With the exponential growth of biomedical literature, leveraging Large Language Models (LLMs) for automated medical knowledge understanding has become increasingly critical for advancing precision medicine. However, current approaches face significant challenges in reliability, verifiability, and scalability when extracting complex biological relationships from scientific literature using LLMs. To overcome the obstacles of LLM development in biomedical literature understating, we propose LORE, a novel unsupervised two-stage reading methodology with LLM that models literature as a knowledge graph of verifiable factual statements and, in turn, as semantic embeddings in Euclidean space. LORE captured essential gene pathogenicity information when applied to PubMed abstracts for large-scale understanding of disease-gene relationships. We demonstrated that modeling a latent pathogenic flow in the semantic embedding with supervision from the ClinVar database led to a 90% mean average precision in identifying relevant genes across 2097 diseases. This work provides a scalable and reproducible approach for leveraging LLMs in biomedical literature analysis, offering new opportunities for researchers to identify therapeutic targets efficiently.

Since the American College of Medical Genetics and Genomics (ACMG) published recommendations for reporting incidental findings in clinical exome and genome sequencing in 2013, the list of suggested actionable genes appear more and more frequently in the reports of genetic testing. However, the existence of low-coverage regions, no matter happening systematically or randomly, in whole genome sequencing (WGS) or whole exome sequencing (WES) raise the chances of false negatives when reporting pathogenicity in variant discovery. To evaluate the effectivity of different sequencing methods, this study developed a workflow to examine low-coverage regions in ACMG actionable genes (version 2 contains 59 genes) on three groups of sequencing data, including 5 samples of 30X WGS, 41 samples of 40X WGS, and 11 samples of WES. By extracting the ACMG regions of the alignment files, we first found that per-base mean coverage in WES is significantly higher in some chromosomes (2, 5, 6, 9, 13, 14, 17, 18) than the others, and the overall average (44.5) is higher than 30X WGS (32.4) and 40X WGS (36.2). However, the total length of the low-coverage regions in ACMG actionable genes is higher in WES samples: 23 times higher than 40X WGS, and 8 times higher than 30X WGS. When looking into ClinVar significance of pathogenic/likely pathogenic (P/LP) variants (the 2020/03/16 release contains 21,388 P/LP variants falling in the 59 actionable genes), 100% (11/11) WES samples and 80% (4/5) 30X WGS samples have more than 20 P/LP variants in the low-coverage regions, while only 2.4% (1/41) 40X WGS samples have such a risk of missing >20 P/LP variants. We conclude that in order to reduce the possibility of missing clinically significant variants, 40X WGS are more recommended than 30X WGS and WES to achieve satisfied quality of characterizing ACMG actionable genes in genetic testing.

MOTIVATION: To facilitate the process of tailor-making a deep neural network for exploring the dynamics of genomic DNA, we have developed a hands-on package called ezGeno. ezGeno automates the search process of various parameters and network structures and can be applied to any kind of 1D genomic data. Combinations of multiple abovementioned 1D features are also applicable. RESULTS: For the task of predicting TF binding using genomic sequences as the input, ezGeno can consistently return the best performing set of parameters and network structure, as well as highlight the important segments within the original sequences. For the task of predicting tissue-specific enhancer activity using both sequence and DNase feature data as the input, ezGeno also regularly outperforms the hand-designed models. Furthermore, we demonstrate that ezGeno is superior in efficiency and accuracy compared to the one-layer DeepBind model and AutoKeras, an open-source AutoML package. AVAILABILITY AND IMPLEMENTATION: The ezGeno package can be freely accessed at https://github.com/ailabstw/ezGeno. SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

Homology-based transfer is frequently used to predict protein functions of unannotated sequences through similarity analysis between the target and previously annotated sequences. The most direct and accessible homology-based transfer approach is sequence alignment. To assess the reliability of alignment-based prediction, we applied a 10-fold cross-validation test in SWISS-Prot database. We compared Matthews correlation coefficient, sensitivity, as well as precision, and examined different parameter settings used in the alignment-based methods, with BLASTp and PSI-BLAST. As the results shown in this paper, in the categories of domain, ligand, molecular function, biological process, cellular component, and PTM, the keywords can be confidently used for protein function predictions, whereas the others are less reliable.

BACKGROUND: Genome-wide association studies (GWAS) provide a powerful means to identify associations between genetic variants and phenotypes. However, GWAS techniques for detecting epistasis, the interactions between genetic variants associated with phenotypes, are still limited. We believe that developing an efficient and effective GWAS method to detect epistasis will be a key for discovering sophisticated pathogenesis, which is especially important for complex diseases such as Alzheimer's disease (AD). RESULTS: In this regard, this study presents GenEpi, a computational package to uncover epistasis associated with phenotypes by the proposed machine learning approach. GenEpi identifies both within-gene and cross-gene epistasis through a two-stage modeling workflow. In both stages, GenEpi adopts two-element combinatorial encoding when producing features and constructs the prediction models by L1-regularized regression with stability selection. The simulated data showed that GenEpi outperforms other widely-used methods on detecting the ground-truth epistasis. As real data is concerned, this study uses AD as an example to reveal the capability of GenEpi in finding disease-related variants and variant interactions that show both biological meanings and predictive power. CONCLUSIONS: The results on simulation data and AD demonstrated that GenEpi has the ability to detect the epistasis associated with phenotypes effectively and efficiently. The released package can be generalized to largely facilitate the studies of many complex diseases in the near future.

In the past few years, many variants in the non-coding regions of the human genome have been reported by personal whole genome sequencing. It is a challenge to distinguish pathogenic non-coding variants from such a large number of benign non-coding variants. Many machine learning methods for predicting pathogenic non-coding variants have been proposed. However, the precision and recall rates of the currently existing methods decline rapidly when the number of negative samples in the data increases. Both under-and over-sampling techniques have been employed in the field of machine learning to resolve the poor performance of classification methods on imbalanced data. Even though, we observed that a more sophisticated method with better performance is still largely desired for the problem of predicting pathogenic non-coding variants. In this regard, this study aims at presenting a general framework for imbalanced data learning, CE-SMURF, which incorporates both Cluster Ensemble (CE) sampling and hyper-ensemble techniques to further improve the prediction accuracy of detecting pathogenic non-coding variants. The results demonstrate that the final setting of CE-SMURF (f = 0, r = 0.1) is superior in training, and outperforms other existing methods on the testing data, providing a valuable insight to tackle the imbalanced learning issue for many future applications in the field of genomic precision medicine.

Face detection is one of the most popular topics in computer vision. There are several well-known techniques for face detection, such as the Viola-Jones detector. However, the performance of the Viola-Jones detector is limited since it mainly applies the simple Haar-based features. Many advanced methods, especially the convolutional neural network (CNN) based method, have very good performance in face detection. However, they require huge amount of training data. Moreover, most of existing algorithms are not robust to rotation, head-up, and head-down cases. In this paper, we find that, with some modifications, the Viola-Jones detector can also have very good performance in face detection. In addition to the Haar features, we also apply the prominent features and the color information. With the contour information, the edge-aware filter, the background smoother, the fuzzy classifier, and the relative locations, the prominent features, such as eyes, mouths, noses, and ears, can be extracted accurately. With these features, the accuracy of face detection can be much improved. Simulations show that, even if huge amount of training data is not applied, the proposed algorithm has better performance than state-of-the-art face detection methods, including the CNN-based method.

MOTIVATION: Gene regulation involves complicated mechanisms such as cooperativity between a set of transcription factors (TFs). Previous studies have used target genes shared by two TFs as a clue to infer TF-TF interactions. However, this task remains challenging because the target genes with low binding affinity are frequently omitted by experimental data, especially when a single strict threshold is employed. This article aims at improving the accuracy of inferring TF-TF interactions by incorporating motif discovery as a fundamental step when detecting overlapping targets of TFs based on ChIP-chip data. RESULTS: The proposed method, simTFBS, outperforms three naïve methods that adopt fixed thresholds when inferring TF-TF interactions based on ChIP-chip data. In addition, simTFBS is compared with two advanced methods and demonstrates its advantages in predicting TF-TF interactions. By comparing simTFBS with predictions based on the set of available annotated yeast TF binding motifs, we demonstrate that the good performance of simTFBS is indeed coming from the additional motifs found by the proposed procedures. SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

DNA-binding proteins perform their functions through specific or non-specific sequence recognition. Although many sequence-based approaches have been proposed to identify DNA-binding residues on proteins or protein-binding sites on DNA sequences with satisfied performance, it remains a challenging task to unveil the exact mechanism of protein-DNA interactions without crystal complex structures. Without information from complexes, the linkages between DNA-binding proteins and their binding sites on DNA are still missing. While it is still difficult to acquire co-crystallized structures in an efficient way, this study proposes a knowledge-based learning method to effectively predict DNA orientation and base locations around the protein's DNA-binding sites when given a protein structure. First, the functionally important residues of a query protein are predicted by a sequential pattern mining tool. After that, surface residues within the functional regions are determined based on the given structure. These residues are then clustered based on their spatial coordinates and the resultant clusters are ranked by a proposed DNA-binding propensity function. Clusters with high DNA-binding propensities are treated as DNA-binding units (DBUs) and each DBU is analyzed by principal component analysis (PCA) to predict potential orientation of DNA grooves. This paper proposes a knowledge-based learning procedure to determine the location of the bound groove by considering geometric propensity between protein side chains and DNA bases. The 11 test cases used in this study reveal that the location and orientation of the DNA groove around a selected DBU can be predicted with satisfied errors. How the proposed method can be incorporated with protein-DNA docking tools to investigate protein-DNA interactions deserve further studies in the near future.

BACKGROUND: A variety of high-throughput techniques are now available for constructing comprehensive gene regulatory networks in systems biology. In this study, we report a new statistical approach for facilitating in silico inference of regulatory network structure. The new measure of association, coefficient of intrinsic dependence (CID), is model-free and can be applied to both continuous and categorical distributions. When given two variables X and Y, CID answers whether Y is dependent on X by examining the conditional distribution of Y given X. In this paper, we apply CID to analyze the regulatory relationships between transcription factors (TFs) (X) and their downstream genes (Y) based on clinical data. More specifically, we use estrogen receptor alpha (ERalpha) as the variable X, and the analyses are based on 48 clinical breast cancer gene expression arrays (48A). RESULTS: The analytical utility of CID was evaluated in comparison with four commonly used statistical methods, Galton-Pearson's correlation coefficient (GPCC), Student's t-test (STT), coefficient of determination (CoD), and mutual information (MI). When being compared to GPCC, CoD, and MI, CID reveals its preferential ability to discover the regulatory association where distribution of the mRNA expression levels on X and Y does not fit linear models. On the other hand, when CID is used to measure the association of a continuous variable (Y) against a discrete variable (X), it shows similar performance as compared to STT, and appears to outperform CoD and MI. In addition, this study established a two-layer transcriptional regulatory network to exemplify the usage of CID, in combination with GPCC, in deciphering gene networks based on gene expression profiles from patient arrays. CONCLUSION: CID is shown to provide useful information for identifying associations between genes and transcription factors of interest in patient arrays. When coupled with the relationships detected by GPCC, the association predicted by CID are applicable to the construction of transcriptional regulatory networks. This study shows how information from different data sources and learning algorithms can be integrated to investigate whether relevant regulatory mechanisms identified in cell models can also be partially re-identified in clinical samples of breast cancers. AVAILABILITY: the implementation of CID in R codes can be freely downloaded from (http://homepage.ntu.edu.tw/~lyliu/BC/).

BACKGROUND: Identification of protein interacting sites is an important task in computational molecular biology. As more and more protein sequences are deposited without available structural information, it is strongly desirable to predict protein binding regions by their sequences alone. This paper presents a pattern mining approach to tackle this problem. It is observed that a functional region of protein structures usually consists of several peptide segments linked with large wildcard regions. Thus, the proposed mining technology considers large irregular gaps when growing patterns, in order to find the residues that are simultaneously conserved but largely separated on the sequences. A derived pattern is called a cluster-like pattern since the discovered conserved residues are always grouped into several blocks, which each corresponds to a local conserved region on the protein sequence. RESULTS: The experiments conducted in this work demonstrate that the derived long patterns automatically discover the important residues that form one or several hot regions of protein-protein interactions. The methodology is evaluated by conducting experiments on the web server MAGIIC-PRO based on a well known benchmark containing 220 protein chains from 72 distinct complexes. Among the tested 218 proteins, there are 900 sequential blocks discovered, 4.25 blocks per protein chain on average. About 92% of the derived blocks are observed to be clustered in space with at least one of the other blocks, and about 66% of the blocks are found to be near the interface of protein-protein interactions. It is summarized that for about 83% of the tested proteins, at least two interacting blocks can be discovered by this approach. CONCLUSION: This work aims to demonstrate that the important residues associated with the interface of protein-protein interactions may be automatically discovered by sequential pattern mining. The detected regions possess high conservation and thus are considered as the computational hot regions. This information would be useful to characterizing protein sequences, predicting protein function, finding potential partners, and facilitating protein docking for drug discovery.

BACKGROUND: More and more disordered regions have been discovered in protein sequences, and many of them are found to be functionally significant. Previous studies reveal that disordered regions of a protein can be predicted by its primary structure, the amino acid sequence. One observation that has been widely accepted is that ordered regions usually have compositional bias toward hydrophobic amino acids, and disordered regions are toward charged amino acids. Recent studies further show that employing evolutionary information such as position specific scoring matrices (PSSMs) improves the prediction accuracy of protein disorder. As more and more machine learning techniques have been introduced to protein disorder detection, extracting more useful features with biological insights attracts more attention. RESULTS: This paper first studies the effect of a condensed position specific scoring matrix with respect to physicochemical properties (PSSMP) on the prediction accuracy, where the PSSMP is derived by merging several amino acid columns of a PSSM belonging to a certain property into a single column. Next, we decompose each conventional physicochemical property of amino acids into two disjoint groups which have a propensity for order and disorder respectively, and show by experiments that some of the new properties perform better than their parent properties in predicting protein disorder. In order to get an effective and compact feature set on this problem, we propose a hybrid feature selection method that inherits the efficiency of uni-variant analysis and the effectiveness of the stepwise feature selection that explores combinations of multiple features. The experimental results show that the selected feature set improves the performance of a classifier built with Radial Basis Function Networks (RBFN) in comparison with the feature set constructed with PSSMs or PSSMPs that adopt simply the conventional physicochemical properties. CONCLUSION: Distinguishing disordered regions from ordered regions in protein sequences facilitates the exploration of protein structures and functions. Results based on independent testing data reveal that the proposed predicting model DisPSSMP performs the best among several of the existing packages doing similar tasks, without either under-predicting or over-predicting the disordered regions. Furthermore, the selected properties are demonstrated to be useful in finding discriminating patterns for order/disorder classification.

Discriminating outer membrane proteins from proteins with other subcellular localizations and with other folding classes are both important to predict farther their functions and structures. In this paper, we propose a method for discriminating outer membrane proteins from other proteins by support vector machines using combinations of gapped amino acid pair compositions. Using 5-fold cross-validation, the method achieves 95% precision and 92% recall on the dataset of proteins with well-annotated subcellular localizations, consisting of 471 outer membrane proteins and 1,120 other proteins. When applied on another dataset of 377 outer membrane proteins and 674 globular proteins belonging to four typical structural classes, the method reaches 96% precision and recall and correctly excludes 98% of the globular proteins. Our method outperforms the OM classifier of PSORTb v.2.0 and a method based on dipeptide composition.

This paper proposes a novel method to construct a radial basis function network (RBFN) classifier. Our contribution consists of two parts. The first one is an incremental hierarchical clustering algorithm for constructing the hidden layer, and the second one is to improve the least mean square error method that calculates the weights between the hidden and the output layers of an RBFN. This paper discusses the effects of incorporating an incremental hierarchical clustering algorithm for constructing an RBFN optimized for data classification applications. The formation of clusters is controlled by the class labels of training samples and therefore the clusters identified are well adapted to the local distributions of training instances. In addition, the incremental framework largely reduces the requirement of memory space when the training data set is large. In regard to the calculation of weights, we employ the regularization theory to solve the singular matrix problem that might happen in determining the optimal weights. Experimental results show that the data classifier constructed is capable of delivering comparable classification accuracy as the support vector machine (SVM) and the kernel density estimation based classifier that we have recently proposed, while enjoying significant execution efficiency in handling data sets that contains a high percentage of redundant training instances.

In recent years, kernel density estimation has been exploited by computer scientists to model several important problems in machine learning, bioinformatics, and computer vision. However, in case the dimension of the data set is high, then the conventional kernel density estimators suffer poor convergence rates of the pointwise mean square error (MSE) and the integrated mean square error (IMSE). Therefore, design of a novel kernel density estimator that overcomes this problem has been a great challenge for many years. This paper proposes a relaxed model of the variable kernel density estimation and analyzes its performance in data classification applications. It is proved in this paper that, in terms of pointwise MSE, the convergence rate of the relaxed variable kernel density estimator can approach O(n/sup -1/) regardless of the dimension of the data set, where n is the number of sampling instances. Experiments with the data classification applications have shown that the improved convergence rate of the pointwise MSE leads to higher prediction accuracy. In fact, the experimental results have also shown that the data classifier constructed based on the relaxed variable kernel density estimator is capable of delivering the same level of prediction accuracy as the SVM with the Gaussian kernel.

This paper presents a novel learning algorithm for efficient construction of the radial basis function (RBF) networks that can deliver the same level of accuracy as the support vector machines (SVMs) in data classification applications. The proposed learning algorithm works by constructing one RBF subnetwork to approximate the probability density function of each class of objects in the training data set. With respect to algorithm design, the main distinction of the proposed learning algorithm is the novel kernel density estimation algorithm that features an average time complexity of O(n log n), where n is the number of samples in the training data set. One important advantage of the proposed learning algorithm, in comparison with the SVM, is that the proposed learning algorithm generally takes far less time to construct a data classifier with an optimized parameter setting. This feature is of significance for many contemporary applications, in particular, for those applications in which new objects are continuously added into an already large database. Another desirable feature of the proposed learning algorithm is that the RBF networks constructed are capable of carrying out data classification with more than two classes of objects in one single run. In other words, unlike with the SVM, there is no need to resort to mechanisms such as one-against-one or one-against-all for handling datasets with more than two classes of objects. The comparison with SVM is of particular interest, because it has been shown in a number of recent studies that SVM generally are able to deliver higher classification accuracy than the other existing data classification algorithms. As the proposed learning algorithm is instance-based, the data reduction issue is also addressed in this paper. One interesting observation in this regard is that, for all three data sets used in data reduction experiments, the number of training samples remaining after a naive data reduction mechanism is applied is quite close to the number of support vectors identified by the SVM software. This paper also compares the performance of the RBF networks constructed with the proposed learning algorithm and those constructed with a conventional cluster-based learning algorithm. The most interesting observation learned is that, with respect to data classification, the distributions of training samples near the boundaries between different classes of objects carry more crucial information than the distributions of samples in the inner parts of the clusters.

Analysis of protein-ligand interactions is a fundamental issue in drug design. As the detailed and accurate analysis of protein-ligand interactions involves calculation of binding free energy based on thermodynamics and even quantum mechanics, which is highly expensive in terms of computing time, conformational and structural analysis of proteins and ligands has been widely employed as a screening process in computer-aided drug design. In this paper, an efficient mechanism for identifying possible protein-ligand interactions based on analysis of protein tertiary substructures is proposed. In one experiment reported in this paper, the proposed prediction mechanism has been exploited to obtain some clues about a hypothesis that the biochemists have been speculating. The main distinction in the design of the prediction mechanism is the filtering process incorporated to expedite the analysis. The filtering process extracts the residues located in a cave of the protein tertiary structure for analysis and operates with O(nlogn) time complexity, where n is the number of residues in the protein. In comparison, the /spl alpha/hull algorithm, which is a widely used algorithm in computer graphics for identifying those instances that are on the contour of a 3-dimensional object, features O(n/sup 2/) time complexity. Experimental results show that the filtering process presented in this paper is able to speed up the analysis by a factor ranging from 3.11 to 9.79 times.

After an analog ECG signal is transferred into digital format, a suitable digital filter can be used to suppress the high-frequency embedded noise. In this paper, we use the equiripple FIR low-pass filter by superimposing of the optimal method, the Butterworth IIR low-pass filter, the 8-point moving-average filter and the FIR filter designed by using a Kaiser window. Furthermore, we combine the 8-point moving-average filter with the FIR filter designed by using a Kaiser window. In addition, we use the mean square error (M. S. E.) to estimate the effect of the digital filters in order to compare the reduction of the embedded high-frequency noise. Hence, we compute the mean square error with respect to the order, N, of these filters and plot the relationship between M. S. E. and N. Finally, we find the relationship between the CPU time and N.

MOTIVATION: Protein sequence clustering has been widely exploited to facilitate in-depth analysis of protein functions and families. For some applications of protein sequence clustering, it is highly desirable that a hierarchical structure, also referred to as dendrogram, which shows how proteins are clustered at various levels, is generated. However, as the sizes of contemporary protein databases continue to grow at rapid rates, it is of great interest to develop some summarization mechanisms so that the users can browse the dendrogram and/or search for the desired information more effectively. RESULTS: In this paper, the design of a novel incremental clustering algorithm aimed at generating summarized dendrograms for analysis of protein databases is described. The proposed incremental clustering algorithm employs a statistics-based model to summarize the distributions of the similarity scores among the proteins in the database and to control formation of clusters. Experimental results reveal that, due to the summarization mechanism incorporated, the proposed incremental clustering algorithm offers the users highly concise dendrograms for analysis of protein clusters with biological significance. Another distinction of the proposed algorithm is its incremental nature. As the sizes of the contemporary protein databases continue to grow at fast rates, due to the concern of efficiency, it is desirable that cluster analysis of a protein database can be carried out incrementally, when the protein database is updated. Experimental results with the Swiss-Prot protein database reveal that the time complexity for carrying out incremental clustering with k new proteins added into the database containing n proteins is O(n2betalogn), where beta congruent with 0.865, provided that k << n. AVAILABILITY: The Linux executable is available on the following supplementary page.

In this paper we propose a technique that reconstructs high-resolution images with improved super-resolution algorithms, based on Irani and Peleg iterative method, and employs our suggested initial interpolation, robust image registration, automatic image selection and image enhancement post-processing. When the target of reconstruction is a moving object with respect to a stationary camera, high-resolution images can still be reconstructed, whereas previous systems only work well when we move the camera and the displacement of the whole scene is the same.

In recent years, Support Vector Machines (SVM) have been extensively applied to deal with various data classification problems. However, in some cases, the application of SVM is limited due to the time taken to conduct model selection for SVM. This issue is of particular significant for some modern applications, such as web mining, in which the large-scale database is frequently updated. This paper proposes a data reduction based mechanism aimed at expediting the model selection process in SVM. Experimental results show that the proposed mechanism is able to greatly reduce the time taken to carry out model selection at minimum cost.

We determine the actual parameters for a class of one-point codes of length 64 and 65 over F/sub 8/ defined by Hansen and Stichtenoth (1990). Several codes have a minimum distance that exceeds the Feng-Rao bound. The codes with parameters [64,5,51], [64,10,42], [64,11,42], [64,12,40], [65,5,52], [65,10,43], [65,11,42], [65,12,41], and [65,13,40] are better than any known code.

Many VLSI architectures for computing the discrete wavelet transform (DWT) were presented, but the parallel input data sequence and the programmability of the 2-D DWT were rarely mentioned. In this paper, we present a parallel-processing VLSI architecture to compute the programmable 2-D DWT, including various wavelet filter lengths and various wavelet transform levels. The proposed architecture is very regular and easy to extend. To eliminate high frequency components, the pixel values outside the boundary of the image are mirror-extended as the symmetric wavelet transform and the mirror-extension is realized via the routing network. This design has been implemented and fabricated in a 0.35 /spl mu/m 1P4M CMOS technology and the working frequency is 50 MHz. The chip size is about 5200 /spl mu/m/spl times/2500 /spl mu/m. For a 256/spl times/256 image, the chip can perform 30 frames per second with the filter length varying from 2 to 20 and with various levels. Hence the proposed architecture is suitable for real-time applications such as JPEG 2000.