1: Precision medicine

Sir William Osler described, “Variability is the law of life, and as no two faces are the same, so no two bodies are alike, and no two individuals react alike and behave alike under the abnormal conditions which we know as a disease” (Osler, 1903). This impeccably summarizes the challenge of understanding the overlaying mechanism for disease, its regulation, and treatment in medicine or associated sciences. Diseases like cancer involve complex underlying causes among the biological processes ranging from the molecular to the cellular level. These genetic variations may not always be one-to-one with the patient phenotype. Different genetic aberrations may result in similar or different phenotype, making it difficult to recognize the specific cause of the disease (Kim et al., 2016; Mardinoglu and Nielsen, 2012). Therefore, often a treatment used against a particular disease may not work on all the patients (Fig. 1).

In 1999, Francis Collins, one of the pioneers of the Human Genome Project, gave the foundation document for precision medicine (Collins, 1999). The following year, he briefed in a news conference, how the completion of the Human Genome Project is going to accelerate precision medicine leading to the complete transformation of therapeutic medicine (Wade, 2010). This gave the base concept for improvising patient care, by planning the treatment based on the maximum information gained for a patient along with the existing knowledge base of the disease. In the last 20 years or so, the study of genomes and other omics disciplines (proteomics, metabolomics, transcriptomics, etc.) has rapidly advanced the scenario and is being routinely adapted along with the clinical examinations for decision-making.

Precision Medicine can be best described as “an emerging approach for disease treatment and prevention that takes into account individual variability in genes, environment, and lifestyle for each person” (Ashley, 2015; Collins and Varmus, 2015; Hodson, 2016). Besides the patient-doctor relationship being the core, the new biomedical information might add substantial information beyond the observable signs and symptoms (König et al., 2017). Thus, Precision Medicine implies the novelty of harnessing this wide array of patient’s data, including clinical, poly-omics (genomic, transcriptomic, proteomic, metabolomic, epigenomic, etc.), and lifestyle information (Pinho, 2017) (Fig. 2). This gives the opportunity of identifying sub-populations who vary in their prognosis and treatment response, by understanding their difference in biology for that particular disease (Uddin et al., 2019).

The major contribution for this has been the rapid advancement of genomic and computational technologies, and with the costs of genetic tests plunging, the new targeted therapies are making it increasingly possible to prevent or treat illnesses based on an individual patient’s characteristics (Love-Koh et al., 2018; Weil, 2018). It provides the opportunity to the clinicians, bioinformaticians, biomedicine, and associated researchers to develop new approaches for detection and diagnosis, prognosis, and other applications by analyzing a wide range of biomedical data—including molecular, genomic, cellular, clinical, behavioral, physiological, and environmental features. For example, in “precision oncology,” i.e., precision medicine for cancer, a few of the obstacles currently been addressed are: the unexplained drug resistance observed in certain patient cohorts, genomic heterogeneity of tumor, risk assessment and means for monitoring responses, tumor recurrence, and drug combinations to be used for treatment (Collins and Varmus, 2015).

Machine learning combines the strengths of computer science, mathematics, and statistics, providing the computational capabilities to learn from the data, and thus has the competency to provide a platform aiding precision medicine for learning from massive sets of clinical, poly-omics, and other datasets. Thus, in this chapter, we review and discuss the basic applications of machine learning on the heterogeneous datasets comprising poly-omics and clinical data, focusing on methods in use and its opportunities in precision medicine. In the following sections, firstly we discuss machine learning in brief; thereafter, the use of machine learning in precision medicine and along with discussing example applications showing its potential use in the detection and diagnosis, prognosis, therapy response, and in the discovery of new biomarkers and drug candidates; we conclude looking into the opportunities for addressing a few other challenges that can be addressed using precision medicine.

2: Machine learning

As modern computing evolved in the early part of the 20th century, the inception of “thinking machines” was aroused by the advent and evolution of the “electronic or digital computer” (Turing, 1950). The aim was to apply this computational capacity toward elucidating patterns and inferences from the dataset which is difficult to achieve by conventional statistical methods. In 1959, Arthur Samuel coined the term “Machine Learning,” explaining how from a given set of minimum parameters a computer can be programmed, so that it will learn to play a better game of checkers in comparison to the humans (Samuel, 1959). Thereafter, since the 1960s, machine learning has been at the forefront, enabling learning from data. Furthermore, in the 1990s, Tom Mitchell explained, the focus for machine learning to be, “developing a computer program, which is said to learn from experience E with respect to some class of tasks T and performance measure P, if its performance at tasks in T, as measured by P, improves with experience E” (Mitchell, 1997). A simple example of this would be a computer program that could predict what kind of television program a person likes, based on the parameters like age, gender, occupation, and geographical location. Or based on customer’s bank transaction history, a program could identify and/or predict fraudulent transactions. In both these scenarios, the computer program is considered to learn from the available data and its prediction performance improves as more data are made available to it.

Machine learning aims to provide learning algorithms (computer programs), which can be applied for analysis on the input data, for predicting the corresponding output values, identify patterns and trends within an acceptable performance range, and improvise based on experience (Fig. 3).

Depending on how the algorithm makes this learning, machine learning can be broadly categorized into supervised, unsupervised, and semisupervised learning (Ayodele, 2010; Brownlee, 2016; Maetschke et al., 2014). Supervised learning can be defined as learning a function, which given a sample of labeled data (i.e., data with known outputs), approximates a function that maps inputs to outputs (example—classification or regression-based learning) (Caruana et al., 2008; Caruana and Niculescu-Mizil, 2006). Unsupervised learning can be defined as learning a function in data that does not have any labeled outputs; therefore, it goals to infer the natural structure present within a set of data points [example—clustering learning] (Celebi and Aydin, 2016; Ghahramani, 2004). Whereas, in semisupervised learning, the learning can be defined as labeling the unlabeled data points using knowledge learned from a few labeled data points [example—reinforcement learning] (Sinha, 2014; Xu et al., 2012).

3: Machine learning in precision medicine

Technological advancements aiding the development of new omics-based diagnostics and therapeutics have the potential of creating the unprecedented ability for detection, prevention, treatment planning, and monitoring of diseases. Cancer is one such heterogeneous disease having complex biological relationships, involving key molecules across the omics space, some of which are uncharacterized while some have unknown context-specific functions. In almost all cancer types, the available treatments are beneficial only for a patient sub-population, whereas for others, it may have adverse effects, or no improvements are observed. Considering the genetic variations at the granularity of an individual can thereof help in a better understanding of the disease and lead in the facilitation of better patient management. Thus, the need for this hour is the technological interventions that can guide for individualized prevention diagnostics and therapeutics leading to improved outcomes for all.

In the last 15–20 years, with the completion of the Human Genome Project and advancement in omics-based technologies, the data have been ever-increasing. Like, the cancer genome atlas program, a joint initiative by National Cancer Institute, USA and the National Human Genome Research Institute, USA, which commenced in 2006, and over a period has generated ~ 2.5 PB (PetaBytes) of genomic, epigenomic, transcriptomic, proteomic, and clinical data for different cancer types. For clinical advancements and research objectives, the data in this program have been made publicly available via the genomics data portal [https://portal.gdc.cancer.gov/] that handles genomic, epigenomic, transcriptomic, and clinical data (Akbani et al., 2014; Liu et al., 2018; Weinstein et al., 2013); and the cancer proteome atlas portal [https://www.tcpaportal.org/], which manages the protein expression datasets (Li et al., 2013). This poly-omics data is complex in nature, and along with clinical data provides an opportunity for exploring new insights via data-driven approaches (Filipp, 2019). Learning from this data can have multifold applications in diagnostics, prognostics, and as well as the discovery of new biomarkers or drug candidates. Spurred by the advancement in computer technologies, machine learning for precision medicine has therefore been a growing area of interest (Handelman et al., 2018; Holzinger, 2014). Supporting these objectives are initiatives like Project Data Sphere [https://data.projectdatasphere.org/projectdatasphere/html/home], which is promoting the development of new cancer therapies by providing an open access data sharing platform for sharing and analysis of patient-level data, giving access to more than 150 datasets (Green et al., 2015; Hede, 2013). It also offers an opportunity to collaborate in research programs using machine learning and big data analytics for oncology (Fig. 4).

Machine learning can be used for analyzing the omics (genome, epigenome, transcriptome, proteome, metabolome, and microbiome) data together with the clinical data and prior knowledge, inferring relationships or finding patterns or deciphering causal associations, giving insight into pleiotropy, complex interactions, and context-specific behavior. The multidimensional poly-omics datasets along with the clinical and other relevant data can be trained using machine learning algorithms to find the relevant genotypic structures which could be subsequently mapped to the observed phenotype. This model may then be used for diagnostic (predict risk, stage of disease), prognostic (chances of the patient treated successfully), and other outcomes for individual patients based on their characteristics (Fig. 5).

In comparison to the current process of treatment based on the symptoms-based classification (Fig. 1), this would provide clinicians the opportunity to tailor-made interventions for patients (Fig. 2).

Globally, various research groups and companies like Google and IBM are already exploring the opportunities of machine learning-centered advances for precision medicine. And, of late, there have been ground-breaking studies describing these approaches illustrating exemplary outcomes, for example, the prognosis of lung cancer patients based on 9879 histopathological and image-based features, distinguishing short-term from long-term survivors (Yu et al., 2016). In further sub-sections, we embellish applications discussing machine learning in precision medicine for the detection and diagnosis, prognosis, and the discovery of new biomarkers and drug candidates.

4: Future opportunities

Application of machine learning in precision medicine presents the distinctive challenges of having clinical interpretability of a model for its adoption under clinical settings. Computationally, this provides a unique research challenge of forming a balance in-between the performance and interpretability of the model as well as the results. For example, a deep learning model may effectively be able to identify a diseased from a healthy person, but clinically it will be imperative to know, based on what features or parameters, these predictions are been made and why.

Also, based on what is known about precision medicine, it is clear that for tailoring patient care, the domain needs to explore varied data dimensionalities. This depends on multiple factors including clinical (patient history, observed features, diagnostic measures, etc.), genetic (genome, proteome, metabolome, etc.), and, less clear nevertheless to be studied in-depth, wellness (behavior, emotional problems, stress, social support, etc.) along with the lifestyle and environmental factors (smoking, alcohol, drugs, malnutrition, etc.). These additional factors may assist in understanding, why some people even with a disease genotype do not develop a disease or why a sub-group shows a healthier prognosis. The role of a few factors like smoking (active and passive) are well-established with triggering, expression, and progression of cancer, cardiovascular, or other diseases. Nevertheless can emotions like “happiness or hope” be also playing a role, which can be associated with the genetic or clinical expression of a disease? Only time and further research in this domain may probably answer such questions but for sure it gives an opportunity for an integrative approach to explore and analyze these data dimensionalities (Fig. 6).

5: Conclusions

By the advancements made in computational power, theoretical understanding, and an ever-increasing amount of data, the last decade has witnessed widespread applications of machine learning in every major field of human society, including medicine and healthcare. In the 21st century, heterogeneous and complex diseases like cancer, cardiovascular, and rare genetic disorders are some major and pressing medical problems. The holistic approach offered by precision medicine will likely become increasingly important in order to provide effective treatment and management strategies for such diseases. Machine learning is currently a vital tool in precision medicine and is almost certain to remain so in the future. It can effectively handle the massive poly-omics datasets and aid to solve a number of analytical problems within precision medicine and in numerous ways better than historically conventional statistical methods. As discussed in this chapter, the state-of-the-art algorithms and analytical techniques offered by machine learning have shown a wide range of applications in precision medicine. Machine learning thus presently offers a diverse and effective tool set for precision medicine research; a tool set that will grow and improve in the future.