PaperlayerView exact revision on arXiv

Source-linked AI summary

Applications of Deep Learning and Reinforcement Learning to Biological Data

Mufti Mahmud, M. Shamim Kaiser, Amir Hussain, Stefano Vassanelli

arXiv:1711.03985v2cs.LGstat.ML

TL;DR

Biological data are multimodal, high-dimensional, diverse, and noisy, creating a need for robust analytical methods. This review surveys DL, RL, and Deep RL applications and compares DL performance across domains, finding varied dataset-specific outcomes with reported advantages in several comparisons.

  • Problem

    The complexity, dimensionality, diversity, and noise of biological data make extracting meaningful conclusions a major analytical challenge.

  • Method

    The paper surveys DL, RL, and Deep RL applications for mining biological data and compares DL performance across datasets and application domains.

  • Results

    Reported comparisons show DL methods often outperform alternatives across omics and medical-imaging tasks, although performance varies by dataset and method.

  • Takeaways & Limitations

    Method selection should consider the dataset, preprocessing, architecture, feature learning, and parameter optimization because these factors collectively determine accuracy.

Abstract

from arXiv · show

Rapid advances of hardware-based technologies during the past decades have opened up new possibilities for Life scientists to gather multimodal data in various application domains (e.g., Omics, Bioimaging, Medical Imaging, and [Brain/Body]-Machine Interfaces), thus generating novel opportunities for development of dedicated data intensive machine learning techniques. Overall, recent research in Deep learning (DL), Reinforcement learning (RL), and their combination (Deep RL) promise to revolutionize Artificial Intelligence. The growth in computational power accompanied by faster and increased data storage and declining computing costs have already allowed scientists in various fields to apply these techniques on datasets that were previously intractable for their size and complexity. This review article provides a comprehensive survey on the application of DL, RL, and Deep RL techniques in mining Biological data. In addition, we compare performances of DL techniques when applied to different datasets across various application domains. Finally, we outline open issues in this challenging research area and discuss future development perspectives.

Introduction

Advances in Life Sciences have enabled holistic, multimodal study of biological systems, but their complexity and high-dimensional, diverse, noisy data make meaningful inference difficult. This has driven demand for robust analytical instruments and multidisciplinary applications of artificial intelligence to biological data.

  • Motivation: Technological advances have expanded biological research by enabling holistic study of biological systems and generating complex, high-dimensional, diverse, and noisy data.These developments have created new opportunities while making inference from biological data more difficult.
  • Motivation: Meaningful conclusions from biological big data remain challenging, requiring analytical instruments that are robust, reliable, reusable, and accurate.The challenge has encouraged collaboration between life and computing scientists, contributing to progress in biological and biomedical research.
  • Machine learning paradigms: Conventional machine learning includes supervised methods that use known annotations, attributes, or features and unsupervised methods that identify similarities to form clusters.These paradigms support classification of objects and unknowns in biological data pools.
  • Machine learning paradigms: Reinforcement learning enables systems to learn from experience gained through interactions with their environments.RL is presented as a distinct learning category alongside supervised and unsupervised methods.
  • Applications: Machine-learning methods and their variants can be applied successfully to biological data from sources including Omics and Bioimaging.Omics encompasses genetics, transcriptomics, epigenomics, proteomics, and metabolomics data.

A F E

This review surveys DL, RL, and deep RL applications across biological data domains and addresses the lack of a comprehensive cross-domain performance comparison. It also identifies open issues and future perspectives.

  • DL, RL, and deep RL methods are reviewed for applications across diverse biological data domains.The domains include omics, bioimaging, medical imaging, and brain/body-machine interfaces.
  • The review fills a gap by comparing the performance of different DL approaches across datasets from various biological application domains.It combines conceptual coverage with state-of-the-art applications and performance comparison.
  • The article concludes by identifying open issues and discussing future development perspectives for applying these techniques to biological data.Its organization reserves a later section for open issues and future perspectives.

1 Conceptual Overview · 1.1 Deep Learning

Deep learning learns hierarchical data representations from raw inputs by building increasingly abstract features. The section surveys major architectures used with biological data, emphasizing their structures, data-processing strengths, and limitations.

  • 1.1 Deep Learning: Deep learning builds increasingly abstract representations from less abstract lower-level features, enabling systems to learn complex representations directly from raw data.This hierarchical learning process makes DL useful across many disciplines.
  • 1.1 Deep Learning: The surveyed DL architectures include DNNs, RNNs, CNNs, DAs, DBMs, DBNs, Deep Residual Networks, and Deep Convolutional Inverse Graphics Networks.The review focuses on architectures widely used with biological data and refers readers to cited sources for mathematical details.
  • 1.1.1 Deep Neural Network: DNNs extend standard neural networks with multiple hidden layers and nonlinear modules, inspired by multilevel visual processing, but their training process is slow.The supplied passage describes DNNs as learning portions of representations across hidden layers.
  • 1.1.2 Recurrent Neural Network: RNNs detect structures in data streams by making each current output depend on previous-state outputs, giving them a memory-like property.They are used in streaming-data domains including text mining, time series, and genomes, despite vanishing- and exploding-gradient problems.
  • 1.1.3 Convolutional Neural Network: CNNs combine convolutional and fully connected layers, often with subsampling, to handle multidimensional locally correlated inputs more effectively than DNNs.Convolutional filters learn data-driven kernels, while pooling reduces features before classification.
  • 1.1.4 Deep Autoencoder: Deep Autoencoders stack unsupervised Autoencoders that automatically project inputs into lower-dimensional representations.Autoencoders use equal-sized input and output layers with fewer hidden units for encoding.
  • 1.1.5 [Restricted] Boltzmann Machine ([R]BM): Restricted Boltzmann Machines are undirected probabilistic generative models and nonlinear feature detectors whose parameters are optimized to fit observed probability distributions.Their learning process uses Gibbs sampling, and their units have symmetrical connections with visible and hidden layers.
  • 1.1.5 [Restricted] Boltzmann Machine ([R]BM): DBMs stack undirected RBMs for bidirectional feedback and concurrent higher- and lower-level inference, whereas DBNs arrange RBMs into a hybrid directed–undirected generative model.DBM parameter estimation remains difficult, while DBN layers are learned greedily and then fine-tuned.

1.2 Reinforcement Learning

Reinforcement learning trains agents through trial-and-error interactions with an environment to maximize cumulative rewards. RL problems are commonly modeled as Markov Decision Processes, while function approximation helps address large state-action spaces.

  • Reinforcement Learning: RL agents learn through trial-and-error interactions, selecting actions that maximize cumulative rewards from the environment.Learning proceeds continuously through interactions at discrete time steps.
  • Reinforcement Learning: RL problems are generally modeled as Markov Decision Processes and solved using Monte Carlo or dynamic programming methods.The MDP framework specifies states, actions, policies, and reward functions.
  • Reinforcement Learning: At each time step, the agent observes state s_t, selects action a_t, and receives reward r_t+1 as the environment transitions to state s_t+1.The reward is associated with the transition (s_t, a_t, s_t+1).
  • Reinforcement Learning: State-value and action-value functions estimate expected discounted future rewards under a policy, with dynamic programming supporting policy or value iteration.The discount factor satisfies γ ∈ [0, 1].
  • Reinforcement Learning: Large real-world state-action spaces require function approximation, such as neural networks, to generalize value estimates to unknown states.Q-function approximation uses features and parameters updated through Monte Carlo or temporal-difference learning, with loss minimized using methods such as gradient descent.

1.3 Deep Reinforcement Learning

Deep reinforcement learning combines reinforcement learning with deep neural networks to address limitations involving high-dimensional data and complex pattern learning. DQN and Double DQN exemplify this integration through deep-network value approximation and reduced Q-learning overestimation.

  • RL learns without feature crafting but struggles with large data dimensionality and non-stationary environments.
  • Deep reinforcement learning combines RL algorithms with deep neural networks, while DL can remain prone to misclassification.
  • DQN combines Q-learning with deep neural networks to learn policies from high-dimensional inputs, approximating action values with deep CNNs.Experience replay and a target network help overcome instability and divergence encountered with shallow neural networks.
  • Double DQN extends DQN by replacing Q-learning with double Q-learning to overcome substantial overestimations.It learns two value functions with separate weight sets; one determines the greedy policy while the other determines its value during each update.
  • Other deep RL algorithms include Deep Deterministic Policy Gradient, Continuous DQN, Asynchronous N-step Q-learning, and Dueling network DQN.

2 Applications to Biological Data

Deep learning and reinforcement learning are applied across omics, biological imaging, and medical imaging to extract biological information, analyze images, and support prediction tasks. Omics applications span sequence analysis, molecular structure and interaction prediction, while imaging applications include segmentation, classification, denoising, and disease detection.

  • Omics: DL methods mine omics data to extract features, functions, structures, and molecular dynamics from DNA, RNA, and amino-acid sequences.Applications include gene-expression profiling, splicing-junction prediction, sequence-specificity prediction, and transcription-factor determination.
  • Omics: Omics studies use DBNs, DNNs, CNNs, autoencoders, and RNNs for gene-expression profiling, RNA-binding analysis, variant annotation, gene selection, and miRNA prediction.These methods address binding sites, pathogenicity, noncoding variants, DNA methylation, and miRNA targets.
  • Omics: Proteomics applications address protein-structure prediction and compound-protein interaction prediction, including decomposition of complete structure estimation into parallel sub-problems.The complete protein-structure problem is described as complex and NP hard, motivating secondary-structure and interaction-focused models.
  • Omics: RL applications in omics include bacterial-genome prediction, improved biological-sequence annotation, DNA-fragment assembly, and protein-protein interaction-network construction.The reported approaches combine RL with binary particle swarm optimization or text mining in some tasks.
  • Biological imaging: DL analyzes biological images at pixel, cell, and tissue levels for segmentation, classification, protein localization detection, and bacterial-colony counting.Examples include nucleus, cytoplasm, cell, and nuclear-membrane segmentation; mitosis identification; label-free cell classification; and yeast microscopy analysis.
  • Medical imaging: DL and RL architectures analyze MRI, CT, PET, radiography, fundus, microscope, and ultrasound images to denoise, segment, classify, and detect anomalies or diseases.Reported medical-image applications include tumor, traumatic-injury, and ischemic-stroke lesion segmentation, image denoising, and Alzheimer’s disease or mild cognitive impairment detection.

3 Performance Analysis and Comparison

Deep learning methods generally outperformed non-DL approaches across biological sequence, cellular imaging, medical image segmentation and denoising, and disease detection tasks. Performance advantages were dataset- and task-dependent, with template-based protein prediction and some mammography settings favoring alternative methods.

  • Biological sequence and molecular prediction: DBN achieved higher splice-junction F1-scores than SVM, Gene Splicer, and Splice Machine across Whole Human Genome and UCSC genome datasets.On GWH-donor and GWH-acceptor, DBN reached 0.81 and 0.75 versus SVM-Radial Basis at 0.77 and 0.67.
  • Bioimaging: Deep learning improved cellular and histopathology image classification, including DNN cellular-compartment accuracy of 87% versus 75% for RF and DNN-MIL colon classification accuracy of 97.44%.Deep CNN also classified flow-cytometry cell-cycle phases at 98.73±0.16% versus boosting at 93.1±0.5%.
  • Medical image analysis: Deep learning methods outperformed comparison methods in medical-image segmentation, including InputCascadeCNN Dice coefficient 0.88 versus Tustison 0.87 and StAE DSC 0.79 versus SVM 0.73.A multi-scale dual-pathway 3D CNN with Conditional Random Fields also exceeded RF, with DSC values of 63.0 ± 16.3 versus 54.
  • Medical image analysis: DL-based denoising methods surpassed alternatives for radiographs and brain CT, with StDAE-CNN reaching SSIM scores of 0.89 and 0.9 under Gaussian and Poisson noise.Adaptive MC-StSDA also exceeded MC-StSDA in CT denoising for Gaussian, salt-and-pepper, and speckle noise using the reported SSIM4 scores.
  • Medical image analysis: CNN methods often led mammography and Alzheimer’s classification benchmarks, although DWT-GMB outperformed SVM-MLP on MIAS and performance varied across ADNI tasks.DSA-3D-CNN achieved 99.3% for AD and 94.2% for MCI, while CNN-LSVM reached 96.7% on DDSM mammograms.

4 Open Issues and Future Perspectives

The section identifies computational, theoretical, privacy, security, and real-time-processing limitations of DL and RL for biological data. It highlights future directions including stronger theoretical foundations, deep-RL methods for multimodal biological data, unsupervised learning, multitasking, multiagent learning, and tailored distributed infrastructures.

  • Limitations: DL and RL require substantial computing power and memory, limiting their suitability for moderate-size datasets.The passage also notes that cloud, distributed, and parallel computing raise ongoing data-privacy and security concerns.
  • Limitations: Real-time processing of the enormous volumes of experimentally acquired biological data remains underdeveloped.
  • Future Perspectives: DL theory should be strengthened using experimental data to quantify individual neural-network models’ computational complexity and learning efficiency under well-defined parameter-tuning strategies.
  • Future Perspectives: Deep RL could derive dynamic information from biological data across multiple levels, reducing redundancy and enabling discovery of novel biomarkers for disease detection and prevention.
  • Future Perspectives: Unsupervised deep-RL learning methods are needed to reduce dependence on large labeled training datasets, while multitasking and multiagent paradigms should advance further.
  • Future Perspectives: Distributed and parallel computing infrastructures tailored to biological applications are needed to keep pace with rapid growth in biological data.

5 Conclusion

Technological advances have enabled Life Sciences to collect complex multimodal and multidimensional biological data, creating a major mining challenge. Interdisciplinary work has responded by developing cutting-edge machine-learning analytical tools, building on advances in artificial neural networks, deep architectures, and reinforcement learning.

  • 5 Conclusion: Technological advancement has created a major challenge in mining multimodal, multidimensional, and complex biological data.The passage identifies this challenge as a consequence of recent advances in Life Sciences.
  • 5 Conclusion: Interdisciplinary approaches have produced cutting-edge machine-learning-based analytical tools to address biological-data mining.These approaches were triggered by the challenge of analyzing complex biological data.
  • 5 Conclusion: Artificial neural networks, deep architectures, and reinforcement learning have established success stories in making machines intelligent.The passage presents these methods as the established foundations motivating continued machine-learning applications.
Loading 1711.03985v2…