Artificial intelligence (AI) is the talk of the town nowadays, but it needs to be utilized for a better life, and that’s where the application of intelligence is important. Genomics is a field where large amounts of data are generated and evaluating and identifying biomarkers and their application for improved healthcare is important. In specific, when we talk about incorporating genomics into healthcare, it is important to establish credible intelligent platforms using data generated in laboratories. These intelligent platforms, derived from large scale genome studies with complicated data ranging from DNA sequence to patient information, help to develop computational models.
The promise of AI is to provide a set of tools to augment and extend the effectiveness of a physician. With more data and information from whole-genome sequencing and biometrics, there are expectations from patients about getting faster and personalized care. Physicians have always been in need of identifying and interpreting relationships between different variables for improving patient care. Using AI-based platforms makes this happen. AI and machine learning use several methodologies which allow algorithmic learning and efficient representation of the data. The difference between the usual statistical methods used by a physician and machine learning is that, with the former, we can conduct inferences about a given set of samples or population, but with the latter, we can predict or develop the classification approaches. However, these two approaches are intertwined and are used together to develop applied intelligence.
DL approaches with genomics
Deep learning is a subfield of machine learning, where large data-driven rules are automatically used to improve diagnosis or therapy. This process of machine learning happens by feature extraction from raw data to use these features to develop a learning algorithm to detect patterns in new inputs. In contrast, deep learning develops its own pattern recognition approaches using multiple hidden layers of sequentially arranged, primitive, non-linear representations or features. As data flows through these layers, the system iteratively identifies patterns and associations. Deep learning models are usually large-scale datasets due to their ability to run on specialized computing hardware and continuous improvements. Convolutional neural networks (CNNs) are a type of deep learning which can process data with spatial invariance, like images. Image-based diagnostics has successfully implemented CNN-based methods. Genomics is the field where deep-learning approaches are adapted beyond conventional approaches like NLP (Natural Language Processing), CNN and RL (Reinforcement Learning). Genomic technologies can output a wide variety of data, from an individual’s DNA sequence to the quantity of functional proteins in different tissues in different medical conditions. These kinds of results help clinicians provide more accurate diagnostics and treatments.
A usual deep-learning genomics pipeline starts with raw data (a DNA sequence, gene expression etc..), converts this raw data into input data tensors, which in turn are fed through neural networks to power biomedical needs. One of the major opportunities are large-scale genome-wide association (GWA) studies which aim to discover causal or associated genetic markers. Applying such algorithms to GWA studies in large patient cohorts can help in dealing with latent confounders. Understanding genetic modifications can allow the clinicians to recommend treatments based on associated factors. One major step ahead for deep learning is integrating different external data sources into GWA studies like medical imaging or other phenotypes, resulting in improved identification of disease-associated causal mutations. However, for a better understanding and easier utilization of genomic intelligence, data visualization tools are important.
Intuitive visualization tools are needed in order to effectively interpret the data and extract new biological knowledge and insights integrating previous biological and clinical knowledge. Therefore, it is important to access large repositories of often structured biological knowledge and facilitate interaction among them. This specific integrative visualization analytics is called panomics for personalized treatment of diseases. For example, cancer tumours which seem similar can be completely different at the molecular level, leading to different types of outcomes or treatment responses. These molecular features are being increasingly used for stratification of patients and can result in more accurate treatment and diagnosis. One such tool extensively used is UCSC Cancer Genomics Browser. It allows scientists and clinicians to identify genomic signatures in cancer subtypes and stratify them to have qualitatively differential treatments. There are more such visualization tools which are being developed, such as iGPSe (Interactive Genomics Patient Stratification explorer), which significantly reduce the computation burden for biomedical fraternity. This system uses unsupervised clustering with graph and parallel sets visualization for direct comparison of clinical outcomes via survival analysis. These kinds of tools and deep learning are a revolution in healthcare which could translate into great benefit to clinicians.