar to how AlphaGo and DeepStack were built to address the challenges in data complexity and imperfect information, respectively. Current ML algorithms are based essentially on analysis of correlation. Correlation-based approaches require high data density relative to dimensionality, which is difficult to satisfy in functional genomic studies.

This challenge is compounded by the n of 1 problem in precision medicine where data relevant to a given patient may be limited to the data available from that patient or a small group of highly similar patients. For this reason, an integrative approach is needed to build more complete training sets that incorporate genomic data, functional genomic data, and relevant information from electronic health records such as demographics and laboratory reports. While it may never be possible to solve the n of 1 problem in precision medicine for most diseases, an integrative approach may allow us to do “deep grouping” for diseases, where we are able to identify a set of markers that can stratify patients to predict outcomes and determine optimal treatment plans.

While correlation-based analysis can be valuable for deep grouping of patients, biological processes underlying disease operate with kinetic properties that may not be easily captured by correlation-based analysis. Construction of kinetic regulatory networks of molecular and physiological events would be complex, but critical for understanding disease and ultimately improving precision medicine, including repurposing of existing therapeutics. However, similar to ML, most of the data analysis methods that have been applied in attempts to construct molecular and physiological regulatory networks are based on correlation or some variant thereof.

There is an urgent need to develop new analytical approaches that can effectively construct regulatory networks utilizing functional genomic readouts. Any analytical approaches developed for this purpose would need to perform well in the context of the disease and disease-applicable tissues, not just in highly reductionist systems. Integration of correlation-based methods and mechanistic modeling could be one such fruitful approach (28, 33). Another approach may be through the study of neuroscience, similar to how convoluted neural networks were developed. The algorithms, architectures, and functions utilized by the human brain can be a source of inspiration for the development of new algorithms that incorporate methods for complex reasoning and decision making (22).

ETHICS OF AI

While AI has many potential benefits in medicine, researchers must also consider the potential social, ethical, and economic impacts of AI (39). The safety of AI is an important concern, especially as AI gains increasing levels of autonomy. A more fundamental challenge that AI poses is the legal definition of personhood (29) and the psychological perception of human identity as AI becomes increasingly human-like. The issue is further compounded by the inevitable development of increasingly complex human-machine hybrids. A full discussion of these issues is beyond the scope of this article.

However, there are more immediate concerns about the use of AI in precision medicine. For example, AI can learn bias and prejudice values when they are present within the data set, leading to unfair or inaccurate predictions (4, 12, 24). Bias can be introduced when a specific patient population is not included in the training set. For instance, a study that predicted the probability of death for pneumonia patients to determine whether the patient were to be admitted into the hospital or treated as an outpatient mistakenly predicted treating asthma patients as outpatients. This error occurred because the training set used included patients who required further care, but a majority of asthma patients were immediately transferred to intensive care, so they were not adequately represented in the data set (7). While the bias involved in this example was related to disease representation, incorporation of prejudice values can occur in a similar manner. AI can learn preexisting inequalities from society represented in the training set, resulting in bias toward historically disadvantaged populations (4). Additionally, data can be manipulated and misinterpreted. Issues like these highlight the importance of integration of data quality and high ethical standards to mitigate bias when AI is used for precision medicine.

Preliminary efforts are underway to address the potential social, ethical, and economic impacts of AI. In 2016, the White House, the European Parliament, and the UK House of Commons each released reports summarizing how government and society should prepare for the development of AI and its widespread use. While these reports identify the potential negative impacts of AI, they fail to outline long-term implementation strategies to assess risk and harm (8). Much work is left to develop strategies to identify and address the negative impacts of AI on society, so that we can work toward the development of a “good AI society” (41b) and unleash the tremendous benefits of AI including advancement in big data-based precision medicine.

GRANTS

The authors acknowledge grant support from the National Institutes of Health (HL-121233, HL-125409, GM-066730, HL-116264, HL-082798), the Advancing a Healthier Wisconsin Endowment, and the American Heart Association (15SFRN23910002).

DISCLAIMERS

The funding sources had no role in the preparation, review, approval, or decision to submit the manuscript for publication.

DISCLOSURES

No conflicts of interest, financial or otherwise, are declared by the authors.

AUTHOR CONTRIBUTIONS

A.M.W. and Y.L. prepared figures; A.M.W. and M.L. drafted manuscript; A.M.W., Y.L., K.R.R., F.J., P.L., and M.L. edited and revised manuscript; A.M.W., Y.L., K.R.R., F.J., P.L., and M.L. approved final version.

Glossary

Area under the curve (AUC)

A measure of the predictability of a statistical model by calculating the probability that the prediction for a random data point will match the actual outcome across all possible thresholds on an ROC curve. An AUC of 1 represents perfect accuracy (21).

Big data

Data sets that are high volume (quantity, variety of sources), high velocity (speed of data generation and processing), and/or high variety (type, variety of formats), making traditional data processing techniques inadequate (38).

Machine learning (ML)

An application of AI used to build models that can learn from data sets to make predictions. The model itself can improve with experience without being explicitly programmed to do so (45).

Precision medicine

A customized medical approach used to prevent, diagnose, or treat disease based on a patient’s unique genetic, environmental, and lifestyle factors (36).

Reinforcement learning

A type of ML that specifies how the AI should take actions for a given environment to work toward maximum benefit of a cumulative reward. This type of machine learning is often used in simulation-based optimization of games (1).

Supervised learning

A type of ML in which the training set is labeled, with the goal of predicting a known outcome (1, 30, 34).

Unsupervised learning

A type of ML in which the training set is unlabeled, with the goal of finding naturally occurring patterns or groupings within the data (1, 34).