Artifical Intelligence in Precision Medicine

From Psyc 40 Wiki
Jump to: navigation, search

Precision medicine, also called personalized medicine, aims to provide individualized treatments for patients with regards to their genes, environments, and lifestyles difference. [1] Artifical intelligence, a branch of computing disciplines, concerns with building computing programs that are able to perform tasks on a level compatible with human intelligence. [2] Artifical intelligence in precision medicine aims to support and improve decision-based medical task performances through real world data mining and hence data-intensive computational solutions. [3]


Artifical Intelligence (AL)

Artifical Intelligence (AL), dated back to 1940s, is originally a field of computer science that targets at developing computational algorithms with advanced analytical or predictive capabilities. [4]

Machine Learning (ML)

Machine Learning (ML), a branch of artificial intelligence, intends to improve the predictive capability and hence the learning capacity of those computation algorithms. [5] Tom Mitchell phases this process in a simplistic manner, “A computer program 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.” Other machine learning techniques include Natural Language Processing (NLP), Computer Vision (CV), Reinforcement Learning (RL) and etc.[4]

Deep Learning (DL)

Deep Learning (DL), a branch of machine learning, utilizes artificial neural networks, which consist of multiple data processing layers, for automated feature extraction and pattern recognition from extremely large datasets.[6]

Convolutional Neural Network (CNN)

Convolutional Neural Network (CNN), an important subsidiary branch of deep learning, exploits multi-layers of nonlinear information processing for feature extraction and transformation, as well as pattern analysis and classification. [6] More specifically, convolutional neural network is widely used in image recognition and classification. Image classification seeks to categorize images into predefined classes, which forms the basis of other computer vision tasks. One major advance took place in the ImageNet Large Scale Visual Recognition Challenge (ILSVRC). The winning entry utilized Deep Convolutional Neural Network (DCNN) to classify approximately 1.2 million images into 1000 classes.

What is Traditional Medicine?

In conventional clinical management, doctors treat patients with a particular illness or disease as a homogenous entity. They make recommendations and treatment plans based on an expected outcome of an average patient who suffers from this illness or disease.

What is Precision Medicine?

According to NIH’s definition, precision medicine is an innovative approach that takes into account individual differences in patients’ genes, environments, and lifestyles.[1] There are four prominent fields within precision medicine: precision oncology, cancer immunotherapy, pharmacogenomics and rare diseases. Precision oncology seeks medical treatment based on the DNA signature of an individual patient’s tumor. [1] Cancer immunotherapy aims to enlist an individual’s own immune system to control his or her cancer cells. [1]For example, CAR T cell immunotherapy utilizes highly specific antibodies or receptors for tumor cells, which are expressed on T cells in a purified form from patient’s blood extract, to re-infuse into the patient. Pharmacogenomics seeks to match DNA patterns in individuals with how they respond to medications, with an ultimate goal of providing the right drug at the right dose at the right time for the right patient. [1]Rare diseases, as the name implies, are diseases which are individually rare and not widely researched upon, especially in terms of medical treatments. However, about 25 to 30 million Americans are collectively affected. Precision medicine provides the hope for them.[1]

Artificial Intelligence Application in Precision Medicine

Precision Oncology

Precision oncology seeks medical treatment based on the DNA signature of an individual patient’s tumor. AIinPrecisionOncology.png [4]

Precision Oncology Using Deep Learning

Detection and Classification of Skin Lesions

Skinlesion.png Skin cancer was primarily diagnosed visually. The first step is usually an initial clinical screening. The following steps might include a dermoscopic analysis, a biopsy and histopathological examination.[7] However, the fine-grained variability in the appearance of skin lesions prevents an automated classification of skin lesions from images. [7] Thus, one study uses the GoogleNet Inception v3 CNN architecture pre-trained on 2014 ImageNet Large Scale Visual Recognition Challenge dataset and trained on a set of dermatologist-labelled images organized in a tree-structured taxonomy of 2032 diseases. [7] In the deep CNN, there are 757 disease training classes with each individual diseases form the leaf nodes of the tree-structured taxonomy. [7]Then, researchers created two critical binary classification tasks: keratinocyte carcinomas versus benign seborrheic keratoses, and malignant melanomas versus benign nevi.[7] The first classification aims to identify common skin cancer type and the second classification aims to identify the worst skin cancer type. The CNN would output a probability distribution over those inference classes, which vary by task. [7]The results are validated using a nine-fold (meaning classes) cross validation and biopsy-proven images comparison. The study is able to conclude that the CNN outperforms average dermatologist at skin cancer classification using photographic and dermoscopic images.[7] A single CNN trained on general skin lesion classification achieves performance on par with professional dermatologists on both identifying common skin cancer types and the worst skin cancer types. The research further proposes that mobile device equipped with CNN would provide large-scale low-cost access to dermatology diagnosis outside hospital.[7]

Identification and Categorization of Lung Cancers

Lung cancer was primarily diagnosed through visual inspection of histopathology slides. [8] Non-Small Lung Cancer (NSLC) is one of the most prevalent cancer types in America. Adenocarcinoma (LUAD) and squamous cell carcinoma (LUSC) are the most prevalent non-small lung cancer subtypes. One study uses the GoogleNet Inception v3 CNN architecture pre-trained on 2014 ImageNet Large Scale Visual Recognition Challenge dataset and trained on datasets obtained from The Cancer Genome Atlas (TCGA). [8]It aims to accurately identify and automatically classify LUAD and LUSC (binary classifications). [8]TCGA’s datasets includes whole-slide images with an automatic classification of ‘solid tissue normal’ and ‘primary tumor.’ [8]There are 459 sets of ‘solid tissue normal’ images and 1175 sets of ‘primary tumor’ images. [8]Within the ‘primary tumor’ sets of images, there are 567 sets of LUAD images and 608 sets of LUSC images. [8]TCGA’s datasets serve as the bench line reference for the test set in the study, which has 170 slide images with correct labels.[8] Those slides were then tiled into a non-overlapping 512*512 pixel windows at a twenty-fold magnification with non-essential information removed. [8]Approximately 1 million tiles are utilized: 70% for training, 15% for validation and 15% for final testing. [8]The research utilizes transfer learning technique, which refines the parameters of the last layer via back propagation. Moreover, it also utilizes cross entropy loss and RMSProp optimization, which not only improves the weights of fully connected layers but also improves the parameters of previous layers and convolution filters of all layers. [8] The training runs for 500,000 iterations and is able to generate relevant classification from the testing set. The final focus of the research is to predict gene mutational status, which helps to identify highly specific gene target therapy available in market. Six out of ten common mutated genes are predictable with the CNN. [8]

Precision Oncology Using Other Machine Learning Techniques

Random Forest: Prediction of Oncogenes and Tumor Suppressors

Cancer immunotherapy

Cancer immunotherapy Using Deep Learning

Prediction of Breast Cancer Proteins Involved in Immunotherapy

Breast cancer is an extremely common and complex cancer type. It is a leading cause cancer-related deaths among women in America. It involves genomic alterations, protein expression deregulation, signal pathway alterations, hormone disruption, ethnicity and environmental determinants. [9] One study uses 13 machine learning methods to choose the best prediction model among all. The research shows that multilayer perceptron, namely Artifical Neural Network (ANN), shows the best Area Under the Receiver Operating Characteristics (AUROC) of 0.980. It can hence be used to classify breast cancer proteins using 300 descriptors. [9] Those invariant descriptors are based on physical and chemical properties of the amino acids, 3D protein conformation, graph topology and protein sequence. [9] The classification links signaling activity (biological function) to protein structure. A total of 4504 sequences of proteins have been screened for breast cancer relation. [9] It identifies some of the best ranked cancer immunotherapy proteins related to breast cancer: RPS27, SUPT4H1, CLPSL2, POLP2K, RPL38, AKT3, CDK3, RPS20, RASL11A and UBTD1 etc. The specific scripts for ANN model can be downloaded at the following link: [9]


Pharmacogenomics seeks gene variants which are correlated with drug effects in populations, cohorts and individual patients. It is essentially an interdisciplinary subject combing genomics and pharmacology with clinical practicality. [10] There are promising applications of pharmacogenomics in drug discovery, medication optimization based on patients’ genotypes and clinical diagnostics.


Pharmacogenomics Using Deep Learning

Pharmacoepigenomic datasets usually contain information on gene regulatory elements like promoters, enhancers and disruption of transcription factor binding sites etc. Current research focuses are on noncoding regulatory genome analysis.[10] Deep learning models, in forms of ANNs, CNNs, and RNNs etc, have shown promises in identifying novel regulatory variants on noncoding domains which might impact pharmacogenomic response. [10] For example, as the Tox21 Data Challenge, researchers use a deep learning model (CNNs) called DeepTox to predict drug compounds’ oral toxicity based on ToxAlerts database. Other open-source deep learning software applications in pharmacogenomics include DeepDTI, DeepChem, DeepSynergy and etc. [10]

Rare Disease

Rare Disease basic and clinical research are severely underrepresented in the science community. There are currently 7000 rare diseases identified worldwide but only 5% have some form of treatment options.

Policy Support

FDA is considering a total product lifecycle-based regulatory framework to allow real world data platform to advance precision medicine. [11] Currently, the regulatory focus is on medical device using adaptive artifical intelligence. For detailed information, FDA has published a discussion paper on the topic on April2, 2019 which entails its foundational approach. The discussion paper is titled as "Proposed Regulatory Framework for Modifications to Artificial Intelligence/Machine Learning (AI/ML)-Based Software as a Medical Device (SaMD) - Discussion Paper and Request for Feedback." [11]

Commercial Implementation

Application Drawbacks


  1. 1.0 1.1 1.2 1.3 1.4 1.5 The Promise of Precision Medicine. (2020, February 12). from
  2. Artificial intelligence. (n.d.). Retrieved November 18, 2020, from
  3. Artificial Intelligence in Medicine. (n.d.). from
  4. 4.0 4.1 4.2 Azuaje, F. Artificial intelligence for precision oncology: beyond patient stratification. npj Precision Onc 3, 6 (2019).
  5. Rawat, W., & Wang, Z. (2017). Deep Convolutional Neural Networks for Image Classification: A Comprehensive Review. Neural Computation, 29(9), 2352-2449. doi:10.1162/neco_a_00990
  6. 6.0 6.1 LeCun, Y., Bengio, Y. & Hinton, G. Deep learning. Nature 521, 436–444 (2015).
  7. 7.0 7.1 7.2 7.3 7.4 7.5 7.6 7.7 Esteva, A., Kuprel, B., Novoa, R. et al. Dermatologist-level classification of skin cancer with deep neural networks. Nature 542, 115–118 (2017).
  8. 8.00 8.01 8.02 8.03 8.04 8.05 8.06 8.07 8.08 8.09 8.10 Coudray, N., Ocampo, P.S., Sakellaropoulos, T. et al. Classification and mutation prediction from non–small cell lung cancer histopathology images using deep learning. Nat Med 24, 1559–1567 (2018).
  9. 9.0 9.1 9.2 9.3 9.4 López-Cortés, A., Cabrera-Andrade, A., Vázquez-Naya, J.M. et al. Prediction of breast cancer proteins involved in immunotherapy, metastasis, and RNA-binding using molecular descriptors and artificial neural networks. Sci Rep 10, 8515 (2020).
  10. 10.0 10.1 10.2 10.3 Kalinin, Alexandr A et al. “Deep learning in pharmacogenomics: from gene regulation to patient stratification.” Pharmacogenomics vol. 19,7 (2018): 629-650. doi:10.2217/pgs-2018-0008
  11. 11.0 11.1 Center for Devices and Radiological Health. (n.d.). Artificial Intelligence and Machine Learning in Software. from