Structural Basis for Mitotic Centrosome Assembly in Flies.

In flies, Centrosomin (Cnn) forms a phosphorylation-dependent scaffold that recruits proteins to the mitotic centrosome, but how Cnn assembles into a scaffold is unclear. We show that scaffold assembly requires conserved leucine zipper (LZ) and Cnn-motif 2 (CM2) domains that co-assemble into a 2:2 complex in vitro. We solve the crystal structure of the LZ:CM2 complex, revealing that both proteins form helical dimers that assemble into an unusual tetramer. A slightly longer version of the LZ can form micron-scale structures with CM2, whose assembly is stimulated by Plk1 phosphorylation in vitro. Mutating individual residues that perturb LZ:CM2 tetramer assembly perturbs the formation of these micron-scale assemblies in vitro and Cnn-scaffold assembly in vivo. Thus, Cnn molecules have an intrinsic ability to form large, LZ:CM2-interaction-dependent assemblies that are critical for mitotic centrosome assembly. These studies provide the first atomic insight into a molecular interaction required for mitotic centrosome assembly.

CatLearning: highly accurate gene expression prediction from histone mark.

Histone modifications, known as histone marks, are pivotal in regulating gene expression within cells. The vast array of potential combinations of histone marks presents a considerable challenge in decoding the regulatory mechanisms solely through biological experimental approaches. To overcome this challenge, we have developed a method called CatLearning. It utilizes a modified convolutional neural network architecture with a specialized adaptation Residual Network to quantitatively interpret histone marks and predict gene expression. This architecture integrates long-range histone information up to 500Kb and learns chromatin interaction features without 3D information. By using only one histone mark, CatLearning achieves a high level of accuracy. Furthermore, CatLearning predicts gene expression by simulating changes in histone modifications at enhancers and throughout the genome. These findings help comprehend the architecture of histone marks and develop diagnostic and therapeutic targets for diseases with epigenetic changes.

Improved prediction of DNA and RNA binding proteins with deep learning models.

Nucleic acid-binding proteins (NABPs), including DNA-binding proteins (DBPs) and RNA-binding proteins (RBPs), play important roles in essential biological processes. To facilitate functional annotation and accurate prediction of different types of NABPs, many machine learning-based computational approaches have been developed. However, the datasets used for training and testing as well as the prediction scopes in these studies have limited their applications. In this paper, we developed new strategies to overcome these limitations by generating more accurate and robust datasets and developing deep learning-based methods including both hierarchical and multi-class approaches to predict the types of NABPs for any given protein. The deep learning models employ two layers of convolutional neural network and one layer of long short-term memory. Our approaches outperform existing DBP and RBP predictors with a balanced prediction between DBPs and RBPs, and are more practically useful in identifying novel NABPs. The multi-class approach greatly improves the prediction accuracy of DBPs and RBPs, especially for the DBPs with ~12% improvement. Moreover, we explored the prediction accuracy of single-stranded DNA binding proteins and their effect on the overall prediction accuracy of NABP predictions.

miWords: transformer-based composite deep learning for highly accurate discovery of pre-miRNA regions across plant genomes.

Discovering pre-microRNAs (miRNAs) is the core of miRNA discovery. Using traditional sequence/structural features, many tools have been published to discover miRNAs. However, in practical applications like genomic annotations, their actual performance has been very low. This becomes more grave in plants where unlike animals pre-miRNAs are much more complex and difficult to identify. A huge gap exists between animals and plants for the available software for miRNA discovery and species-specific miRNA information. Here, we present miWords, a composite deep learning system of transformers and convolutional neural networks which sees genome as a pool of sentences made of words with specific occurrence preferences and contexts, to accurately identify pre-miRNA regions across plant genomes. A comprehensive benchmarking was done involving >10 software representing different genre and many experimentally validated datasets. miWords emerged as the best one while breaching accuracy of 98% and performance lead of ~10%. miWords was also evaluated across Arabidopsis genome where also it outperformed the compared tools. As a demonstration, miWords was run across the tea genome, reporting 803 pre-miRNA regions, all validated by small RNA-seq reads from multiple samples, and most of them were functionally supported by the degradome sequencing data. miWords is freely available as stand-alone source codes at https://scbb.ihbt.res.in/miWords/index.php.

iAMP-Attenpred: a novel antimicrobial peptide predictor based on BERT feature extraction method and CNN-BiLSTM-Attention combination model.

As a kind of small molecule protein that can fight against various microorganisms in nature, antimicrobial peptides (AMPs) play an indispensable role in maintaining the health of organisms and fortifying defenses against diseases. Nevertheless, experimental approaches for AMP identification still demand substantial allocation of human resources and material inputs. Alternatively, computing approaches can assist researchers effectively and promptly predict AMPs. In this study, we present a novel AMP predictor called iAMP-Attenpred. As far as we know, this is the first work that not only employs the popular BERT model in the field of natural language processing (NLP) for AMPs feature encoding, but also utilizes the idea of combining multiple models to discover AMPs. Firstly, we treat each amino acid from preprocessed AMPs and non-AMP sequences as a word, and then input it into BERT pre-training model for feature extraction. Moreover, the features obtained from BERT method are fed to a composite model composed of one-dimensional CNN, BiLSTM and attention mechanism for better discriminating features. Finally, a flatten layer and various fully connected layers are utilized for the final classification of AMPs. Experimental results reveal that, compared with the existing predictors, our iAMP-Attenpred predictor achieves better performance indicators, such as accuracy, precision and so on. This further demonstrates that using the BERT approach to capture effective feature information of peptide sequences and combining multiple deep learning models are effective and meaningful for predicting AMPs.

MITNet: a fusion transformer and convolutional neural network architecture approach for T-cell epitope prediction.

Classifying epitopes is essential since they can be applied in various fields, including therapeutics, diagnostics and peptide-based vaccines. To determine the epitope or peptide against an antibody, epitope mapping with peptides is the most extensively used method. However, this method is more time-consuming and inefficient than using present methods. The ability to retrieve data on protein sequences through laboratory procedures has led to the development of computational models that predict epitope binding based on machine learning and deep learning (DL). It has also evolved to become a crucial part of developing effective cancer immunotherapies. This paper proposes an architecture to generalize this case since various research strives to solve a low-performance classification problem. A proposed DL model is the fusion architecture, which combines two architectures: Transformer architecture and convolutional neural network (CNN), called MITNet and MITNet-Fusion. Combining these two architectures enriches feature space to correlate epitope labels with the binary classification method. The selected epitope-T-cell receptor (TCR) interactions are GILG, GLCT and NLVP, acquired from three databases: IEDB, VDJdb and McPAS-TCR. The previous input data was extracted using amino acid composition, dipeptide composition, spectrum descriptor and the combination of all those features called AADIP composition to encode the input data to DL architecture. For ensuring consistency, fivefold cross-validations were performed using the area under curve metric. Results showed that GILG, GLCT and NLVP received scores of 0.85, 0.87 and 0.86, respectively. Those results were compared to prior architecture and outperformed other similar deep learning models.

Predicting miRNA-disease associations based on lncRNA-miRNA interactions and graph convolution networks.

Increasing studies have proved that microRNAs (miRNAs) are critical biomarkers in the development of human complex diseases. Identifying disease-related miRNAs is beneficial to disease prevention, diagnosis and remedy. Based on the assumption that similar miRNAs tend to associate with similar diseases, various computational methods have been developed to predict novel miRNA-disease associations (MDAs). However, selecting proper features for similarity calculation is a challenging task because of data deficiencies in biomedical science. In this study, we propose a deep learning-based computational method named MAGCN to predict potential MDAs without using any similarity measurements. Our method predicts novel MDAs based on known lncRNA-miRNA interactions via graph convolution networks with multichannel attention mechanism and convolutional neural network combiner. Extensive experiments show that the average area under the receiver operating characteristic values obtained by our method under 2-fold, 5-fold and 10-fold cross-validations are 0.8994, 0.9032 and 0.9044, respectively. When compared with five state-of-the-art methods, MAGCN shows improvement in terms of prediction accuracy. In addition, we conduct case studies on three diseases to discover their related miRNAs, and find that all the top 50 predictions for all the three diseases have been supported by established databases. The comprehensive results demonstrate that our method is a reliable tool in detecting new disease-related miRNAs.

APEX-pHLA: A novel method for accurate prediction of the binding between exogenous short peptides and HLA class I molecules.

Human leukocyte antigen (HLA) molecules play critically significant role within the realm of immunotherapy due to their capacities to recognize and bind exogenous antigens such as peptides, subsequently delivering them to immune cells. Predicting the binding between peptides and HLA molecules (pHLA) can expedite the screening of immunogenic peptides and facilitate vaccine design. However, traditional experimental methods are time-consuming and inefficient. In this study, an efficient method based on deep learning was developed for predicting peptide-HLA binding, which treated peptide sequences as linguistic entities. It combined the architectures of textCNN and BiLSTM to create a deep neural network model called APEX-pHLA. This model operated without limitations related to HLA class I allele variants and peptide segment lengths, enabling efficient encoding of sequence features for both HLA and peptide segments. On the independent test set, the model achieved Accuracy, ROC_AUC, F1, and MCC is 0.9449, 0.9850, 0.9453, and 0.8899, respectively. Similarly, on an external test set, the results were 0.9803, 0.9574, 0.8835, and 0.7863, respectively. These findings outperformed fifteen methods previously reported in the literature. The accurate prediction capability of the APEX-pHLA model in peptide-HLA binding might provide valuable insights for future HLA vaccine design.

Machine learning aided single cell image analysis improves understanding of morphometric heterogeneity of human mesenchymal stem cells.

The multipotent stem cells of our body have been largely harnessed in biotherapeutics. However, as they are derived from multiple anatomical sources, from different tissues, human mesenchymal stem cells (hMSCs) are a heterogeneous population showing ambiguity in their in vitro behavior. Intra-clonal population heterogeneity has also been identified and pre-clinical mechanistic studies suggest that these cumulatively depreciate the therapeutic effects of hMSC transplantation. Although various biomarkers identify these specific stem cell populations, recent artificial intelligence-based methods have capitalized on the cellular morphologies of hMSCs, opening a new approach to understand their attributes. A robust and rapid platform is required to accommodate and eliminate the heterogeneity observed in the cell population, to standardize the quality of hMSC therapeutics globally. Here, we report our primary findings of morphological heterogeneity observed within and across two sources of hMSCs namely, stem cells from human exfoliated deciduous teeth (SHEDs) and human Wharton jelly mesenchymal stem cells (hWJ MSCs), using real-time single-cell images generated on immunophenotyping by imaging flow cytometry (IFC). We used the ImageJ software for identification and comparison between the two types of hMSCs using statistically significant morphometric descriptors that are biologically relevant. To expand on these insights, we have further applied deep learning methods and successfully report the development of a Convolutional Neural Network-based image classifier. In our research, we introduced a machine learning methodology to streamline the entire procedure, utilizing convolutional neural networks and transfer learning for binary classification, achieving an accuracy rate of 97.54%. We have also critically discussed the challenges, comparisons between solutions and future directions of machine learning in hMSC classification in biotherapeutics.

Deepm6A-MT: A deep learning-based method for identifying RNA N6-methyladenosine sites in multiple tissues.

N6-methyladenosine (m6A) is the most prevalent, abundant, and conserved internal modification in the eukaryotic messenger RNA (mRNAs) and plays a crucial role in the cellular process. Although more than ten methods were developed for m6A detection over the past decades, there were rooms left to improve the predictive accuracy and the efficiency. In this paper, we proposed an improved method for predicting m6A modification sites, which was based on bi-directional gated recurrent unit (Bi-GRU) and convolutional neural networks (CNN), called Deepm6A-MT. The Deepm6A-MT has two input channels. One is to use an embedding layer followed by the Bi-GRU and then by the CNN, and another is to use one-hot encoding, dinucleotide one-hot encoding, and nucleotide chemical property codes. We trained and evaluated the Deepm6A-MT both by the 5-fold cross-validation and the independent test. The empirical tests showed that the Deepm6A-MT achieved the state of the art performance. In addition, we also conducted the cross-species and the cross-tissues tests to further verify the Deepm6A-MT for effectiveness and efficiency. Finally, for the convenience of academic research, we deployed the Deepm6A-MT to the web server, which is accessed at the URL http://www.biolscience.cn/Deepm6A-MT/.

Deep learning predicts DNA methylation regulatory variants in the human brain and elucidates the genetics of psychiatric disorders.

There is growing evidence for the role of DNA methylation (DNAm) quantitative trait loci (mQTLs) in the genetics of complex traits, including psychiatric disorders. However, due to extensive linkage disequilibrium (LD) of the genome, it is challenging to identify causal genetic variations that drive DNAm levels by population-based genetic association studies. This limits the utility of mQTLs for fine-mapping risk loci underlying psychiatric disorders identified by genome-wide association studies (GWAS). Here we present INTERACT, a deep learning model that integrates convolutional neural networks with transformer, to predict effects of genetic variations on DNAm levels at CpG sites in the human brain. We show that INTERACT-derived DNAm regulatory variants are not confounded by LD, are concentrated in regulatory genomic regions in the human brain, and are convergent with mQTL evidence from genetic association analysis. We further demonstrate that predicted DNAm regulatory variants are enriched for heritability of brain-related traits and improve polygenic risk prediction for schizophrenia across diverse ancestry samples. Finally, we applied predicted DNAm regulatory variants for fine-mapping schizophrenia GWAS risk loci to identify potential novel risk genes. Our study shows the power of a deep learning approach to identify functional regulatory variants that may elucidate the genetic basis of complex traits.

An efficient hybrid deep learning architecture for predicting short antimicrobial peptides.

Short-length antimicrobial peptides (AMPs) have been demonstrated to have intensified antimicrobial activities against a wide spectrum of microbes. Therefore, exploration of novel and promising short AMPs is highly essential in developing various types of antimicrobial drugs or treatments. In addition to experimental approaches, computational methods have been developed to improve screening efficiency. Although existing computational methods have achieved satisfactory performance, there is still much room for model improvement. In this study, we proposed iAMP-DL, an efficient hybrid deep learning architecture, for predicting short AMPs. The model was constructed using two well-known deep learning architectures: the long short-term memory architecture and convolutional neural networks. To fairly assess the performance of the model, we compared our model with existing state-of-the-art methods using the same independent test set. Our comparative analysis shows that iAMP-DL outperformed other methods. Furthermore, to assess the robustness and stability of our model, the experiments were repeated 10 times to observe the variation in prediction efficiency. The results demonstrate that iAMP-DL is an effective, robust, and stable framework for detecting promising short AMPs. Another comparative study of different negative data sampling methods also confirms the effectiveness of our method and demonstrates that it can also be used to develop a robust model for predicting AMPs in general. The proposed framework was also deployed as an online web server with a user-friendly interface to support the research community in identifying short AMPs.

O-GlcNAcPRED-DL: Prediction of Protein O-GlcNAcylation Sites Based on an Ensemble Model of Deep Learning.

O-linked ß-N-acetylglucosamine (O-GlcNAc) is a post-translational modification (i.e., O-GlcNAcylation) on serine/threonine residues of proteins, regulating a plethora of physiological and pathological events. As a dynamic process, O-GlcNAc functions in a site-specific manner. However, the experimental identification of the O-GlcNAc sites remains challenging in many scenarios. Herein, by leveraging the recent progress in cataloguing experimentally identified O-GlcNAc sites and advanced deep learning approaches, we establish an ensemble model, O-GlcNAcPRED-DL, a deep learning-based tool, for the prediction of O-GlcNAc sites. In brief, to make a benchmark O-GlcNAc data set, we extracted the information on O-GlcNAc from the recently constructed database O-GlcNAcAtlas, which contains thousands of experimentally identified and curated O-GlcNAc sites on proteins from multiple species. To overcome the imbalance between positive and negative data sets, we selected five groups of negative data sets in humans and mice to construct an ensemble predictor based on connection of a convolutional neural network and bidirectional long short-term memory. By taking into account three types of sequence information, we constructed four network frameworks, with the systematically optimized parameters used for the models. The thorough comparison analysis on two independent data sets of humans and mice and six independent data sets from other species demonstrated remarkably increased sensitivity and accuracy of the O-GlcNAcPRED-DL models, outperforming other existing tools. Moreover, a user-friendly Web server for O-GlcNAcPRED-DL has been constructed, which is freely available at http://oglcnac.org/pred_dl.

Miffi: Improving the accuracy of CNN-based cryo-EM micrograph filtering with fine-tuning and Fourier space information.

Efficient and high-accuracy filtering of cryo-electron microscopy (cryo-EM) micrographs is an emerging challenge with the growing speed of data collection and sizes of datasets. Convolutional neural networks (CNNs) are machine learning models that have been proven successful in many computer vision tasks, and have been previously applied to cryo-EM micrograph filtering. In this work, we demonstrate that two strategies, fine-tuning models from pretrained weights and including the power spectrum of micrographs as input, can greatly improve the attainable prediction accuracy of CNN models. The resulting software package, Miffi, is open-source and freely available for public use (https://github.com/ando-lab/miffi).

A systematic review on deep learning-based automated cancer diagnosis models.

Deep learning is gaining importance due to its wide range of applications. Many researchers have utilized deep learning (DL) models for the automated diagnosis of cancer patients. This paper provides a systematic review of DL models for automated diagnosis of cancer patients. Initially, various DL models for cancer diagnosis are presented. Five major categories of cancers such as breast, lung, liver, brain and cervical cancer are considered. As these categories of cancers have a very high percentage of occurrences with high mortality rate. The comparative analysis of different types of DL models is drawn for the diagnosis of cancer at early stages by considering the latest research articles from 2016 to 2022. After comprehensive comparative analysis, it is found that most of the researchers achieved appreciable accuracy with implementation of the convolutional neural network model. These utilized the pretrained models for automated diagnosis of cancer patients. Various shortcomings with the existing DL-based automated cancer diagnosis models are also been presented. Finally, future directions are discussed to facilitate further research for automated diagnosis of cancer patients.

Fully Automated Hippocampus Segmentation using T2-informed Deep Convolutional Neural Networks.

Hippocampal atrophy (tissue loss) has become a fundamental outcome parameter in clinical trials on Alzheimer's disease. To accurately estimate hippocampus volume and track its volume loss, a robust and reliable segmentation is essential. Manual hippocampus segmentation is considered the gold standard but is extensive, time-consuming, and prone to rater bias. Therefore, it is often replaced by automated programs like FreeSurfer, one of the most commonly used tools in clinical research. Recently, deep learning-based methods have also been successfully applied to hippocampus segmentation. The basis of all approaches are clinically used T1-weighted whole-brain MR images with approximately 1 mm isotropic resolution. However, such T1 images show low contrast-to-noise ratios (CNRs), particularly for many hippocampal substructures, limiting delineation reliability. To overcome these limitations, high-resolution T2-weighted scans are suggested for better visualization and delineation, as they show higher CNRs and usually allow for higher resolutions. Unfortunately, such time-consuming T2-weighted sequences are not feasible in a clinical routine. We propose an automated hippocampus segmentation pipeline leveraging deep learning with T2-weighted MR images for enhanced hippocampus segmentation of clinical T1-weighted images based on a series of 3D convolutional neural networks and a specifically acquired multi-contrast dataset. This dataset consists of corresponding pairs of T1- and high-resolution T2-weighted images, with the T2 images only used to create more accurate manual ground truth annotations and to train the segmentation network. The T2-based ground truth labels were also used to evaluate all experiments by comparing the masks visually and by various quantitative measures. We compared our approach with four established state-of-the-art hippocampus segmentation algorithms (FreeSurfer, ASHS, HippoDeep, HippMapp3r) and demonstrated a superior segmentation performance. Moreover, we found that the automated segmentation of T1-weighted images benefits from the T2-based ground truth data. In conclusion, this work showed the beneficial use of high-resolution, T2-based ground truth data for training an automated, deep learning-based hippocampus segmentation and provides the basis for a reliable estimation of hippocampal atrophy in clinical studies.

Heartbeat-related spectral perturbation of electroencephalogram reflects dynamic interoceptive attention states in the trial-by-trial classification analysis.

Attending to heartbeats for interoceptive awareness initiates distinct electrophysiological responses synchronized with the R-peaks of an electrocardiogram (ECG), such as the heartbeat-evoked potential (HEP). Beyond HEP, this study proposes heartbeat-related spectral perturbation (HRSP), a time-frequency map of the R-peak locked electroencephalogram (EEG), and explores its characteristics in identifying interoceptive attention states using a classification approach. HRSPs of EEG brain components specified by independent component analysis (ICA) were used for the offline and online classification of interoceptive states. A convolutional neural network (CNN) designed specifically for HRSP was applied to publicly available data from a binary-state experiment (attending to self-heartbeats and white noise) and data from our four-state classification experiment (attending to self-heartbeats, white noise, time passage, and toe) with diverse input feature conditions of HRSP. From the dynamic state perspective, we evaluated the primary frequency bands of HRSP and the minimal number of averaging epochs required to reflect changing interoceptive attention states without compromising accuracy. We also assessed the utility of group ICA and models for classifying HRSP in new participants. The CNN for trial-by-trial HRSP with actual R-peaks demonstrated significantly higher classification accuracy than HRSP with sham, i.e., randomly positioned, R-peaks. Gradient-weighted class activation mapping highlighted the prominent role of theta and alpha bands between 200-600 ms post-R-peak-features absent in classifications using sham HRSPs. Online classification benefits from employing a group ICA and classification model, ensuring reliable accuracy without individual EEG precollection. These results suggest HRSP's potential to reflect interoceptive attention states, proposing transformative implications for clinical applications.

Light3DHS: A lightweight 3D hippocampus segmentation method using multiscale convolution attention and vision transformer.

The morphological analysis and volume measurement of the hippocampus are crucial to the study of many brain diseases. Therefore, an accurate hippocampal segmentation method is beneficial for the development of clinical research in brain diseases. U-Net and its variants have become prevalent in hippocampus segmentation of Magnetic Resonance Imaging (MRI) due to their effectiveness, and the architecture based on Transformer has also received some attention. However, some existing methods focus too much on the shape and volume of the hippocampus rather than its spatial information, and the extracted information is independent of each other, ignoring the correlation between local and global features. In addition, many methods cannot be effectively applied to practical medical image segmentation due to many parameters and high computational complexity. To this end, we combined the advantages of CNNs and ViTs (Vision Transformer) and proposed a simple and lightweight model: Light3DHS for the segmentation of the 3D hippocampus. In order to obtain richer local contextual features, the encoder first utilizes a multi-scale convolutional attention module (MCA) to learn the spatial information of the hippocampus. Considering the importance of local features and global semantics for 3D segmentation, we used a lightweight ViT to learn high-level features of scale invariance and further fuse local-to-global representation. To evaluate the effectiveness of encoder feature representation, we designed three decoders of different complexity to generate segmentation maps. Experiments on three common hippocampal datasets demonstrate that the network achieves more accurate hippocampus segmentation with fewer parameters. Light3DHS performs better than other state-of-the-art algorithms.

Role of Optimization in RNA-Protein-Binding Prediction.

RNA-binding proteins (RBPs) play an important role in regulating biological processes, such as gene regulation. Understanding their behaviors, for example, their binding site, can be helpful in understanding RBP-related diseases. Studies have focused on predicting RNA binding by means of machine learning algorithms including deep convolutional neural network models. One of the integral parts of modeling deep learning is achieving optimal hyperparameter tuning and minimizing a loss function using optimization algorithms. In this paper, we investigate the role of optimization in the RBP classification problem using the CLIP-Seq 21 dataset. Three optimization methods are employed on the RNA-protein binding CNN prediction model; namely, grid search, random search, and Bayesian optimizer. The empirical results show an AUC of 94.42%, 93.78%, 93.23% and 92.68% on the ELAVL1C, ELAVL1B, ELAVL1A, and HNRNPC datasets, respectively, and a mean AUC of 85.30 on 24 datasets. This paper's findings provide evidence on the role of optimizers in improving the performance of RNA-protein binding prediction.

Machine/deep learning-assisted hemoglobin level prediction using palpebral conjunctival images.

Palpebral conjunctival hue alteration is used in non-invasive screening for anaemia, whereas it is a qualitative measure. This study constructed machine/deep learning models for predicting haemoglobin values using 150 palpebral conjunctival images taken by a smartphone. The median haemoglobin value was 13.1 g/dL, including 10 patients with <11 g/dL. A segmentation model using U-net was successfully constructed. The segmented images were subjected to non-convolutional neural network (CNN)-based and CNN-based regression models for predicting haemoglobin values. The correlation coefficients between the actual and predicted haemoglobin values were 0.38 and 0.44 in the non-CNN-based and CNN-based models, respectively. The sensitivity and specificity for anaemia detection were 13% and 98% for the non-CNN-based model and 20% and 99% for the CNN-based model. The performance of the CNN-based model did not improve with a mask layer guiding the model's attention towards the conjunctival regions, however, slightly improved with correction by the aspect ratio and exposure time of input images. The gradient-weighted class activation mapping heatmap indicated that the lower half area of the conjunctiva was crucial for haemoglobin value prediction. In conclusion, the CNN-based model had better results than the non-CNN-based model. The prediction accuracy would improve by using more input data with anaemia.