thesis about convolutional neural network

International Conference on Theory and Application of Soft Computing, Computing with Words and Perceptions

ICSCCW 2021: 11th International Conference on Theory and Application of Soft Computing, Computing with Words and Perceptions and Artificial Intelligence - ICSCCW-2021 pp 116–123 Cite as

Malaria Detection Using Convolutional Neural Network

• Khaled Almezhghwi   ORCID: orcid.org/0000-0001-5755-7297 15  
• Conference paper
• First Online: 04 January 2022

379 Accesses

2 Citations

Part of the book series: Lecture Notes in Networks and Systems ((LNNS,volume 362))

Malaria detection and classification is still time and money costly. Identification of malaria cells can be done through some costly techniques. Those techniques are good, but they require time and high cost. Hence, there is a need of discovering alternative techniques to identify blood cells, that saves both time and reduce cost. In addition to time and cost, those new techniques should also be accurate and effective. Thus, in this work, we propose a transfer learning based GoogleNet approach for the classification of Malaria cells. The depth and inception of GoogleNet made it a very robust deep network that can classify accurately if trained and fine-tuned on enough amount of data. Thus, in this study, 27558 of the 2 types of cells are used for fine-tuning and testing the pre-trained network GoogleNet. Experimentally, the employed GoogleNet fine-tuned to classify Malaria, showed a great capability in generalizing accurate and correct diagnosis of images that were not seen during training, in which it achieved a testing accuracy of 95% with a relatively short time and small number of epochs.

This is a preview of subscription content, log in via an institution .

Buying options

• Available as PDF
• Read on any device
• Instant download
• Own it forever
• Available as EPUB and PDF
• Compact, lightweight edition
• Dispatched in 3 to 5 business days
• Free shipping worldwide - see info

Tax calculation will be finalised at checkout

Purchases are for personal use only

Razzak, M.I.: Malarial parasite classification using recurrent neural network. Int. J. Image Process 9 (2), 27–32 (2015)

Google Scholar  

World Health Organization. Disease burden of malaria. http://www.who.int/mediacentre/factsheets/fs094/en/

Almezhghwi, K., Serte, S., Al-Turjman, F.: Convolutional neural networks for the classification of chest X-rays in the IoT era. Multim. Tool. Appl. 80 (19), 29051–29065 (2021). https://doi.org/10.1007/s11042-021-10907-y

Kaymak, S., Almezhghwi, K., Shelag, A.A.S.: Classification of diseases on chest X-rays using deep learning. In: Aliev, R.A., Kacprzyk, J., Pedrycz, W., Jamshidi, M.O., Sadikoglu, F.M. (eds.) ICAFS 2018. AISC, vol. 896, pp. 516–523. Springer, Cham (2019). https://doi.org/10.1007/978-3-030-04164-9_69

Abiyev, R.H., Maaitah, M.K.S.: Deep convolutional neural networks for chest diseases detection. J. Healthc. Eng. 2018 , 4168538 (2018). https://doi.org/10.1155/2018/4168538

Nguyen, D., et al.: A comparison of Monte Carlo dropout and bootstrap aggregation on the performance and uncertainty estimation in radiation therapy dose prediction with deep learning neural networks. Phys. Med. Biol. 66 (5), 054002 (2021)

Malaria Cell Images Dataset. https://www.kaggle.com/iarunava/cell-images-for-detecting-malaria

Szegedy, C., et al.: Going deeper with convolutions. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR), p. 12 (2015). https://doi.org/10.1109/CVPR.2015.7298594

Download references

Author information

Authors and affiliations.

Electrical and Electronics Engineering, College of Electronics Technology, Tripoli, Libya

Khaled Almezhghwi

You can also search for this author in PubMed   Google Scholar

Editor information

Editors and affiliations.

Department of Control Systems, Azerbaijan State Oil and Industry University, Baku, Azerbaijan

Rafik A. Aliev

System Research Institute, Polish Academy of Sciences, Warsaw, Poland

Janusz Kacprzyk

Department of Electrical and Computer Engineering, University of Alberta, Edmonton, AB, Canada

Witold Pedrycz

Department of Electrical and Computer Engineering, University of Texas at San Antonio, San Antonio, TX, USA

Mo Jamshidi

Azerbaijan State Oil and Industry University, Baku, Azerbaijan

Mustafa Babanli

Department of Mechatronics, Near East University, North Cyprus, Turkey

Fahreddin M. Sadikoglu

Rights and permissions

Reprints and permissions

Copyright information

© 2022 The Author(s), under exclusive license to Springer Nature Switzerland AG

About this paper

Cite this paper.

Almezhghwi, K. (2022). Malaria Detection Using Convolutional Neural Network. In: Aliev, R.A., Kacprzyk, J., Pedrycz, W., Jamshidi, M., Babanli, M., Sadikoglu, F.M. (eds) 11th International Conference on Theory and Application of Soft Computing, Computing with Words and Perceptions and Artificial Intelligence - ICSCCW-2021. ICSCCW 2021. Lecture Notes in Networks and Systems, vol 362. Springer, Cham. https://doi.org/10.1007/978-3-030-92127-9_19

Download citation

DOI : https://doi.org/10.1007/978-3-030-92127-9_19

Published : 04 January 2022

Publisher Name : Springer, Cham

Print ISBN : 978-3-030-92126-2

Online ISBN : 978-3-030-92127-9

eBook Packages : Intelligent Technologies and Robotics Intelligent Technologies and Robotics (R0)

Share this paper

Anyone you share the following link with will be able to read this content:

Sorry, a shareable link is not currently available for this article.

Provided by the Springer Nature SharedIt content-sharing initiative

• Publish with us

Policies and ethics

• Find a journal
• Track your research

• Bibliography
• More Referencing guides Blog Automated transliteration Relevant bibliographies by topics
• Automated transliteration
• Relevant bibliographies by topics
• Referencing guides

Convolutional Neural Network for COVID-19 Detection in Chest X-Rays

Joshua Elliot Henderson Follow

Date of Award

Spring 2022

Document Type

Department/major.

Computer Science

First Advisor

KC Santosh, PhD

Second Advisor

Douglas R Goodman, PhD

Third Advisor

Arun Singh, PhD

Convolutional Neural Network, Binary Image Classification, COVID-19, Chest X-rays

Subject Categories

Artificial Intelligence and Robotics

The COVID-19 pandemic has had a large effect on almost every facet of life. As COVID-19 was a disease only discovered in recent history, there is comparatively little data on the disease, how we detect it, and how we cure it. Deep learning is a powerful tool that can be used to learn to classify information in ways that humans might not be able to. This allows computers to learn on relatively little data and provide exceptional results. In this paper, I propose a novel convolutional neural network (CNN) for the detection of COVID-19 from chest X-rays called basicConv. This network consists of five sets of convolution and pooling layers, a flatten layer, and two dense layers with a total of approximately 9 million parameters. This network achieves an accuracy of 95.8 percent, which is comparable to other high-performing image classification networks. This provides a promising launching point for future research and developing a network that achieves an accuracy higher than that of the leading classification networks. It also demonstrates the incredible power of convolution.

Recommended Citation

Henderson, Joshua Elliot, "Convolutional Neural Network for COVID-19 Detection in Chest X-Rays" (2022). Honors Thesis . 254. https://red.library.usd.edu/honors-thesis/254

Since April 29, 2022

Included in

Artificial Intelligence and Robotics Commons

To view the content in your browser, please download Adobe Reader or, alternately, you may Download the file to your hard drive.

NOTE: The latest versions of Adobe Reader do not support viewing PDF files within Firefox on Mac OS and if you are using a modern (Intel) Mac, there is no official plugin for viewing PDF files within the browser window.

Advanced Search

• Notify me via email or RSS
• Collections
• Disciplines

Author Corner

• Copyright Guidelines
• Scholarly Communication
• Getting Started
• Submit Research
• Honors Program
• Thesis Submission Guide
• University Libraries

Home | About | FAQ | My Account | Accessibility Statement

Privacy Copyright

Help | Advanced Search

Computer Science > Machine Learning

Title: on the efficiency of convolutional neural networks.

Abstract: Since the breakthrough performance of AlexNet in 2012, convolutional neural networks (convnets) have grown into extremely powerful vision models. Deep learning researchers have used convnets to produce accurate results that were unachievable a decade ago. Yet computer scientists make computational efficiency their primary objective. Accuracy with exorbitant cost is not acceptable; an algorithm must also minimize its computational requirements. Confronted with the daunting computation that convnets use, deep learning researchers also became interested in efficiency. Researchers applied tremendous effort to find the convnet architectures that have the greatest efficiency. However, skepticism grew among researchers and engineers alike about the relevance of arithmetic complexity. Contrary to the prevailing view that latency and arithmetic complexity are irreconcilable, a simple formula relates both through computational efficiency. This insight enabled us to co-optimize the separate factors that determine latency. We observed that the degenerate conv2d layers that produce the best accuracy-complexity trade-off also have low operational intensity. Therefore, kernels that implement these layers use significant memory resources. We solved this optimization problem with block-fusion kernels that implement all layers of a residual block, thereby creating temporal locality, avoiding communication, and reducing workspace size. Our ConvFirst model with block-fusion kernels ran approximately four times as fast as the ConvNeXt baseline with PyTorch Inductor, at equal accuracy on the ImageNet-1K classification task. Our unified approach to convnet efficiency envisions a new era of models and kernels that achieve greater accuracy at lower cost.

Submission history

Access paper:.

• Other Formats

References & Citations

• Google Scholar
• Semantic Scholar

BibTeX formatted citation

Bibliographic and Citation Tools

Code, data and media associated with this article, recommenders and search tools.

• Institution

arXivLabs: experimental projects with community collaborators

arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.

Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.

Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs .

Image Classification using Quantum Convolutional Neural Network

Ieee account.

• Change Username/Password
• Update Address

Purchase Details

• Payment Options
• Order History
• View Purchased Documents

Profile Information

• Communications Preferences
• Profession and Education
• Technical Interests
• US & Canada: +1 800 678 4333
• Worldwide: +1 732 981 0060
• Contact & Support
• About IEEE Xplore
• Accessibility
• Terms of Use
• Nondiscrimination Policy
• Privacy & Opting Out of Cookies

A not-for-profit organization, IEEE is the world's largest technical professional organization dedicated to advancing technology for the benefit of humanity. © Copyright 2024 IEEE - All rights reserved. Use of this web site signifies your agreement to the terms and conditions.

• Search Menu
• Advance articles
• Author Guidelines
• Submission Site
• Open Access
• Why publish with this journal?
• About Bioinformatics
• Journals Career Network
• Editorial Board
• Advertising and Corporate Services
• Self-Archiving Policy
• Dispatch Dates
• Journals on Oxford Academic
• Books on Oxford Academic

Article Contents

1 introduction, 2 materials and methods, 3 results and discussions, 4 conclusions, acknowledgements, supplementary data, conflict of interest, data availability.

• < Previous

Contrastive pre-training and 3D convolution neural network for RNA and small molecule binding affinity prediction

• Article contents
• Figures & tables
• Supplementary Data

Saisai Sun, Lin Gao, Contrastive pre-training and 3D convolution neural network for RNA and small molecule binding affinity prediction, Bioinformatics , Volume 40, Issue 4, April 2024, btae155, https://doi.org/10.1093/bioinformatics/btae155

• Permissions Icon Permissions

The diverse structures and functions inherent in RNAs present a wealth of potential drug targets. Some small molecules are anticipated to serve as leading compounds, providing guidance for the development of novel RNA-targeted therapeutics. Consequently, the determination of RNA–small molecule binding affinity is a critical undertaking in the landscape of RNA-targeted drug discovery and development. Nevertheless, to date, only one computational method for RNA–small molecule binding affinity prediction has been proposed. The prediction of RNA–small molecule binding affinity remains a significant challenge. The development of a computational model is deemed essential to effectively extract relevant features and predict RNA–small molecule binding affinity accurately.

In this study, we introduced RLaffinity, a novel deep learning model designed for the prediction of RNA–small molecule binding affinity based on 3D structures. RLaffinity integrated information from RNA pockets and small molecules, utilizing a 3D convolutional neural network (3D-CNN) coupled with a contrastive learning-based self-supervised pre-training model. To the best of our knowledge, RLaffinity was the first deep learning based method for the prediction of RNA–small molecule binding affinity. Our experimental results exhibited RLaffinity’s superior performance compared to baseline methods, revealed by all metrics. The efficacy of RLaffinity underscores the capability of 3D-CNN to accurately extract both global pocket information and local neighbor nucleotide information within RNAs. Notably, the integration of a self-supervised pre-training model significantly enhanced predictive performance. Ultimately, RLaffinity was also proved as a potential tool for RNA-targeted drugs virtual screening.

https://github.com/SaisaiSun/RLaffinity

RNA molecules, traditionally known for their roles in information transfer and protein synthesis, have emerged as key players in gene regulation, cellular signaling, and disease pathways. The functionality of RNA molecules is commonly realized through intricate interactions with various cellular components, encompassing proteins, peptides, DNAs, other RNAs, and small molecules. Notably, RNA–small molecule interactions play a crucial role in various biological processes and have garnered significant research interests in recent years ( Morgan et al. 2017 ). Moreover, as our understanding of RNA structures and functions continues to accumulate, it becomes evident that RNAs are intricately linked to conditions such as cancer and viral diseases, functioning as mediators in neurological diseases, and emerging as potential targets for therapeutic interventions ( Bernat and Disney 2015 ). The study of RNA–small molecule interactions can provide valuable insights into the mechanisms underlying RNA-mediated processes. On one hand, these interactions can occur through multiple modes, including hydrogen bonding, electrostatic interactions, hydrophobic interactions, and pi-stacking interactions ( Szulc et al. 2022 ). On the other hand, small molecules can comprise a diverse range of compounds, including metabolites, drugs, and chemical probes. These molecules can bind to specific regions of RNA, such as hairpins, loops, or bulges, inducing conformational changes that can impact RNA folding, stability, and activity ( Costales et al. 2020 ). They can also target functional RNA elements, such as ribozymes, riboswitches, or non-coding RNA domains, to regulate gene expression or modulate cellular processes. These multifaceted interactions allow for the design of small molecules with the capacity to selectively bind to RNA, enabling the inhibition or activation of specific RNA functions, offering a unique approach for precision medicine ( Aboul-Ela 2010 ).

A plenty of studies have been conducted to identify RNA–small molecule interactions experimentally. An illustrative example is the work by Serganov et al. , where they presented the crystal structures of the Fusobacterium nucleatum riboswitch bound to riboflavin and the antibiotic roseoflavin ( Serganov et al. 2009 ). And some studies have elucidated the roles of microRNA and long non-coding RNA in various human diseases ( Tang et al. 2021 , Gao et al. 2022 , Fan et al. 2023 ). Simultaneously, some recent investigations have introduced small molecules and oligonucleotides into clinical trials for cancer treatment, with a focus on modulating the mRNA spliceosome ( Lee et al. 2016 , Effenberger et al. 2017 ). Subsequently, RNA has emerged as a spivotal category of potential therapeutic targets, as evidenced by comprehensive reviews ( Thomas and Hergenrother 2008 , Aboul-Ela 2010 ). Notably, to date, the landscape of FDA-approved drugs interacting with RNAs remains relatively limited, such as: risdiplam, patisiran, fomivirsen, inotersen, mipomersen, and eteplirsen ( Zogg et al. 2022 ).

The analysis of RNA–ligand interactions is significantly enhanced by the existence of high-resolution structures of RNA–ligand complexes. Nevertheless, the experimental determination of structures for RNA and its complexes poses multiple challenges and currently cannot be achieved in a high-throughput manner. Consequently, the need to overcome these challenges has driven the development of computer software aimed at modeling RNA–ligand complex structures, based on the available structures of RNA receptors. Several of these advancements have been inspired by methods previously devised for modeling protein–ligand complexes ( Bottegoni 2011 ). While recent years have witnessed significant strides in the development of automatic docking tools for predicting protein–ligand interactions ( Gohlke and Klebe 2002 , Sousa et al. 2006 ), the modeling of RNA–ligand interactions has seen comparatively less progress ( Pfeffer and Gohlke 2007 ). Only a few appropriate docking and scoring methods for RNA–ligand interactions have been developed, such as rDOCK ( Morley and Afshar 2004 , Ruiz-Carmona et al. 2014 ), AutoDock ( Trott and Olson 2010 , Goodsell et al. 2021 ), MORDOR ( Guilbert and James 2008 ), Dock6 ( Lang et al. 2009 ), as well as NLDock ( Feng et al. 2021 ). Correspondingly, the scoring functions for docking include LigandRNA ( Philips et al. 2013 ), SPA-LN ( Yan and Wang 2017 ), ITScore-NL ( Feng and Huang 2020 ), and DrugScoreRNA ( Pfeffer and Gohlke 2007 ). Additionally, a few databases focusing on RNA–ligand interactions have been established, such as SMMRNA ( Mehta et al. 2014 ), HARIBOSS ( Panei et al. 2022 ), RNALID ( Fan et al. 2023 ), and RPocket ( Zhou et al. 2021 ). Besides, several RNA–ligand binding sites prediction algorithms have been proposed, such as MetalionRNA ( Philips et al. 2012 ), Rsite ( Zeng et al. 2015 ), RBind ( Wang et al. 2018 ), RNAsite ( Su et al. 2021 ), RNALigand ( Sun et al., 2022 ) and RLBind ( Wang et al. 2023 ).

Despite the usefulness of these databases and methods, to date there has not been very active research in cataloging and predicting RNA and small molecule ligand interaction strength or affinity. To the best of our knowledge, only one computational approach predicting RNA–ligand binding strength or affinity has been developed. The method was named RSAPred, which was developed based on traditional machine learning ( Krishnan et al. 2024 ). Comprehensive understanding of binding affinities to substrates, inhibitors, and cofactors is essential for unraveling the intermolecular interactions that drive biological processes, structural biology, and the intricate connections between structure and function in proteins, nucleic acids, and other biomolecules. However, several challenges exist in predicting RNA–small molecule binding affinities, such as the inherent flexibility and conformational dynamics of both RNA and small molecules impeding accurately capturing their interactions. Additionally, the limited availability of high-quality experimental data on RNA–small molecule binding affinities hinders the development of robust and generalizable prediction models.

In this study, we aim to address these challenges and contribute to the field of RNA–small molecule binding affinity prediction. We have developed a reliable and accurate prediction model named RLaffinity, by combing a contrastive learning-based self-supervised pre-training model and a 3D-CNN model. Firstly, our method was compared with some baseline methods and demonstrated superior performances. Additionally, we did a blind test on RNA–small molecule pairs, revealing consistent results with other literatures. Furthermore, our method was proved utility in drug discovery for ranking compound hits binding to the target, which demonstrated that RLaffinity can aid in the development of medicines that bind their targets selectively and precisely and enhance our understanding of RNA functions.

2.1 RNA–ligand complexes

In our study, we employed the standardized data, specifically consisting of the structures of RNA and ligands in their bound state (i.e. complex state), archived in individual Protein Data Bank (.pdb) format files ( Berman 2008 , wwPDB consortium 2019 ). The strength of ligand binding to RNA was defined as binding affinity, which was determined for each complex structure through experimental techniques such as spectroscopic shift assays and isothermal titration calorimetry (Ballester and Mitchell). The experimental binding affinity data (dissociation constants k d ⁠ ) were used as the ground truth labels for the supervised training and testing procedure. And following the works in protein–ligand binding affinity prediction ( Jiménez et al. 2018 , Jones et al. 2021 ), the dissociation constants k d   were normalized by the negative log function [i.e. - log ⁡ ( k d ) ]. In addition, for the effectiveness input of the neural networks, all labels were normalized to (0, 1) using the min-max normalization method.

In the initial phase, RNA–small molecule pairs lacking binding affinity data were downloaded and filtered from Protein Data Bank (PDB) for the self-supervised pre-training process. This phase involved exclusive training with RNA–ligand structures, devoid of affinity labels. Firstly, all structures containing both RNA and small molecules were downloaded from PDB, obtaining 1429 RNA–small molecule complex structures. Subsequently, structures containing proteins or RNAs with lengths less than 10 nucleotides were selectively removed. Following this refinement, all retained structures exclusively featured ions as their ligands, and any structures containing only artifact ligands were excluded. Ultimately, 1415 complex structures with interatomic distances between RNA and ligand smaller than 4 Å were preserved for subsequent pre-training procedure.

In the second phase, RNA–small molecule pairs with binding affinity labels were obtained from the PDBbind database for the supervised training and testing processes ( Wang et al. 2004 , 2005 ). Specifically, 149 RNA–ligand pairs were retrieved from PDBbind NL2020 set, including 144 compounds, one DNA and four peptides as the ligands. The distribution of these RNA lengths was shown in Supplementary Fig. S3 . Subsequently, 144 RNA–ligand pairs were remained after removing the DNA ligand and peptide ligands. After that, 100, 29, and 13 pairs were randomly selected for training, validation, and testing, respectively. Given the limited dataset, this procedure was iterated 10 times to generate 10 distinct training sets, 10 distinct validation sets, and 10 corresponding test sets. During the training procedure, the optimal parameters were obtained according to the performances of the validation sets. And then the final 10 models with the optimal parameters were tested on the 10 corresponding test sets. Notably, no redundancy elimination procedure was applied, considering the absence of substantial similarity between RNA–ligand pairs. Additionally, to augment the original data, rotational and translational transformations were incorporated into the training procedure to expand the labeled dataset.

2.2 Binding pocket, features, box size, and grid resolution

To obtain accurate three-dimensional representations for the deep learning models, a standardized preprocessing procedure was implemented on the binding complex structures deposited in Protein Data Bank (.pdb) format.

Firstly, all RNA–ligand binding complexes were charged and protonated by UCSF Chimera ( Couch et al. 2006 ) with AMBER bsc1 ( Oweida et al. 2021 ), with the default settings of the program. Following this, for each RNA–ligand complex, KD-Trees algorithm was utilized to detect of the binding pocket, with a designated query radius of 10 Å (distance from the ligand atoms). The KD-Tree algorithm is a data structure designed for organizing points in k -dimensional space in a hierarchical manner. This structure greatly simplifies and enhances efficiency in tasks such as closest-neighbor searches, which involve identifying the nearest neighbors of a given point within the space ( Bentley 1975 ).

Then, a standardized atomic representation was employed as input for the 3D-structure-based deep learning models. Specifically, only the heavy atoms from each biological structure were taken into consideration. This involved employing a one-hot encoding scheme for eight major elements, including C, N, O, P, F, Cl, Br, and I. The cheminformatic tool OpenBabel (version 2.4.1) ( O’Boyle et al. 2008 ) was employed to extract features for all binding complexes. Subsequently, to generate the spatial representation of the binding complex, all atomic coordinates were centered by each ligand.

Furthermore, to determine the voxel box size, the distribution of end-to-end distances for all ligands in our datasets were calculated, as shown in Supplementary Figs S1 and S2 . According to the sum of the end-to-end distance for the longest ligand in these two datasets and the query radius, the voxel box size was ultimately defined as 41 Å, which is sufficient to cover the entire pocket region while minimizing the collisions between atoms. If the dimensions of the box be set smaller than the length of a ligand, there exists the possibility that the terminal sides of the ligand to extend beyond the space of the box, resulting in data loss about molecular components. Even in cases where the initial orientation permits the ligand to fit within the box, subsequent data augmentation procedures are likely to generate input structures with consequential data loss.

The Van der Waals radius of the eight major heavy atoms (C, N, O, P, F, Cl, Br, and I) used in our study are greater than 1.4 Å. This radius is considered a measure of an atom’s size, defined as half of the internuclear separation of two non-bonded atoms of the same element at their closest possible approach. A grid resolution larger than two times of the radius would be insufficient to differentiate two atoms from each other. Conversely, opting for a finer resolution would result in a significantly higher computational cost. To ensure an optimal balance between accuracy and efficiency, we set the 3D voxel grid resolution at 1.0 Å. Hence, each atom can be allocated to at least one voxel grid, depending on its Van der Waals radius. Subsequently, after the voxelization of all atoms, a Gaussian blur with σ = 1 was employed to diffuse the atom features into neighboring voxels. This strategy was implemented to prevent an excessively sparse representation within the 3D voxel grid, similar to that employed in the previous work ( Kuzminykh et al. 2018 ).

2.3 Model architecture

To sufficiently utilize the unlabeled RNA–ligand structures data, an unsupervised pre-training model was designed for the RNA–ligand interaction information and pattern extraction. Subsequently, a supervised regression model was employed and fine-tuned to predict RNA–ligand binding affinities. For the unsupervised pre-training, a contrastive learning model was utilized in this study, demonstrating its ability to generate efficient 3D representations for the subsequent supervised regression model. Furthermore, the prediction of binding affinity was performed through a three-dimensional convolutional neural network model. An illustration of the overall model architecture is presented in Fig. 1 .

The flowchart of RLaffinity. Firstly, RNA-ligand complex structures without labels were represented as cubes with a dimension of 41 × 41 × 41 × 8 and input into a self-supervised pre-training model, here referring to a contrastive learning model, for RNA–ligand interaction information extraction. After the pre-training model, RNA–ligand complex structures with labels were input into the pre-trained model to generate structural representations ( ⁠ h i ⁠ ), which then were input to the 3DCNN-based regression models for binding affinity prediction.

2.4 Three-dimensional convolutional neural networks

3D-CNNs have been extensively applied to diverse computer vision fields, such as gestures or shapes recognition in videos and 3D image segmentation, due to its capability of shift invariance and representation learning. Lately, 3D-CNNs have been substantiated for their efficacy in predicting binding affinity and elucidating protein–ligand interactions within the domain of drug discovery ( Jiménez et al. 2017 , 2018 , Ragoza et al. 2017 , Kuzminykh et al. 2018 , Jones et al. 2021 ). The inherent capability of the 3D-CNN model lies in its ability to capture intricate three-dimensional atomic features and implicit atomic interactions. This is achieved through the use of 3D volume representations, wherein atoms and their respective features are voxelized into a three-dimensional voxel grid. The 3D atomic representation dimension was N  ×  N × N  ×  C , where N denoted the voxel box size in each axis, and C represented the number of atomic features described in the preceding section (eight in our study).

To build the supervised regression model, a straightforward 3D-CNN architecture was implemented. This model consisted of only four convolutional layers, complemented by four dropout layers to avoid overfitting, as visually represented in Fig. 1 (top). To enhance the model’s nonlinear capacity, a rectified linear unit (RELU) layer was integrated within each convolutional and fully connected layer. Additionally, to ensure translation invariance, MaxPooling layer was applied to the second and fourth convolutional layers. The adaptive moment estimation (Adam) optimizer, with a learning rate set to 0.0001, was employed for optimization. The batch size, denoting the number of samples processed per batch, was configured as 16. The overall parameters of the 3D-CNN model can be found in Supplementary Table S1 , including details such as filter size, kernel size and stride.

2.5 Contrastive learning framework

Incorporating insights from recent advances in computer vision ( Wang et al. 2021 , Xie et al. 2021 , Zhao et al. 2021 , Dave et al. 2022 , Denize et al. 2023 ), our framework was trained with the objective of maximizing the similarity between representations derived from augmented structures of the same pocket–ligand complex while minimizing the similarity between augmented structures from different pocket–ligand complexes, referring to Fig. 1 (bottom). Generally, for a given pocket–ligand complex cube, we performed rotations and translations of the complex within the cube to obtain two augmented structures, S i and S j ⁠ . Subsequently, we computed the latent representations of these augmented structures, h i and h j ⁠ , utilizing a 3D ResNet-based encoder, h i = R ( S i ) ⁠ . Following the previous approach in protein representation ( Xia et al. 2022 ), these latent representations were further projected into smaller latent representations, z i and z j ⁠ , using a multilayer perceptron with a single hidden layer, z i = P ( h i ) ⁠ . Ultimately, the similarity between these representations was quantified using the cosine distance, denoted as s ( z i , z j ) and the associated training loss was determined through the computation of the NT-Xent loss ( Denize et al. 2023 ). The detailed parameters involving in the contrastive learning model can be found in Supplementary Table S2 .

Specifically, the 3D residual neural network (ResNet18) model served as the encoder within the contrastive learning framework, comprising a total of eight residual blocks ( Alaeddine and Jihene 2021 ). The inclusion of the residual short connection facilitates the transmission of gradients to subsequent layers without engaging nonlinear activation. Each residual block was structured with three convolution layers, three batch normalization layers, and three RELU activation layers. The final output dimension of this module was 23 × 23 × 23 × 32.

2.6 Evaluation metrics

3.1 comparison with other baseline methods.

To evaluate the robustness of our method, we trained 10 models and conducted testing on 10 distinct sets randomly sampled from the entire dataset. Table 1 presents the mean metrics values across the 10 test sets for different methods, including Vina ( Trott and Olson 2010 ), RF-score ( Ballester and Mitchell 2010 ), RSAPred ( Krishnan et al. 2024 ), our 3D-CNN model, and our pre-trained 3D-CNN model (RLaffinity). Evidently, the 3D-CNN models, both in its conventional form and when augmented with pre-training, exhibit heightened predictive capabilities (PCCs: 0.466 and 0.559) compared to Vina, RF-score, and RSAPred (PCCs: −0.386, 0.445, and 0.399). The increased correlation coefficients (PCCs/SPCCs) and reduced RMSE values associated with 3D-CNN indicate its efficacy in providing accurate predictions for binding affinities of RNA and small molecules. Furthermore, the self-supervised pre-training before the 3D-CNN model demonstrates potential enhancements, as indicated by the increased correlation coefficients (PCCs/SPCCs). This underscores the viability of leveraging pre-existing structural information or representations to enhance predictive performance.

The results from different methods on the benchmark test set. a

The value of each metric is the mean value of metrics on the 10 test sets. For Vina, we took the docking score as the predicted affinity. And the method of 3DCNN represents only using the 3DCNN model for prediction. For the RSAPred method, the labels were not normalized leading to RMSE and MAE values exceeding 1. The highest value of each metric from different methods is highlighted in bold type.

To further explore the details of the results, we analyzed the metric distributions across different methods and parameters on 10 test sets. Figure 2A–C present violin plots illustrating the distributions, scopes, and median values of PCC, SPCC, and RMSE for various methods. Notably, our pre-trained 3D-CNN model (RLaffinity) demonstrates the narrowest deviation scope for each metric and the highest median values of correlation coefficients (PCC: 0.540, SPCC: 0.500), compared to 3D-CNN model (PCC: 0.460, SPCC: 0.360), and RF-score (PCC: 0.480, SPCC: 0.320). This indicates that our pre-training model provided more stable and accurate predictions of RNA–ligand binding affinity. Additionally, the 3D-CNN method exhibits a more robust result in PCC distribution and slightly higher SPCC values compared to the RF-score method, which are consistent with the results showed in Table 1 . As for the RMSE distributions, RF-score method has a higher maximum value than the other two methods. And most RMSE values focus on similar ranges for RLaffinity method and RF-score method (0.120–0.180), with 0.155/0.144 as the medium value. In summary, these observations suggest that 3D convolutional neural networks can effectively extract valuable information about binding strength from complex structures for the prediction of binding affinities between RNA and small molecules.

The metrics distributions of different methods (A–C) and different parameters (D–F). Here, RLaffinity, 3DCNN and RF-score represent our method with pre-training, our method without pre-training and the method based on random forest, respectively. And the numbers in horizontal axis of figure (D–F) represent different combinations of box size and pocket size.

Furthermore, for a more direct comparison of methods, we did a one-to-one comparison of the pre-trained 3D-CNN method with the original 3D-CNN method and the RF-score method. Figure 3 presents the scatter plots of the AE values of different methods on 10 test sets. In Fig. 3A , the majority of dots concentrate in the upper region, indicating that most samples exhibit higher AE values generated by 3D-CNN method compared to the pre-trained 3D-CNN method. This highlights the substantial improvement in performance achieved through the pre-training procedure. However, Fig. 3B indicates a comparable number of dots falling in both the upper and lower regions, including some notable dots with large AE values. For these specific samples, further analysis and improvement are warranted.

One-to-one AE comparison of different methods on 10 test sets. Here, the horizontal axis represents pre-trained 3DCNN method (RLaffinity) and the vertical axis represents 3DCNN method without pre-training (A) and the RF-score method (B).

3.2 Pocket size and box size selection

In the investigation of protein–small molecule interactions, a conventional radius of 6 Å has been widely employed to extract binding pockets. In this study, we conducted a systematic examination of different pocket sizes (6 and 10 Å) to identify the most suitable parameter for predicting RNA–small molecule binding affinity. Furthermore, considering that 3D-CNNs process four-dimensional input data ( N × N × N × C , where N denotes the box size and C represents the channels), variations in box sizes can impact feature extraction through 3D-CNN. In consideration of the ligand lengths distribution, we opted for box sizes of 31 and 41 Å, chosen to encompass the entirety of both pocket and ligand structures. The comparative results based on diverse pocket sizes and box sizes are presented in Table 2 . Notably, the highest PCC of 0.559 was observed when using a pocket size of 10 Å and a box size of 41 Å. Additionally, distinct parameters yielded similar RMSEs around 0.15.

The results from different parameters on the benchmark test set. a

The value of each metric is the mean value of metrics on the 10 test sets.  The highest value of each metric from different parameters is highlighted in bold type.

For a more comprehensive understanding of the results obtained with different box sizes and pocket sizes, we delved into the metrics distributions across all test sets with various parameter combinations. The outcomes are visually depicted in Fig. 2D–F through violin plots. In Fig. 2 , a pocket size of 10 Å exhibits a narrower range across both correlation coefficients (PCCs and SPCCs) compared to a pocket size of 6 Å. This observation suggests that a pocket size of 6 Å might inadvertently exclude some crucial atoms involved in essential chemical bonds, potentially leading to less stable predictions. Regarding box sizes, a size of 41 Å, in contrast to a size of 31 Å, achieved a higher median value of both PCCs (0.534, 0.483) and SPCCs (0.500, 0.360) with the pocket size of 10 Å. These results suggest that larger box sizes (41 Å) coupled with smaller pocket size (6 Å) resulted in broader ranges of PCCs/SPCCs might be due to sparser embeddings within the grid box. Conversely, box size of 31 Å coupled with pocket size of 10 Å may result in information omitting. In addition, all distinct size parameters demonstrated similar performances in RMSE distributions, with a range around 0.08–0.20 and a medium value of 0.154.

To further assess the impact of different box and pocket size combinations, we conducted one-to-one comparisons on all metrics across various test sets. Figure 4 illustrates scatter diagrams depicting PCCs, SPCCs, and RMSEs for different box size and pocket size combinations. From Fig. 4A and B , it is evident that the combination of a box size of 41 Å and a pocket size of 10 Å outperforms other combinations in terms of PCC and SPCC metrics. This suggests that a box size of 41 Å and a pocket size of 10 Å exhibit superior performance across most test sets. Additionally, Fig. 4C indicates that the combination of a box size of 41 Å and a pocket size of 10 Å demonstrates similar RMSE values when compared to other combinations on each test set.

Metrics comparisons (PCC, SPCC, and RMSE) of different combinations of box size and pocket size on 10 test sets.

3.3 Blind test on structure unknown complexes

To further validate the effectiveness of our method, we conducted a blind test on three RNA–small molecule pairs from literatures ( Yan et al. 2018 , Tran et al. 2020 ). Firstly, the RNA and small molecule structures were docked by Autodock Vina to generate the complex structures with default parameters ( Trott and Olson 2010 ). Subsequently, leveraging the known structures as input, we predicted the binding pockets and binding affinities by our method. Figure 5 exhibits the docking results of the three RNA–ligand pairs, which include a “G-riboswitch-guanine analog” complex ( Fig. 5A ) and two “Fusibacterium ulcerans ZTP riboswitch-5-aminoimidazole-4-carboxamide (AICA) analog” complexes ( Fig. 5B and C ). They are reported involving in bacterial growth and infection related diseases ( Tran et al. 2020 ). The guanine derivatives were deciphered structure–activity relationship with the guanine riboswitches and inhibitory effect on bacterial growth through in-line probing experiments ( Tran et al. 2020 ). And the AICA derivatives were found to bind and activate ZTP riboswitches in vitro through transcription termination assays ( Yan et al. 2018 ). As depicted in Fig. 5 , the “G-riboswitch-guanine analog” complex, with an assay-measured k d of 6.5 µM, exhibited a medium-strength affinity. Our method predicted its binding affinity with a k d of 5.5 µM, indicating a slight deviation. Specifically, the head of its ligand was positioned outside its pocket, resulting in limited interactions between the tail of the ligand and the receptor. In the case of the “Fusibacterium ulcerans ZTP riboswitch-5-aminoimidazole-4-carboxamide (AICA) analog” complexes, their binding affinities were measured by assay with k d ∼5.7 and ∼2.9 µM. And their prediction values were 5.4 and 3.8 µM, respectively, demonstrating the capability of our approach to provide effective information for RNA drug selections. To be more specific, the AICA analog ligands were entirely enveloped by their pockets, facilitating well-proportioned and high-density interactions between the ligands and their respective binding sites. Moreover, AICA analog five, in comparison to AICA analog 13, features a greater number of hydroxide radicals. This characteristic contributes to the formation of more hydrogen bonds with the receptor, resulting in a slightly higher binding strength. Altogether, these findings could serve as guidance for the future design of enhanced riboswitch activators and offer valuable insights into the potential trajectory of RNA-targeted ligand discovery.

Blind tests on three RNA–small molecule pairs. Thereinto, the 3D interaction structures were generated by Vina. (A) is a “G-riboswitch-guanine analogue” complex, (B) and (C) are two “Fusibacterium ulcerans ZTP riboswitch-5-aminoimidazole-4-carboxamide (AICA) analog” complexes.

3.4 Potential compounds screening

Furthermore, we conducted a virtual screening of small molecules targeting the entire structural landscape of the transactivation response element (TAR) from the human immunodeficiency type 1 virus (HIV-1) ( Stelzer et al. 2011 ). We performed quantitative predictions of binding affinities for small molecules that interact with HIV-1 TAR and reported the top five ranking compounds that bind TAR with near record affinities. Figure 6 provides comprehensive binding information of potential compounds [5-( N , N )-dimethyl amiloride, netilmicin, amikacin, sisomicin, mitoxantrone] with the target (HIV-1 TAR). In Fig. 6A , 2D chemical structures of the five small molecule hits are depicted along with the 3D interaction structures of each compound and its corresponding pocket within the HIV-1 TAR (generated from Vina). Figure 6A indicates that each RNA–ligand complex formed a distinct binding pocket wrapping the corresponding compound with a specific binding mode.

Compounds selection against HIV-1 TAR target. (A) 2D structures of the five compound hits targeting HIV-1 TAR and their 3D interaction structures visualized by Pymol (slate: pocket, yellow: interaction sites, magenta: ligand, black: hydrogen bonds). (B) Ranking results of the five compounds from the experimental assay (left) and from our method (right).

Additionally, Fig. 6B exhibits the binding affinities ( ⁠ - log ⁡ ( k d ) ⁠ ) of the five compounds and their order ranked by the fluorescence intensity assay and RLaffinity. Among the tested compounds, 5-( N , N )-dimethyl amiloride displayed the highest k d value (121.85 µM), indicating the weakest binding strength with HIV-1 TAR. Netilmicin, Amikacin, and Sisomicin showed similar k d values at a moderate level (1.35, 1.54, and 0.73 µM). Mitoxantrone exhibited the lowest k d value of 0.076 µM, designating it as the top-ranked compound for targeting HIV-1 TAR RNA. Importantly, our method yielded a comparable ranking order for the five compounds based on the predicted binding affinities. This outcome underscores the utility of our method in facilitating virtual screening efforts effectively.

In conclusion, our comparative analysis provided significant insights into the strengths and limitations of diverse prediction methods of RNA and small molecule binding affinities. These findings could serve as a valuable contribution to the continual refinement of computational approaches in the domain of molecular binding predictions. It is crucial to contextualize these observations within the specific attributes and requirements of the investigated RNA–small molecule interactions. Subsequent research endeavors and validation studies are warranted to affirm these conclusions and potentially enhance the predictive models.

Notably, this study pioneers a 3D convolutional neural network (3D-CNN) model coupled with a contrastive learning-based self-supervised pre-training model for the precise prediction of RNA–small molecule binding affinities, presented in the RLaffinity framework. Leveraging structural information from both RNA pockets and small molecules, RLaffinity outperforms established baseline methods in the realm of binding affinity prediction. Combing with the appropriate box size and pocket size, the 3D-CNN model could extract global pocket information and local neighbor nucleotide information within RNAs. Additionally, the incorporation of a self-supervised pre-training model emerges as a key contributor to the heightened predictive performance, underscoring the efficacy of this approach. The versatility of RLaffinity is further highlighted by its potential application as a tool for virtual screening of RNA-targeted drugs. This research represents a significant stride forward in the computational prediction of RNA–small molecule interactions, offering valuable insights and paving the way for enhanced approaches to RNA-targeted drug discovery and development.

We proposed the first computational method named RLaffinity to predict the RNA–small molecule binding affinity.

RLaffinity was constructed through a fusion model, including a contrastive learning-based pre-training model and a 3D-CNN regression model.

Experimental results demonstrated that RLaffinity significantly outperformed other baseline methods and showed its potential use in RNA-targeted drug virtual screening.

We are grateful to Dr Pengyong Li for helping about the usage of the Autodock Vina package.

Supplementary data are available at Bioinformatics online.

None declared.

Young Scientists Fund of the National Natural Science Foundation of China [62302357]; the Natural Science Basic Research Program of Shaanxi Province [S2023-JC-QN-0727]; and the Fundamental Research Funds for the Central Universities [ZYTS24088].

The codes and datasets are available online at https://github.com/SaisaiSun/RLaffinity .

Aboul-Ela F. Strategies for the design of RNA-binding small molecules . Future Med Chem 2010 ; 2 : 93 – 119 .

Google Scholar

Alaeddine H , Jihene M. Deep residual network in network . Comput Intell Neurosci 2021 ; 2021 : 6659083 .

Ballester PJ , Mitchell JB. A machine learning approach to predicting protein-ligand binding affinity with applications to molecular docking . Bioinformatics 2010 ; 26 : 1169 – 75 .

Bentley JL. Multidimensional binary search trees used for associative searching . Commun ACM 1975 ; 18 : 509 – 17 .

Berman HM. The protein data bank: a historical perspective . Acta Crystallogr A 2008 ; 64 : 88 – 95 .

Bernat V , Disney MD. RNA structures as mediators of neurological diseases and as drug targets . Neuron 2015 ; 87 : 28 – 46 .

Bottegoni G. Protein-ligand docking . Front Biosci (Landmark Ed) 2011 ; 16 : 2289 – 306 .

Costales MG , Childs-Disney JL , Haniff HS et al.  How we think about targeting RNA with small molecules . J Med Chem 2020 ; 63 : 8880 – 900 .

Couch GS , Hendrix DK , Ferrin TE. Nucleic acid visualization with UCSF chimera . Nucleic Acids Res 2006 ; 34 : e29 .

Dave I , Gupta R , Rizve MN et al.  TCLR: Temporal contrastive learning for video representation . Computer Vision and Image Understanding 2022 ; 219 : 103406 .

Denize J , Rabarisoa J , Orcesi A et al.  Similarity contrastive estimation for self-supervised soft contrastive learning . IEEE Wint Conf Appl 2023 ; 2705 – 15 .

Effenberger KA , Urabe VK , Jurica MS. Modulating splicing with small molecular inhibitors of the spliceosome . Wiley Interdiscip Rev RNA 2017 ; 8 : e1381.

Fan C , Wang X , Ling T et al.  Characterizing RNA-binding ligands on structures, chemical information, binding affinity and drug-likeness . RNA Biol 2023 ; 20 : 431 – 43 .

Feng Y , Huang SY. ITScore-NL: an iterative knowledge-based scoring function for nucleic acid-ligand interactions . J Chem Inf Model 2020 ; 60 : 6698 – 708 .

Feng Y , Zhang K , Wu Q et al.  NLDock: a fast nucleic acid-ligand docking algorithm for modeling RNA/DNA-ligand complexes . J Chem Inf Model 2021 ; 61 : 4771 – 82 .

Gao Y , Mi N , Zhang Y et al.  Uterine macrophages as treatment targets for therapy of premature rupture of membranes by modified ADSC-EVs through a circRNA/miRNA/NF-kappaB pathway . J Nanobiotechnology 2022 ; 20 : 487 .

Gohlke H , Klebe G. Approaches to the description and prediction of the binding affinity of small-molecule ligands to macromolecular receptors . Angew Chem Int Ed 2002 ; 41 : 2644 – 76 .

Goodsell DS , Sanner MF , Olson AJ et al.  The AutoDock suite at 30 . Protein Sci 2021 ; 30 : 31 – 43 .

Guilbert C , James TL. Docking to RNA via root-mean-square-deviation-driven energy minimization with flexible ligands and flexible targets . J Chem Inf Model 2008 ; 48 : 1257 – 68 .

Jiménez J , Doerr S , Martínez-Rosell G et al.  DeepSite: protein-binding site predictor using 3D-convolutional neural networks . Bioinformatics 2017 ; 33 : 3036 – 42 .

Jiménez J , Škalič M , Martínez-Rosell G et al.  K(DEEP): protein-ligand absolute binding affinity prediction via 3D-convolutional neural networks . J Chem Inf Model 2018 ; 58 : 287 – 96 .

Jones D , Kim H , Zhang X et al.  Improved protein-ligand binding affinity prediction with structure-based deep fusion inference . J Chem Inf Model 2021 ; 61 : 1583 – 92 .

Krishnan SR , Roy A , Gromiha MM. Reliable method for predicting the binding affinity of RNA-small molecule interactions using machine learning . Brief Bioinform 2024 ; 25 :bbae002.

Kuzminykh D , Polykovskiy D , Kadurin A et al.  3D molecular representations based on the wave transform for convolutional neural networks . Mol Pharm 2018 ; 15 : 4378 – 85 .

Lang PT , Brozell SR , Mukherjee S et al.  DOCK 6: combining techniques to model RNA-small molecule complexes . RNA 2009 ; 15 : 1219 – 30 .

Lee SC-W , Dvinge H , Kim E et al.  Modulation of splicing catalysis for therapeutic targeting of leukemia with mutations in genes encoding spliceosomal proteins . Nat Med 2016 ; 22 : 672 – 8 .

Mehta A , Sonam S , Gouri I et al.  SMMRNA: a database of small molecule modulators of RNA . Nucleic Acids Res 2014 ; 42 : D132 – 41 .

Morgan BS , Forte JE , Culver RN et al.  Discovery of key physicochemical, structural, and spatial properties of RNA-targeted bioactive ligands . Angew Chem Int Ed Engl 2017 ; 56 : 13498 – 502 .

Morley SD , Afshar M. Validation of an empirical RNA-ligand scoring function for fast flexible docking using Ribodock . J Comput Aided Mol Des 2004 ; 18 : 189 – 208 .

O’Boyle NM , Morley C , Hutchison GR. Pybel: a Python wrapper for the OpenBabel cheminformatics toolkit . Chem Cent J 2008 ; 2 : 5 .

Oweida TJ , Kim HS , Donald JM et al.  Assessment of AMBER force fields for simulations of ssDNA . J Chem Theory Comput 2021 ; 17 : 1208 – 17 .

Panei FP , Torchet R , Ménager H et al.  HARIBOSS: a curated database of RNA-small molecules structures to aid rational drug design . Bioinformatics 2022 ; 38 : 4185 – 93 .

Pfeffer P , Gohlke H. DrugScoreRNA—knowledge-based scoring function to predict RNA-ligand interactions . J Chem Inf Model 2007 ; 47 : 1868 – 76 .

Philips A , Milanowska K , Lach G et al.  MetalionRNA: computational predictor of metal-binding sites in RNA structures . Bioinformatics 2012 ; 28 : 198 – 205 .

Philips A , Milanowska K , Lach G et al.  LigandRNA: computational predictor of RNA-ligand interactions . RNA 2013 ; 19 : 1605 – 16 .

Ragoza M , Hochuli J , Idrobo E et al.  Protein-ligand scoring with convolutional neural networks . J Chem Inf Model 2017 ; 57 : 942 – 57 .

Ruiz-Carmona S , Alvarez-Garcia D , Foloppe N et al.  rDock: a fast, versatile and open source program for docking ligands to proteins and nucleic acids . PLoS Comput Biol 2014 ; 10 : e1003571 .

Serganov A , Huang L , Patel DJ. Coenzyme recognition and gene regulation by a flavin mononucleotide riboswitch . Nature 2009 ; 458 : 233 – 7 .

Sousa SF , Fernandes PA , Ramos MJ. Protein-ligand docking: current status and future challenges . Proteins 2006 ; 65 : 15 – 26 .

Stelzer AC , Frank AT , Kratz JD et al.  Discovery of selective bioactive small molecules by targeting an RNA dynamic ensemble . Nat Chem Biol 2011 ; 7 : 553 – 9 .

Su H , Peng Z , Yang J. Recognition of small molecule-RNA binding sites using RNA sequence and structure . Bioinformatics 2021 ; 37 : 36 – 42 .

Sun S , , Yang J , , Zhang Z. RNALigands: A database and web server for rna–ligand interactions . RNA 2022 ; 28 : 115 – 22 .

Szulc NA , Mackiewicz Z , Bujnicki JM et al.  fingeRNAt-a novel tool for high-throughput analysis of nucleic acid-ligand interactions . PLoS Comput Biol 2022 ; 18 : e1009783 .

Tang J , Zhang C , Huang Y et al.  CircRNA circ_0124554 blocked the ubiquitination of AKT promoting the skip lymphovascular invasion on hepatic metastasis in colorectal cancer . Cell Death Dis 2021 ; 12 : 270 .

Thomas JR , Hergenrother PJ. Targeting RNA with small molecules . Chem Rev 2008 ; 108 : 1171 – 224 .

Tran B , Pichling P , Tenney L et al.  Parallel discovery strategies provide a basis for riboswitch ligand design . Cell Chem Biol 2020 ; 27 : 1241 – 9.e4 .

Trott O , Olson AJ. AutoDock Vina: improving the speed and accuracy of docking with a new scoring function, efficient optimization, and multithreading . J Comput Chem 2010 ; 31 : 455 – 61 .

Wang R , Fang X , Lu Y et al.  The PDBbind database: collection of binding affinities for protein-ligand complexes with known three-dimensional structures . J Med Chem 2004 ; 47 : 2977 – 80 .

Wang R , Fang X , Lu Y et al.  The PDBbind database: methodologies and updates . J Med Chem 2005 ; 48 : 4111 – 9 .

Wang K , Jian Y , Wang H et al.  RBind: computational network method to predict RNA binding sites . Bioinformatics 2018 ; 34 : 3131 – 6 .

Wang XL , Zhang RF , Shen CH et al.  Dense contrastive learning for self-supervised visual pre-training . Proc Cvpr Ieee 2021 ; 3023 – 32 .

Wang KL , Zhou RY , Wu YF et al.  RLBind: a deep learning method to predict RNA-ligand binding sites . Brief Bioinform 2023 ; 24 :bbac486.

wwPDB consortium . Protein Data Bank: the single global archive for 3D macromolecular structure data . Nucleic Acids Res 2019 ; 47 : D520 – 8 .

Xia C , Feng S-H , Xia Y et al.  Fast protein structure comparison through effective representation learning with contrastive graph neural networks . PLoS Comput Biol 2022 ; 18 : e1009986 .

Xie EZ , Ding J , Wang WH et al.  DetCo: unsupervised contrastive learning for object detection. In: 2021 IEEE/CVF International Conference on Computer Vision (ICCV 2021) , pp. 8372 – 81 . 2021 .

Yan L-H , Le Roux A , Boyapelly K et al.  Purine analogs targeting the guanine riboswitch as potential antibiotics against Clostridioides difficile . Eur J Med Chem 2018 ; 143 : 755 – 68 .

Yan Z , Wang J. SPA-LN: a scoring function of ligand-nucleic acid interactions via optimizing both specificity and affinity . Nucleic Acids Res 2017 ; 45 : e110 .

Zeng P , Li J , Ma W et al.  Rsite: a computational method to identify the functional sites of noncoding RNAs . Sci Rep 2015 ; 5 : 9179 .

Zhao XY , Vemulapalli R , Mansfield PA et al.  Contrastive learning for label efficient semantic segmentation. In: 2021 IEEE/CVF International Conference on Computer Vision (ICCV 2021) , pp. 10603 – 13 . 2021 .

Zhou T , Wang H , Zeng C et al.  RPocket: an intuitive database of RNA pocket topology information with RNA-ligand data resources . BMC Bioinformatics 2021 ; 22 : 428 .

Zogg H , Singh R , Ro S. Current advances in RNA therapeutics for human diseases . Int J Mol Sci 2022 ; 23 : 2736 .

Email alerts

Citing articles via, looking for your next opportunity.

• Recommend to your Library

Affiliations

• Online ISSN 1367-4811
• Copyright © 2024 Oxford University Press
• About Oxford Academic
• Publish journals with us
• University press partners
• What we publish
• New features  
• Open access
• Institutional account management
• Rights and permissions
• Get help with access
• Accessibility
• Advertising
• Media enquiries
• Oxford University Press
• Oxford Languages
• University of Oxford

Oxford University Press is a department of the University of Oxford. It furthers the University's objective of excellence in research, scholarship, and education by publishing worldwide

• Copyright © 2024 Oxford University Press
• Cookie settings
• Cookie policy
• Privacy policy
• Legal notice

This Feature Is Available To Subscribers Only

Sign In or Create an Account

This PDF is available to Subscribers Only

For full access to this pdf, sign in to an existing account, or purchase an annual subscription.

• Previous Article
• Next Article

Implementation of transfer learning in convolutional neural network architecture for android-based handwriting quality detection

• Split-Screen
• Article contents
• Figures & tables
• Supplementary Data
• Peer Review
• Open the PDF for in another window
• Reprints and Permissions
• Cite Icon Cite
• Search Site

Muhammad Zidni Subarkah , Winita Sulandari , Respatiwulan Respatiwulan; Implementation of transfer learning in convolutional neural network architecture for android-based handwriting quality detection. AIP Conf. Proc. 9 April 2024; 3095 (1): 070003. https://doi.org/10.1063/5.0204724

Download citation file:

• Ris (Zotero)
• Reference Manager

Handwriting is a number, letter, word, or sentence written on a piece of paper that comes from human or individual handwriting. Handwriting has benefits that can help fine motor coordination, memory, and cognitive development in multi-sensory activities. There are several problems in the world of education that involve handwriting. First, handwriting that is difficult to read can cause students to get low grades and even fail to graduate from school. Second, sloppy handwriting makes students lose confidence. Third, students who have bad handwriting have the potential to be bullied by friends at school. The description of these problems causes the author to be interested in developing a digital technology innovation in detecting the quality of handwriting. The purpose of this research is to detect handwriting quality using a Convolutional Neural Network in transfer learning process of EfficientNet B0 model architecture on fine tuning. The data used are secondary data from the CSAFE Handwriting Database at Iowa State University collected by researchers at the Center for Statistics and Applications in Forensic Evidence. This data consists of 27 scanned handwriting samples from each of the 90 participants for a total of 2430 handwritten image samples. The model generated from EfficientNet B0 SGD optimization using learning rate of 0.01 at 28th epochs is very good model obtained. Moreover, the model was evaluated for accuracy, precision, recall, and F 1 −score of 92%. Next, the model is integrated into cloud computing. Furthermore, implementing the best model, android application is developed called Rayuan (Rate Your Handwriting).

Citing articles via

Publish with us - request a quote.

Sign up for alerts

• Online ISSN 1551-7616
• Print ISSN 0094-243X
• For Researchers
• For Librarians
• For Advertisers
• Our Publishing Partners  
• Physics Today
• Conference Proceedings
• Special Topics

pubs.aip.org

• Privacy Policy
• Terms of Use

Connect with AIP Publishing

This feature is available to subscribers only.

Sign In or Create an Account