Institute of Information Science, Academia Sinica

Research Overview


Press Ctrl+P to print from browser

Recent Research Results


Accelerating Video Captioning on Heterogeneous System Architectures

ACM Transactions on Architecture and Code Optimization (TACO), March To Appear

Horng-Ruey Huang, Ding-Yong Hong, Jan-Jan Wu, Kung-Fu Chen, Pangfeng Liu, and Wei-Chung Hsu

Horng-Ruey Huang Ding-Yong Hong Jan-Jan Wu Kung-Fu Chen


Video captioning is a core technology to many important applications such as AI-assisted medical diagnosis, video question answering, storytelling through videos, and lip-reading. Video captioning employs a hybrid CNN+RNN model. Accelerating such a hybrid model on a heterogeneous system is challenging because (1) CNN and RNN exhibit very different computing behaviors, making the mapping between computation and heterogeneous devices difficult. (2) Data dependency exists between the CNN and RNN within a video frame and between adjacent RNNs across video frames. These data dependencies prohibit the full parallelization of the hybrid model. The issues also include the utilization of accelerator resources, which is critical to maximizing the performance. In this work, we propose a fine-grained scheduling scheme for mapping computation and devices within a video frame, and a pipeline scheduling scheme for exploiting maximum parallelism between the execution of the video frames. In addition, we propose two capacity-guided scheduling methods. On the server, the concurrent kernel execution mechanism is exploited for improving GPU utilization. On the edge platform, we re-arrange CNN computation among the CPU and EdgeTPUs guided by the EdgeTPU’s SRAM capacity, so that balanced computation is achieved and off-chip memory overhead is minimized. Experimental results show that our scheduling scheme improves video captioning performance by up to 3.24× with CPU+GPU collaboration over the GPU-only execution. On an edge platform with an ARM CPU and two EdgeTPUs, our CPU+EdgeTPU scheduling exhibits outstanding performance, which achieves up to 54.9× speedup compared to using ARM CPU only and can perform video captioning of 59 frames per second.

Multi-aspect examinations of possible alternative mappings of identified variant peptides: a case study on the HEK293 cell line

ACS Omega, To Appear

Wai-Kok Choong and Ting-Yi Sung*

Wai-Kok Choong Ting-Yi Sung


Adopting proteogenomics approach to validate single nucleotide variation events by identifying corresponding single amino acid variant peptides from mass spectrometry (MS)-based proteomic data facilitates translational and clinical research. Although variant peptides are usually identified from MS data with a stringent false discovery rate (FDR), FDR control could fail to eliminate dubious results caused by several issues; thus, post-examination to eliminate dubious results is required. However, comprehensive post-examinations of identification results are still lacking. Therefore, we propose a framework of three bottom-up levels, peptide-spectrum match, peptide, and variant event levels, that consists of rigorous eleven-aspect examinations from the MS perspective to further confirm the reliability of variant events. As a proof of concept and showing feasibility, we demonstrate the eleven examinations on the identified variant peptides from an HEK293 cell line data set, where various database search strategies were applied to maximize the number of identified variant PSMs with an FDR <1% for post-examinations. The results showed that only FDR criterion is insufficient to validate identified variant peptides and the eleven post-examinations can reveal low-confidence variant events detected by shotgun proteomics experiments. Therefore, we suggest that post-examinations of identified variant events based on the proposed framework are necessary in proteogenomics studies.

Sparse Trigger Pattern Guided Deep Learning Model Watermarking

ACM Workshop on Information Hiding and Multimedia Security (ACM IH&MMSec), June 2022

Chun-Shien Lu

Chun-Shien Lu


Watermarking neural networks (NNs) for ownership protection has received considerable attention recently. Resisting both model pruning and fine-tuning is commonly considered to evaluate the robustness of a watermarked NN. However, the rationale behind such a robustness is still relatively unexplored in the literature. In this paper, we study this problem to propose a so-called sparse trigger pattern (STP) guided deep learning model watermarking method. We provide empirical evidence to show that trigger patterns are able to make the distribution of model parameters compact, and thus exhibit interpretable resilience to model pruning and fine-tuning. We find the effect of STP can also be technically interpreted as the first layer dropout. Extensive experiments demonstrate the robustness of our method.

Galectin-1 orchestrates an inflammatory tumor-stroma crosstalk in hepatoma by enhancing TNFR1 protein stability and signaling in carcinoma-associated fibroblasts

Oncogene, April 2022

Yao-Tsung Tsai, Chih-Yi Li, Yen-Hua Huang, Te-Sheng Chang, Chung-Yen Lin, Chia-Hsien Chuang, Chih-Yang Wang, Gangga Anuraga, Tzu-Hao Chang, Tsung-Chieh Shih, Zu-Yau Lin, Yuh-Ling Chen, Ivy Chung, Kuen-Haur Lee, Che-Chang Chang, Shian-Ying Sung, Kai-Huei Yang, Wan-Lin Tsui, Chee-Voon Yap, Ming-Heng Wu

Chun-Yen Lin Chia-Hsien Chuang


Most cases of hepatocellular carcinoma (HCC) arise with the fibrotic microenvironment where hepatic stellate cells (HSCs) and carcinoma-associated fibroblasts (CAFs) are critical components in HCC progression. Therefore, CAF normalization could be a feasible therapy for HCC. Galectin-1 (Gal-1), a β-galactoside-binding lectin, is critical for HSC activation and liver fibrosis. However, few studies has evaluated the pathological role of Gal-1 in HCC stroma and its role in hepatic CAF is unclear. Here we showed that Gal-1 mainly expressed in HCC stroma, but not cancer cells. High expression of Gal-1 is correlated with CAF markers and poor prognoses of HCC patients. In co-culture systems, targeting Gal-1 in CAFs or HSCs, using small hairpin (sh)RNAs or an therapeutic inhibitor (LLS30), downregulated plasminogen activator inhibitor-2 (PAI-2) production which suppressed cancer stem-like cell properties and invasion ability of HCC in a paracrine manner. The Gal-1-targeting effect was mediated by increased a disintegrin and metalloprotease 17 (ADAM17)-dependent TNF-receptor 1 (TNFR1) shedding/cleavage which inhibited the TNF-α → JNK → c-Jun/ATF2 signaling axis of pro-inflammatory gene transcription. Silencing Gal-1 in CAFs inhibited CAF-augmented HCC progression and reprogrammed the CAF-mediated inflammatory responses in a co-injection xenograft model. Taken together, the findings uncover a crucial role of Gal-1 in CAFs that orchestrates an inflammatory CSC niche supporting HCC progression and demonstrate that targeting Gal-1 could be a potential therapy for fibrosis-related HCC.

Efficient Dual Batch Size Deep Learning for Distributed Parameter Server Systems

IEEE Annual Computer Software and Applications Conference (COMPSAC), June 2022

Kuan-Wei Lu, Pangfeng Liu, Ding-Yong Hong and Jan-Jan Wu

Ding-Yong Hong Jan-Jan Wu


Distributed machine learning is essential for applying deep learning models with many data and parameters. Current researches on distributed machine learning focus on using more hardware devices powerful computing units for fast training. Consequently, the model training prefers a larger batch size to accelerate the training speed. However, the large batch training often suffers from poor accuracy due to poor generalization ability. Researchers have come up with many sophisticated methods to address this accuracy issue due to large batch sizes. These methods usually have complex mechanisms, thus making training more dif cult. In addition, powerful training hardware for large batch sizes is expensive, and not all researchers can afford it. We propose dual batch size learning scheme to address the batch size issue. We use the maximum batch size of our hardware for maximum training ef ciency we can afford. In addition, we introduce a smaller batch size during the training to improve the model generalization ability. Using two different batch sizes in the same training simultaneously will reduce the testing loss and obtain a good generalization ability, with only a slight increase in the training time. We implement our dual batch size learning scheme and conduct experiments. By increasing 5% of the training time, we can reduce the loss from 1.429 to 1.246 in some cases. In addition, by appropriately adjusting the percentage of large and small batch sizes, we can increase the accuracy by 2.8% in some cases. With the additional 10% increase in training time, we can reduce the loss from 1.429 to 1.193. And after moderately adjusting the number of large batches and small batches used by GPUs, the accuracy can increase by 2.9%. Distributed machine learning is essential for ap- plying deep learning models with many data and parameters. Current researches on distributed machine learning focus on using more hardware devices powerful computing units for fast training. Consequently, the model training prefers a larger batch size to accelerate the training speed. However, the large batch training often suffers from poor accuracy due to poor generaliza- tion ability. Researchers have come up with many sophisticated methods to address this accuracy issue due to large batch sizes. These methods usually have complex mechanisms, thus making training more dif cult. In addition, powerful training hardware for large batch sizes is expensive, and not all researchers can afford it. We propose a dual batch size learning scheme to address the batch size issue. We use the maximum batch size of our hardware for maximum training ef ciency we can afford. In addition, we introduce a smaller batch size during the training to improve the model generalization ability. Using two different batch sizes in the same training simultaneously will reduce the testing loss and obtain a good generalization ability, with only a slight increase in the training time. We implement our dual batch size learning scheme and conduct experiments. By increasing 5% of the training time, we can reduce the loss from 1.429 to 1.246 in some cases. In addition, by appropriately adjusting the percentage of large and small batch sizes, we can increase the accuracy by 2.8% in some cases. With the additional 10% increase in training time, we can reduce the loss from 1.429 to 1.193. And after moderately adjusting the number of large batches and small batches used by GPUs, the accuracy can increase by 2.9%. Using two different batch sizes in the same training introduces two complications. First, the data processing speeds for two different batch sizes are different, so we must assign the data proportionally to maximize the overall processing speed. In addition, since the smaller batches will see fewer data due to the overall processing speed consideration, we proportionally adjust their contribution towards the global weight update in the parameter server. We use the ratio of data between the small and large batches to adjust the contribution. Experimental results indicate that this contribution adjustment increases the  nal accuracy by another 0.9%.

A metagenomics study of hexabromocyclododecane degradation with a soil microbial community

Journal of Hazardous Materials, May 2022

Yi-Jie Li, Chia-Hsien Chuang, Wen-Chih Cheng, Shu-Hwa Chen, Wen-Ling Chen, Yu-Jie Lin, Chung-Yen Lin, Yang-hsin Shih

Chia-Hsien Chuang Wen-Chih Cheng Shu-Hwa Chen Chun-Yen Lin


Hexabromocyclododecanes (HBCDs) are globally prevalent and persistent organic pollutants (POPs) listed by the Stockholm Convention in 2013. They have been detected in many environmental media from waterbodies to Plantae and even in the human body. Due to their highly bioaccumulative characterization, they pose an urgent public health issue. Here, we demonstrate that the indigenous microbial community in the agricultural soil in Taiwan could decompose HBCDs with no additional carbon source incentive. The degradation kinetics reached 0.173 day-1 after the first treatment and 0.104 day-1 after second exposure. With additional C-sources, the rate constants decreased to 0.054–0.097 day-1. The hydroxylic debromination metabolites and ring cleavage long-chain alkane metabolites were identified to support the potential metabolic pathways utilized by the soil microbial communities. The metagenome established by Nanopore sequencing showed significant compositional alteration in the soil microbial community after the HBCD treatment. After ranking, comparing relative abundances, and performing network analyses, several novel bacterial taxa were identified to contribute to HBCD biotransformation, including Herbaspirillum, Sphingomonas, Brevundimonas, Azospirillum, Caulobacter, and Microvirga, through halogenated / aromatic compound degradation, glutathione-S-transferase, and hydrolase activity. We present a compelling and applicable approach combining metagenomics research, degradation kinetics, and metabolomics strategies, which allowed us to decipher the natural attenuation and remediation mechanisms of HBCDs.

An integrated metadatabase of 16S rRNA gene amplicon for microbiome taxonomic classification

Bioinformatics, March 2022

Chun-Chieh Liao, Po-Ying Fu, Chih-Wei Huang, Chia-Hsien Chuang, Yun Yen, Chung-Yen Lin*, Shu-Hwa Chen*

Chun-Chieh Liao Chih Wei Huang Chia-Hsien Chuang Chun-Yen Lin Shu-Hwa Chen


Motivation: Taxonomic classification of 16S ribosomal RNA gene amplicon is an efficient and economic approach in microbiome analysis. 16S rRNA sequence databases like SILVA, RDP, EzBioCloud, and HOMD used in downstream bioinformatic pipelines have limitations on either the sequence redundancy or the delay on new sequence recruitment. To improve the 16S rRNA gene-based taxonomic classification, we merged these widely used databases and a collection of novel sequences systemically into an integrated resource. Results: MetaSquare version 1.0 is an integrated 16S rRNA sequence database. It is composed of more than six million sequences and improves taxonomic classification resolution on both long-read and short-read methods. Availability: Accessible at and

Peptide-Based Drug Predictions for Cancer Therapy Using Deep Learning

Pharmaceuticals, March 2022

Yih-Yun Sun†, Tzu-Tang Lin†, Wen-Chih Cheng, I-Hsuan Lu, Chung-Yen Lin,* and Shu-Hwa Chen *

Yih-Yun Sun Tzu-Tang Lin Wen-Chih Cheng I-HSUAN Lu Chun-Yen Lin Shu-Hwa Chen


Anticancer peptides (ACPs) are selective and toxic to cancer cells as new anticancer drugs. Identifying new ACPs is time-consuming and expensive to evaluate all candidates’ anticancer abilities. To reduce the cost of ACP drug development, we collected the most updated ACP data to train a convolutional neural network (CNN) with a peptide sequence encoding method for initial in silico evaluation. Here we introduced PC6, a novel protein-encoding method, to convert a peptide sequence into a computational matrix, representing six physicochemical properties of each amino acid. By integrating data, encoding method, and deep learning model, we developed AI4ACP, a user-friendly web-based ACP distinguisher that can predict the anticancer property of query peptides and promote the discovery of peptides with anticancer activity. The experimental results demonstrate that AI4ACP in CNN, trained using the new ACP collection, outperforms the existing ACP predictors. The 5-fold cross-validation of AI4ACP with the new collection also showed that the model could perform at a stable level on high accuracy around 0.89 without overfitting. Using AI4ACP, users can easily accomplish an early-stage evaluation of unknown peptides and select potential candidates to test their anticancer activities quickly.

TagSeq: Malicious Behavior Discovery Using Dynamic Analysis

PLOS ONE, To Appear

Yi-Ting Huang, Yeali Sun and Meng Chang Chen

Yi-Ting Huang Meng-Chang Chen


In recent years, studies on malware analysis have noticeably increased in the cybersecurity community. Most recent studies concentrate on malware classification and detection or malicious patterns identification, but as to malware activity, it still relies heavily on manual analysis for high-level semantic descriptions. We develop a sequence-to-sequence (seq2seq) neural network, called TagSeq, to investigate a sequence of Windows API calls recorded from malware execution, and produce tags to label their malicious behavior. We propose embedding modules to transform Windows API function parameters, registry, filenames, and URLs into low-dimension vectors, while still preserving the closeness property. Moreover, we utilize an attention mechanism to capture the relations between generated tags and certain API invocation calls. Results show that the most possible malicious actions are identified by TagSeq. Examples and a case study demonstrate that the proposed embedding modules preserve semantic-physical relations and that the predicted tags reflect malicious intentions. We believe this work is suitable as a tool to help security analysts recognize malicious behavior and intent with easy-to-understand tags.

Capturing Humans in Motion: Temporal-Attentive 3D Human Pose and Shape Estimation from Monocular Video

IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022, June 2022

Wen-Li Wei, Jen-Chun Lin, Tyng-Luh Liu, and Hong-Yuan Mark Liao

Wen-Li Wei Jen-Chun Lin Tyng-Luh Liu Hong-Yuan Mark Liao


Learning to capture human motion is essential to 3D human pose and shape estimation from monocular video. However, the existing methods mainly rely on recurrent or convolutional operation to model such temporal information, which limits the ability to capture non-local context relations of human motion. To address this problem, we propose a motion pose and shape network (MPS-Net) to effectively capture humans in motion to estimate accurate and temporally coherent 3D human pose and shape from a video. Specifically, we first propose a motion continuity attention (MoCA) module that leverages visual cues observed from human motion to adaptively recalibrate the range that needs attention in the sequence to better capture the motion continuity dependencies. Then, we develop a hierarchical attentive feature integration (HAFI) module to effectively combine adjacent past and future feature representations to strengthen temporal correlation and refine the feature representation of the current frame. By coupling the MoCA and HAFI modules, the proposed MPS-Net excels in estimating 3D human pose and shape in the video. Though conceptually simple, our MPS-Net not only outperforms the state-of-the-art methods on the 3DPW, MPI-INF-3DHP, and Human3.6M benchmark datasets, but also uses fewer network parameters. The video demos can be found at

DPGEN: Differentially Private Generative Energy-Guided Network for High-Resolution Natural Image Synthesis

IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR), June 2022

Jia-Wei Chen, Chia-Mu Yu, Ching-Chia Kao, Tzai-Wei Pang, and Chun-Shien Lu

Jia-Wei Chen Chia-Mu Yu Ching-Chia Kao Tzai-Wei Pang Chun-Shien Lu


Despite an increased demand for valuable data, the privacy concerns associated with sensitive datasets present a barrier to data sharing. One may use differentially private generative models to generate synthetic data. Unfortunately, generators are typically restricted to generating images of low-resolutions due to the limitation of noisy gradients. Here, we propose DPGEN, a network model designed to synthesize high-resolution natural images while satisfying differential privacy. In particular, we propose an energy-guided network trained on sanitized data to indicate the direction of the true data distribution via Langevin Markov chain Monte Carlo (MCMC) sampling method. In contrast to the state-ofthe-art methods that can process only low-resolution images (e.g., MNIST and Fashion-MNIST), DPGEN can generate differentially private synthetic images with resolutions up to 128X128 with superior visual quality and data utility.

Calibr improves spectral library search for spectrum‑centric analysis of data independent acquisition proteomics

Scientific Reports, February 2022

Jen‑Hung Wang, Wai‑Kok Choong, Ching‑Tai Chen*, and Ting‑Yi Sung*

Jen-Hung Wang Wai-Kok Choong Ching-Tai Chen Ting-Yi Sung


Identifying peptides and proteins from mass spectrometry (MS) data, spectral library searching has emerged as a complementary approach to the conventional database searching. However, for the spectrum-centric analysis of data-independent acquisition (DIA) data, spectral library searching has not been widely exploited because existing spectral library search tools are mainly designed and optimized for the analysis of data-dependent acquisition (DDA) data. We present Calibr, a spectral library search tool for spectrum-centric DIA data analysis. Calibr optimizes spectrum preprocessing for pseudo MS2 spectra, generating an 8.11% increase in spectrum-spectrum match (SSM) number and a 7.49% increase in peptide number over the traditional preprocessing approach. When searching against the DDA-based spectral library, Calibr improves SSM number by 17.6%-26.65% and peptide number by 18.45%-37.31% over two state-of-the-art tools on three different data sets. Searching against the public spectral library from MassIVE, Calibr improves state-of-the-art tools in SSM and peptide numbers by more than 31.49% and 25.24%, respectively, for two data sets. Our analyses indicate  higher sensitivity of Calibr results from the use of various spectral similarity measures and statistical scores, coupled with machine learning-based statistical validation for FDR control. Calibr executable files including a graphical user-interface application are available at and

SVSNet: An End-to-end Speaker Voice Similarity Assessment Model

IEEE Signal Processing Letters, 2022

Cheng-Hung Hu, Yu-Huai Peng, Junichi Yamagishi, Yu Tsao, and Hsin-Min Wang

Hsin-Min Wang


Neural evaluation metrics derived for numerous speech generation tasks have recently attracted great attention. In this paper, we propose SVSNet, the first end-to-end neural network model to assess the speaker voice similarity between converted speech and natural speech for voice conversion tasks. Unlike most neural evaluation metrics that use hand-crafted features, SVSNet directly takes the raw waveform as input to more completely utilize speech information for prediction. SVSNet consists of encoder, co-attention, distance calculation, and prediction modules and is trained in an end-to-end manner. The experimental results on the Voice Conversion Challenge 2018 and 2020 (VCC2018 and VCC2020) datasets show that SVSNet outperforms well-known baseline systems in the assessment of speaker similarity at the utterance and system levels.

Improved Lite Audio-Visual Speech Enhancement

IEEE/ACM Transactions on Audio, Speech, and Language Processing, 2022

Shang-Yi Chuang, Hsin-Min Wang, and Yu Tsao

Hsin-Min Wang


Numerous studies have investigated the effectiveness of audio-visual multimodal learning for speech enhancement (AVSE) tasks, seeking a solution that uses visual data as auxiliary and complementary input to reduce the noise of noisy speech signals. Recently, we proposed a lite audio-visual speech enhancement (LAVSE) algorithm for a car-driving scenario. Compared to conventional AVSE systems, LAVSE requires less online computation and to some extent solves the user privacy problem on facial data. In this study, we extend LAVSE to improve its ability to address three practical issues often encountered in implementing AVSE systems, namely, the additional cost of processing visual data, audio-visual asynchronization, and low-quality visual data. The proposed system is termed improved LAVSE (iLAVSE), which uses a convolutional recurrent neural network architecture as the core AVSE model. We evaluate iLAVSE on the Taiwan Mandarin speech with video dataset. Experimental results confirm that compared to conventional AVSE systems, iLAVSE can effectively overcome the aforementioned three practical issues and can improve enhancement performance. The results also confirm that iLAVSE is suitable for real-world scenarios, where high-quality audio-visual sensors may not always be available.

Semantics-Preserved Data Augmentation for Aspect-Based Sentiment Analysis

Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, November 2021

Ting-Wei Hsu, Chung-Chi Chen, Hen-Hsen Huang, and Hsin-Hsi Chen

Hen-Hsen Huang


Both the issues of data deficiencies and semantic consistency are important for data augmentation. Most of previous methods address the first issue, but ignore the second one. In the cases of aspect-based sentiment analysis, violation of the above issues may change the aspect and sentiment polarity. In this paper, we propose a semantics-preservation data augmentation approach by considering the importance of each word in a textual sequence according to the related aspects and sentiments. We then substitute the unimportant tokens with two replacement strategies without altering the aspect-level polarity. Our approach is evaluated on several publicly available sentiment analysis datasets and the real-world stock price/risk movement prediction scenarios. Experimental results show that our methodology achieves better performances in all datasets.

A Yes-Associated Protein (YAP) and Insulin-like Growth Factor 1 Receptor (IGF-1R) Signaling Loop Is Involved in Sorafenib Resistance in Hepatocellular Carcinoma

Cancers, July 2021

Mai-Huong T. Ngo, Sue-Wei Peng, Yung-Che Kuo, Chun-Yen Lin, Ming-Heng Wu, Chia-Hsien Chuang, Cheng-Xiang Kao, Han-Yin Jeng, Gee-Way Lin, Thai-Yen Ling, Te-Sheng Chang *, Yen-Hua Huang *

Chun-Yen Lin Chia-Hsien Chuang


The role of a YAP-IGF-1R signaling loop in HCC resistance to sorafenib remains unknown. Method: Sorafenib-resistant cells were generated by treating naïve cells (HepG2215 and Hep3B) with sorafenib. Different cancer cell lines from databases were analyzed through the ONCOMINE web server. BIOSTORM_LIHC patient tissues (46 nonresponders and 21 responders to sorafenib) were used to compare YAP mRNA levels. The HepG2215_R-derived xenograft in SCID mice was used as an in vivo model. HCC tissues from a patient with sorafenib failure were used to examine differences in YAP and IGF-R signaling. Results: Positive associations exist among the levels of YAP, IGF-1R, and EMT markers in HCC tissues and the levels of these proteins increased with sorafenib failure, with a trend of tumor-margin distribution in vivo. Blocking YAP downregulated IGF-1R signaling-related proteins, while IGF-1/2 treatment enhanced the nuclear translocation of YAP in HCC cells through PI3K-mTOR regulation. The combination of YAP-specific inhibitor verteporfin (VP) and sorafenib effectively decreased cell viability in a synergistic manner, evidenced by the combination index (CI). Conclusion: A YAP-IGF-1R signaling loop may play a role in HCC sorafenib resistance and could provide novel potential targets for combination therapy with sorafenib to overcome drug resistance in HCC.

Development of Disease-Resistance-Associated Microsatellite DNA Markers for Selective Breeding of Tilapia (Oreochromis spp.) Farmed in Taiwan

Genes, January 2022

Chen, Che-Chun, Chang-Wen Huang, Chung-Yen Lin, Chia-Hui Ho, Hong N. Pham, Te-Hua Hsu, Tzu-Tang Lin, Rong-Hwa Chen, Shuenn-Der Yang, Chin-I. Chang, and Hong-Yi Gong.

Chun-Yen Lin Tzu-Tang Lin Shu-Hwa Chen


There are numerous means to improve the tilapia aquaculture industry, and one is to develop disease resistance through selective breeding using molecular markers. In this study, 11 disease-resistance-associated microsatellite markers including 3 markers linked to hamp2, 4 linked to hamp1, 1 linked to pgrn2, 2 linked to pgrn1, and 1 linked to piscidin 4 (TP4) genes were established for tilapia strains farmed in Taiwan after challenge with Streptococcus inae. The correlation analysis of genotypes and survival revealed a total of 55 genotypes related to survival by the chi-square and Z-test. Although fewer markers were found in B and N2 strains compared with A strain, they performed well in terms of disease resistance. It suggested that this may be due to the low potency of some genotypes and the combinatorial arrangement between them. Therefore, a predictive model was built by the genotypes of the parental generation and the mortality rate of different combinations was calculated. The results show the same trend of predicted mortality in the offspring of three new disease-resistant strains as in the challenge experiment. The present findings is a nonkilling method without requiring the selection by challenge with bacteria or viruses and might increase the possibility of utilization of selective breeding using SSR markers in farms.

Identification of Genes Related to Cold Tolerance and Novel Genetic Markers for Molecular Breeding in Taiwan Tilapia (Oreochromis spp.) via Transcriptome Analysis

Animals, December 2021

Chu, Pei-Yun, Jia-Xian Li, Te-Hua Hsu, Hong-Yi Gong, Chung-Yen Lin, Jung-Hua Wang, and Chang-Wen Huang.

Chun-Yen Lin


In this study, we investigated the brain, gill, liver, and muscle transcriptomic responses of Taiwan tilapia towards cold stress. Some key genes and molecular markers involved in cold biological pathways were screened through differential expression. Among them, energy-related metabolic pathways and nucleotide genotypes were highly correlated with cold tolerance traits. This suggested that single nucleotide polymorphism (SNP) genetic variation can be used as a molecular marker to assist the selection and verification of cold-tolerant populations. Our study results will accelerate the understanding of different farmed tilapia tolerance mechanisms to environmental temperature changes and provide insights for the molecular breeding of cold-tolerant Taiwan tilapia species.

Termite pest identification method based on deep convolution neural networks

Journal of Economic Entomology, August 2021

Huang, J.H., Liu, Y.T., Ni, H.C., Chen, B.Y., Huang, S.Y., Tsai, H.K.*, and Li, H.F.*

Huai-Kuang Tsai


Several species of drywood termites, subterranean termites, and fungus-growing termites cause extensive economic losses annually worldwide. Because no universal method is available for controlling all termites, correct species identification is crucial for termite management. Despite deep neural network technologies’ promising performance in pest recognition, a method for automatic termite recognition remains lacking. To develop an automated deep learning classifier for termite image recognition suitable for mobile applications, we used smartphones to acquire 18,000 original images each of four termite pest species: Kalotermitidae: Cryptotermes domesticus (Haviland); Rhinotermitidae: Coptotermes formosanus Shiraki and Reticulitermes flaviceps (Oshima); and Termitidae: Odontotermes formosanus (Shiraki). Each original image included multiple individuals, and we applied five image segmentation techniques for capturing individual termites. We used 24,000 individual-termite images (4 species   2 castes   3 groups   1,000 images) for model development and testing. We implemented a termite classification system by using a deep learning–based model, MobileNetV2. Our models achieved high accuracy scores of 0.947, 0.946, and 0.929 for identifying soldiers, workers, and both castes, respectively, which is not significantly different from human expert performance. We further applied image augmentation techniques, including geometrical transformations and intensity transformations, to individual-termite images. The results revealed that the same classification accuracy can be achieved by using 1,000 augmented images derived from only 200 individual-termite images, thus facilitating further model development on the basis of many fewer original images. Our image-based identification system can enable the selection of termite control tools for pest management professionals or homeowners.

The degradation mechanisms of Rhodopseudomonas palustris toward hexabromocyclododecane by time-course transcriptome analysis

Chemical Engineering Journal, December 2021

Yi-Jie Li, Reuben Wang, Chung-Yen Lin, Shu-Hwa Chen, Chia-Hsien Chuang, Tzu-Ho Chou, Chi-Fang Ko, Pei-Hsin Chou, Chi-Te Liu, Yang-hsin Shih

Chun-Yen Lin Shu-Hwa Chen Chia-Hsien Chuang


Hexabromocyclododecane (HBCD) is one of the most frequently used brominated flame retardants (BFRs). However, the HBCD degradation method's development has become vital because it readily bioaccumulates and is persistent in the environment. A previous study showed Rhodopseudomonas palustris degrades HBCD through several possible metabolic pathways based on transcriptomic analysis of compared samples. This study introduces multiple time-course transcriptomic analysis approaches to identify the specific HBCD metabolic pathway in R. palustris inbubated at different temperatures. The transcriptome profiles revealed that the addition of HBCD triggered 126 transcripts in cells at 25°C and 35°C. Further KEGG analysis showed several HBCD induced metabolic pathways, including ABC transporter, butanoate metabolism, dephosphorylation, lipid glycosylation pathways, etc. The principal component analysis further provides evidence of genes directly affected by HBCD. The increased expression level of transcriptional regulator LysR, two-component system regulators, HBCD degradation enzymes, including haloacid dehalogenases, glutathione-S-transferase, cytochrome p450, hydrolases, and dioxygenases in R. palustris were confirmed by qRT-PCR analysis. Combining the transcriptomic profiles and gene expression level analysis, we proposed the HBCD metabolic pathway in R. palustris. Briefly, HBCD signal transferred from cell membrane to transcriptional regulator LysR, then further to downstream degradation working enzymes. Overall, our results highlight the value of systematic transcriptomic approaches to discover and elucidate the intrinsic microbial metabolisms for HBCD degradation in R. palustris. The results of this study provide a novel perspective on the degradation of persistent organic pollutants (POPs) such as HBCD using a bio-omics approach.

Beyond Write-reduction Consideration: A Wear-leveling-enabled B+-tree Indexing Scheme over an NVRAM-based Architecture

IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems (TCAD), December 2021

Dharamjeet, Tseng-Yi Chen, Yuan-Hao Chang, Chun-Feng Wu, Chi-Heng Lee, and Wei-Kuan Shih

Tseng-Yi Chen Yuan-Hao Chang Chun-Feng Wu


Recently, non-volatile random-access memory (NVRAM) has been regarded as the most up-and-coming main memory technology in embedded and Internet-of-Thing (IoT) system due to its attractive features: Zero-static power consumption and high memory cell density. However, the endurance issue as a “nightmare” always haunts NVRAM system developers. Worse still, NVRAM’s lifespan will wear out soon in embedded applications because their data management systems usually utilize an indexing scheme to maintain small data. Plus, a node structure within the indexing scheme will be frequently updated because of data creation and deletion. Therefore, many previous works rethink B+-tree indexing scheme on an NVRAM-based system. The most previous studies focused on reducing the amount of write traffic to memory. Unfortunately, they are failed to extend NVRAM lifespan because their solution cannot evenly distribute the amount of write traffic to each memory cell. Additionally, prior solutions have not considered that all nodes within B+-tree indexing structure have different update frequencies. Based on such the observation, this work proposes a wear-leveling-aware B+-tree design, namely waB+-tree, to consider the update frequency of each node within the B+-tree structure, so as to evenly scatter the amount of write traffic to the NVRAM cells. According to our experiments, the proposed waB+-tree shows the encouraging results of endurance improvement.

NK cell receptor and ligand composition influences the clearance of SARS-CoV-2

The Journal of Clinical Investigation, October 2021

Hsieh, W.C., Lai, E.Y., Liu, Y.T., Wang, Y.F., Tzeng, Y.S., Cui, L., Lai, Y.J., Huang, H.C., Huang, J.H., Ni, H.C., Tsai, D.Y., Liang, J.J., Liao, C.C., Lu, Y.T., Jiang, L., Liu, M.T., Wang, J.T., Chang, S.Y., Chen, C.Y., Tsai, H.C., Chang, Y.M., Wernig, G., Li, C.W., Lin, K.I., Lin, Y.L., Tsai, H.K., Huang, Y.T., Chen, S.Y.

Chun-Chieh Liao Jen‑Hung Wang Huai-Kuang Tsai Yi-Ting Huang


To explore how the immune system controls clearance of SARS-CoV-2, we used a single-cell, mass cytometry–based proteomics platform to profile the immune systems of 21 patients who had recovered from SARS-CoV-2 infection without need for admission to an intensive care unit or for mechanical ventilation. We focused on receptors involved in interactions between immune cells and virus-infected cells. We found that the diversity of receptor repertoires on natural killer (NK) cells was negatively correlated with the viral clearance rate. In addition, NK subsets expressing the receptor DNAM1 were increased in patients who more rapidly recovered from infection. Ex vivo functional studies revealed that NK subpopulations with high DNAM1 expression had cytolytic activities in response to target cell stimulation. We also found that SARS-CoV-2 infection induced the expression of CD155 and nectin-4, ligands of DNAM1 and its paired coinhibitory receptor TIGIT, which counterbalanced the cytolytic activities of NK cells. Collectively, our results link the cytolytic immune responses of NK cells to the clearance of SARS-CoV-2 and show that the DNAM1 pathway modulates host-pathogen interactions during SARS-CoV-2 infection.

Investigating the Transcriptomic and Expression Presence-Absence Variation Exist in Japanese Eel (Anguilla japonica), a Primitive Teleost

Marine Biotechnology, October 2021

Yung-Sen Huang, Chung-Yen Lin, Wen-Chih Cheng

Chun-Yen Lin Wen-Chih Cheng


The pan-genome was defined as the complete gene set across strains, and it is built upon genes displaying presence-absence variations (PAVs); the pan-transcriptome is defined by recalling the pan-genome. Indeed, a PAV is reflected from the expression presence-absence variation (ePAV). In this study, treated with androgen, eels, which are a primitive fish from the basal lineage of Teleost, with different ovarian developments were chosen and submitted to RAN-sequencing. Transcriptomes were the assembly against eel genome scaffolds; a pair was the unit (the same eel before and after treatment) to analyze DEGs (differentially expressed genes); the core, unique, or accessory genes were identified, and the list of DEGs was analyzed to investigate ePAV. The results suggest that there was ePAV in Japanese eel, and the ePAV of eel was analyzed by pathway enrichment. These results signify the importance of genetic differential expression on the variations of phenotypes by androgen, and a transcriptomic approach appears to enable extracting multiple layers of genomic data.

Pose Adaptive Dual Mixup for Few-Shot Single-View 3D Reconstruction

Thirty-Sixth AAAI Conference on Artificial Intelligence (AAAI), February 2022

Ta-Ying Cheng, Hsuan-ru Yang, Niki Trigoni, Hwann-Tzong Chen and Tyng-Luh Liu

Tyng-Luh Liu


We present a pose adaptive few-shot learning procedure and a two-stage data interpolation regularization, termed Pose Adaptive Dual Mixup (PADMix), for single-image 3D reconstruction. While augmentations via interpolating feature-label pairs are effective in classification tasks, they fall short in shape predictions potentially due to inconsistencies between interpolated products of two images and volumes when rendering viewpoints are unknown. PADMix targets this issue with two sets of mixup procedures performed sequentially. We first perform an input mixup which, combined with a pose adaptive learning procedure, is helpful in learning 2D feature extraction and pose adaptive latent encoding. The stagewise training allows us to build upon the pose invariant representations to perform a follow-up latent mixup under one-to-one correspondences between features and ground-truth volumes. PADMix significantly outperforms previous literature on few-shot settings over the ShapeNet dataset and sets new benchmarks on the more challenging real-world Pix3D dataset.

Optimal Branch Location Finding for Cost effective Inference on Branchynet

IEEE International Conference on Big Data (top conference), the Machine Learning with Big Data track, December 2021

Chang-Han Chiang, Pangfeng Liu, Da-Wei Wang, Ding-Yong Hong, and Jan-Jan Wu

Da-Wei Wang Ding-Yong Hong Jan-Jan Wu


Deep Neural Networks (DNNs) are very popular in many machine learning domains. To achieve higher accuracy, DNNs have become deeper and larger. However, the improvement in accuracy comes with the price of the longer inference time and energy consumption. The marginal cost to increase a unit of accuracy has become higher as the accuracy itself is rising. The Branchynet, known as early exits, is an architecture to address increasing marginal cost for improving accuracy. The Branchynet adds extra side classifiers to a DNN model. The inference on a significant portion of the samples can exit from the network earlier via these side branches if they already have high confidence in the results. The Branchynet requires manually tuning the learning hyperparameters, e.g., the locations of branches and the confidence threshold for early exiting. The effectiveness of this manual tuning dramatically impacts the efficiency of the tuned networks. To the best of our knowledge, there are no efficient algorithms to find the best branch location, which is a trade-off between the accuracy and inference time on the Branchynet. We propose an algorithm to find the optimal branch locations for the Branchynet. We formulate the problem of finding the optimal branch location for the branchynet as an optimization problem, and prove that the branch placement problem is an NPcomplete problem. We then derive dynamic programming that runs in pseudo-polynomial time and solves the branch placement problem optimally. We also implement our algorithm and solve the branch placement problems on four types of VGG networks. The experiment results indicate that our dynamic programming can find the optimal branch locations for generating the maximum number of correct classifications within a given time budget. We also run the four VGG models on a GeForce RTX-3090 GPU with the branch combination found by the dynamic programming. The experiment results show that our dynamic programming accurately predicts the number of correct classifications and the execution time on the GPU.

Connecting MHC-I-binding motifs with HLA alleles via deep learning

Communications Biology, October 2021

Lee, K.H., Chang, Y.C., Chen, T.F., Juan, H.F., Tsai, H.K., Chen, C.Y.

Yung-Chun Chang Huai-Kuang Tsai


The selection of peptides presented by MHC molecules is crucial for antigen discovery. Previously, several predictors have shown impressive performance on binding affinity. However, the decisive MHC residues and their relation to the selection of binding peptides are still unrevealed. Here, we connected HLA alleles with binding motifs via our deep learning-based framework, MHCfovea. MHCfovea expanded the knowledge of MHC-I binding motifs from 150 to 13,008 alleles. After clustering N-terminal and C-terminal submotifs on both observed and unobserved alleles, MHCfovea calculated the hyper-motifs and the corresponding allele signatures on the important positions to disclose the relation between binding motifs and MHC-I sequences. MHCfovea delivered 32 pairs of hyper-motifs and allele signatures (HLA-A: 13, HLA-B: 12, and HLA-C: 7). The paired hyper-motifs and allele signatures disclosed the critical polymorphic residues that determine the binding preference, which are believed to be valuable for antigen discovery and vaccine design when allele specificity is concerned.