Accurate Prediction of Cancer Prognosis by Exploiting Patient-Specific Cancer Driver Genes.

Accurate prediction of the prognoses of cancer patients and identification of prognostic biomarkers are both important for the improved treatment of cancer patients, in addition to enhanced anticancer drugs. Many previous bioinformatic studies have been carried out to achieve this goal; however, there remains room for improvement in terms of accuracy. In this study, we demonstrated that patient-specific cancer driver genes could be used to predict cancer prognoses more accurately. To identify patient-specific cancer driver genes, we first generated patient-specific gene networks before using modified PageRank to generate feature vectors that represented the impacts genes had on the patient-specific gene network. Subsequently, the feature vectors of the good and poor prognosis groups were used to train the deep feedforward network. For the 11 cancer types in the TCGA data, the proposed method showed a significantly better prediction performance than the existing state-of-the-art methods for three cancer types (BRCA, CESC and PAAD), better performance for five cancer types (COAD, ESCA, HNSC, KIRC and STAD), and a similar or slightly worse performance for the remaining three cancer types (BLCA, LIHC and LUAD). Furthermore, the case study for the identified breast cancer and cervical squamous cell carcinoma prognostic genes and their subnetworks included several pathways associated with the progression of breast cancer and cervical squamous cell carcinoma. These results suggested that heterogeneous cancer driver information may be associated with cancer prognosis.

Cell type-restricted activity of hnRNPM promotes breast cancer metastasis via regulating alternative splicing.

Tumor metastasis remains the major cause of cancer-related death, but its molecular basis is still not well understood. Here we uncovered a splicing-mediated pathway that is essential for breast cancer metastasis. We show that the RNA-binding protein heterogeneous nuclear ribonucleoprotein M (hnRNPM) promotes breast cancer metastasis by activating the switch of alternative splicing that occurs during epithelial-mesenchymal transition (EMT). Genome-wide deep sequencing analysis suggests that hnRNPM potentiates TGFß signaling and identifies CD44 as a key downstream target of hnRNPM. hnRNPM ablation prevents TGFß-induced EMT and inhibits breast cancer metastasis in mice, whereas enforced expression of the specific CD44 standard (CD44s) splice isoform overrides the loss of hnRNPM and permits EMT and metastasis. Mechanistically, we demonstrate that the ubiquitously expressed hnRNPM acts in a mesenchymal-specific manner to precisely control CD44 splice isoform switching during EMT. This restricted cell-type activity of hnRNPM is achieved by competition with ESRP1, an epithelial splicing regulator that binds to the same cis-regulatory RNA elements as hnRNPM and is repressed during EMT. Importantly, hnRNPM is associated with aggressive breast cancer and correlates with increased CD44s in patient specimens. These findings demonstrate a novel molecular mechanism through which tumor metastasis is endowed by the hnRNPM-mediated splicing program.

Binding affinity prediction for protein-ligand complex using deep attention mechanism based on intermolecular interactions.

BACKGROUND: Accurate prediction of protein-ligand binding affinity is important for lowering the overall cost of drug discovery in structure-based drug design. For accurate predictions, many classical scoring functions and machine learning-based methods have been developed. However, these techniques tend to have limitations, mainly resulting from a lack of sufficient energy terms to describe the complex interactions between proteins and ligands. Recent deep-learning techniques can potentially solve this problem. However, the search for more efficient and appropriate deep-learning architectures and methods to represent protein-ligand complex is ongoing. RESULTS: In this study, we proposed a deep-neural network model to improve the prediction accuracy of protein-ligand complex binding affinity. The proposed model has two important features, descriptor embeddings with information on the local structures of a protein-ligand complex and an attention mechanism to highlight important descriptors for binding affinity prediction. The proposed model performed better than existing binding affinity prediction models on most benchmark datasets. CONCLUSIONS: We confirmed that an attention mechanism can capture the binding sites in a protein-ligand complex to improve prediction performance. Our code is available at https://github.com/Blue1993/BAPA .

Coregulation of alternative splicing by hnRNPM and ESRP1 during EMT.

The epithelial-mesenchymal transition (EMT) is a fundamental developmental process that is abnormally activated in cancer metastasis. Dynamic changes in alternative splicing occur during EMT. ESRP1 and hnRNPM are splicing regulators that promote an epithelial splicing program and a mesenchymal splicing program, respectively. The functional relationships between these splicing factors in the genome scale remain elusive. Comparing alternative splicing targets of hnRNPM and ESRP1 revealed that they coregulate a set of cassette exon events, with the majority showing discordant splicing regulation. Discordant splicing events regulated by hnRNPM show a positive correlation with splicing during EMT; however, concordant events do not, indicating the role of hnRNPM in regulating alternative splicing during EMT is more complex than previously understood. Motif enrichment analysis near hnRNPM-ESRP1 coregulated exons identifies guanine-uridine rich motifs downstream from hnRNPM-repressed and ESRP1-enhanced exons, supporting a general model of competitive binding to these cis-elements to antagonize alternative splicing. The set of coregulated exons are enriched in genes associated with cell migration and cytoskeletal reorganization, which are pathways associated with EMT. Splicing levels of coregulated exons are associated with breast cancer patient survival and correlate with gene sets involved in EMT and breast cancer subtyping. This study identifies complex modes of interaction between hnRNPM and ESRP1 in regulation of splicing in disease-relevant contexts.

Novel deep learning model for more accurate prediction of drug-drug interaction effects.

BACKGROUND: Predicting the effect of drug-drug interactions (DDIs) precisely is important for safer and more effective drug co-prescription. Many computational approaches to predict the effect of DDIs have been proposed, with the aim of reducing the effort of identifying these interactions in vivo or in vitro, but room remains for improvement in prediction performance. RESULTS: In this study, we propose a novel deep learning model to predict the effect of DDIs more accurately.. The proposed model uses autoencoders and a deep feed-forward network that are trained using the structural similarity profiles (SSP), Gene Ontology (GO) term similarity profiles (GSP), and target gene similarity profiles (TSP) of known drug pairs to predict the pharmacological effects of DDIs. The results show that GSP and TSP increase the prediction accuracy when using SSP alone, and the autoencoder is more effective than PCA for reducing the dimensions of each profile. Our model showed better performance than the existing methods, and identified a number of novel DDIs that are supported by medical databases or existing research. CONCLUSIONS: We present a novel deep learning model for more accurate prediction of DDIs and their effects, which may assist in future research to discover novel DDIs and their pharmacological effects.

Improved prediction of breast cancer outcome by identifying heterogeneous biomarkers.

MOTIVATION: Identification of genes that can be used to predict prognosis in patients with cancer is important in that it can lead to improved therapy, and can also promote our understanding of tumor progression on the molecular level. One of the common but fundamental problems that render identification of prognostic genes and prediction of cancer outcomes difficult is the heterogeneity of patient samples. RESULTS: To reduce the effect of sample heterogeneity, we clustered data samples using K-means algorithm and applied modified PageRank to functional interaction (FI) networks weighted using gene expression values of samples in each cluster. Hub genes among resulting prioritized genes were selected as biomarkers to predict the prognosis of samples. This process outperformed traditional feature selection methods as well as several network-based prognostic gene selection methods when applied to Random Forest. We were able to find many cluster-specific prognostic genes for each dataset. Functional study showed that distinct biological processes were enriched in each cluster, which seems to reflect different aspect of tumor progression or oncogenesis among distinct patient groups. Taken together, these results provide support for the hypothesis that our approach can effectively identify heterogeneous prognostic genes, and these are complementary to each other, improving prediction accuracy. AVAILABILITY AND IMPLEMENTATION: https://github.com/mathcom/CPR. CONTACT: jgahn@inu.ac.kr. SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

PISTON: Predicting drug indications and side effects using topic modeling and natural language processing.

The process of discovering novel drugs to treat diseases requires a long time and high cost. It is important to understand side effects of drugs as well as their therapeutic effects, because these can seriously damage the patients due to unexpected actions of the derived candidate drugs. In order to overcome these limitations, computational methods for predicting the therapeutic effects and side effects have been proposed. In particular, text mining is a widely used technique in the field of systems biology, because it can discover hidden relationships between drugs, genes and diseases from a large amount of literature data. Compared with in vivo/in vitro experiments, text mining derives meaningful results with less time and cost. In this study, we propose an algorithm for predicting novel drug-phenotype associations and drug-side effect associations using topic modeling and natural language processing (NLP). We extract sentences in which drugs and genes co-occur from the abstracts of the literature and identify words that describe the relationship between them using NLP. Considering the characteristics of the identified words, we determine if the drug has an up-regulation effect or a down-regulation effect on the gene. Based on genes that affect drugs and their regulatory relationships, we group the frequently occurring genes and regulatory relationships into topics, and build a drug-topic probability matrix by calculating the score that the drug will have a topic using topic modeling. Using the matrix, a classifier is constructed for predicting the novel indications and side effects of drugs considering the characteristics of known drug-phenotype associations or drug-side effect associations. The proposed method predicts both indications and side effects with a single algorithm, and it can exclude drugs with serious side effects or side effects that patients do not want to experience from among the candidate drugs provided for the treatment of the phenotype. Furthermore, lists of novel candidate drugs for phenotypes and side effects can be continuously updated with our algorithm every time a document is added. More than a thousand documents are produced per day, and it is possible for our algorithm to efficiently derive candidate drugs because it requires less cost than the existing drug repositioning methods. The resource of PISTON is available at databio.gachon.ac.kr/tools/PISTON.

Drug voyager: a computational platform for exploring unintended drug action.

BACKGROUND: The dominant paradigm in understanding drug action focuses on the intended therapeutic effects and frequent adverse reactions. However, this approach may limit opportunities to grasp unintended drug actions, which can open up channels to repurpose existing drugs and identify rare adverse drug reactions. Advances in systems biology can be exploited to comprehensively understand pharmacodynamic actions, although proper frameworks to represent drug actions are still lacking. RESULTS: We suggest a novel platform to construct a drug-specific pathway in which a molecular-level mechanism of action is formulated based on pharmacologic, pharmacogenomic, transcriptomic, and phenotypic data related to drug response ( http://databio.gachon.ac.kr/tools/ ). In this platform, an adoption of three conceptual levels imitating drug perturbation allows these pathways to be realistically rendered in comparison to those of other models. Furthermore, we propose a new method that exploits functional features of the drug-specific pathways to predict new indications as well as adverse reactions. For therapeutic uses, our predictions significantly overlapped with clinical trials and an up-to-date drug-disease association database. Also, our method outperforms existing methods with regard to classification of active compounds for cancers. For adverse reactions, our predictions were significantly enriched in an independent database derived from the Food and Drug Administration (FDA) Adverse Event Reporting System and meaningfully cover an Adverse Reaction Database provided by Health Canada. Lastly, we discuss several predictions for both therapeutic indications and side-effects through the published literature. CONCLUSIONS: Our study addresses how we can computationally represent drug-signaling pathways to understand unintended drug actions and to facilitate drug discovery and screening.

RASER: reads aligner for SNPs and editing sites of RNA.

MOTIVATION: Accurate identification of genetic variants such as single-nucleotide polymorphisms (SNPs) or RNA editing sites from RNA-Seq reads is important, yet challenging, because it necessitates a very low false-positive rate in read mapping. Although many read aligners are available, no single aligner was specifically developed or tested as an effective tool for SNP and RNA editing prediction. RESULTS: We present RASER, an accurate read aligner with novel mapping schemes and index tree structure that aims to reduce false-positive mappings due to existence of highly similar regions. We demonstrate that RASER shows the best mapping accuracy compared with other popular algorithms and highest sensitivity in identifying multiply mapped reads. As a result, RASER displays superb efficacy in unbiased mapping of the alternative alleles of SNPs and in identification of RNA editing sites. AVAILABILITY AND IMPLEMENTATION: RASER is written in C++ and freely available for download at https://github.com/jaegyoonahn/RASER.

Identification of functional CNV region networks using a CNV-gene mapping algorithm in a genome-wide scale.

MOTIVATION: Identifying functional relation of copy number variation regions (CNVRs) and gene is an essential process in understanding the impact of genotypic variations on phenotype. There have been many related works, but only a few attempts were made to normal populations. RESULTS: To analyze the functions of genome-wide CNVRs, we applied a novel correlation measure called Correlation based on Sample Set (CSS) to paired Whole Genome TilePath array and messenger RNA (mRNA) microarray data from 210 HapMap individuals with normal phenotypes and calculated the confident CNVR-gene relationships. Two CNVR nodes form an edge if they regulate a common set of genes, allowing the construction of a global CNVR network. We performed functional enrichment on the common genes that were trans-regulated from CNVRs clustered together in our CNVR network. As a result, we observed that most of CNVR clusters in our CNVR network were reported to be involved in some biological processes or cellular functions, while most CNVR clusters from randomly constructed CNVR networks showed no evidence of functional enrichment. Those results imply that CSS is capable of finding related CNVR-gene pairs and CNVR networks that have functional significance. AVAILABILITY: http://embio.yonsei.ac.kr/~ Park/cnv_net.php. CONTACT: sanghyun@cs.yonsei.ac.kr SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

Improved method for protein complex detection using bottleneck proteins.

BACKGROUND: Detecting protein complexes is one of essential and fundamental tasks in understanding various biological functions or processes. Therefore accurate identification of protein complexes is indispensable. METHODS: For more accurate detection of protein complexes, we propose an algorithm which detects dense protein sub-networks of which proteins share closely located bottleneck proteins. The proposed algorithm is capable of finding protein complexes which allow overlapping with each other. RESULTS: We applied our algorithm to several PPI (Protein-Protein Interaction) networks of Saccharomyces cerevisiae and Homo sapiens, and validated our results using public databases of protein complexes. The prediction accuracy was even more improved over our previous work which used also bottleneck information of the PPI network, but showed limitation when predicting small-sized protein complex detection. CONCLUSIONS: Our algorithm resulted in overlapping protein complexes with significantly improved F1 score over existing algorithms. This result comes from high recall due to effective network search, as well as high precision due to proper use of bottleneck information during the network search.

Higher Number of Tumor-Infiltrating PD-L1+ Cells Is Related to Better Response to Multikinase Inhibitors in Hepatocellular Carcinoma.

Multikinase inhibitors (MKIs) such as sorafenib and lenvatinib are first-line treatments for unresectable hepatocellular carcinoma (HCC) and are known to have immunomodulatory effects. However, predictive biomarkers of MKI treatment in HCC patients need to be elucidated. In the present study, thirty consecutive HCC patients receiving lenvatinib (n = 22) and sorafenib (n = 8) who underwent core-needle biopsy before treatment were enrolled. The associations of CD3, CD68, and programmed cell death-ligand-1 (PD-L1) immunohistochemistry with patient outcomes, including overall survival (OS), progression-free survival (PFS), and objective response rate (ORR), were evaluated. High and low subgroups were determined according to median CD3, CD68, and PD-L1 values. Median CD3 and CD68 counts were 51.0 and 46.0 per 20,000 µm2, respectively. The median combined positivity score (CPS) of PD-L1 was 2.0. Median OS and PFS were 17.6 and 4.4 months, respectively. ORRs of the total, lenvatinib, and sorafenib groups were 33.3% (10/30), 12.5% (1/8), and 40.9% (9/22), respectively. The high CD68+ group had significantly better PFS than the low CD68+ group. The high PD-L1 group had better PFS than the low subgroup. When we analyzed the lenvatinib subgroup, PFS was also significantly better in the high CD68+ and PD-L1 groups. These findings suggest that high numbers of PD-L1-expressing cells within tumor tissue prior to MKI treatment can serve as a biomarker to predict favorable PFS in HCC patients.

Comparative Analysis of Atezolizumab Plus Bevacizumab and Hepatic Artery Infusion Chemotherapy in Unresectable Hepatocellular Carcinoma: A Multicenter, Propensity Score Study.

This study aimed to compare the prognosis and characteristics of patients with advanced hepatocellular carcinoma treated with first-line atezolizumab plus bevacizumab (AB) combination therapy and hepatic artery infusion chemotherapy (HAIC). We retrospectively assessed 193 and 114 patients treated with HAIC and AB combination therapy, respectively, between January 2018 and May 2023. The progression-free survival (PFS) of patients treated with AB combination therapy was significantly superior to that of patients treated with HAIC (p < 0.05), but there was no significant difference in overall survival (OS). After propensity score matching, our data revealed no significant differences in OS and PFS between patients who received AB combination therapy and those who received HAIC therapy (p = 0.5617 and 0.3522, respectively). In conclusion, our propensity score study reveals no significant differences in OS and PFS between patients treated with AB combination therapy and those treated with HAIC.

Integrative gene network construction for predicting a set of complementary prostate cancer genes.

MOTIVATION: Diagnosis and prognosis of cancer and understanding oncogenesis within the context of biological pathways is one of the most important research areas in bioinformatics. Recently, there have been several attempts to integrate interactome and transcriptome data to identify subnetworks that provide limited interpretations of known and candidate cancer genes, as well as increase classification accuracy. However, these studies provide little information about the detailed roles of identified cancer genes. RESULTS: To provide more information to the network, we constructed the network by incorporating genetic interactions and manually curated gene regulations to the protein interaction network. To make our newly constructed network cancer specific, we identified edges where two genes show different expression patterns between cancer and normal phenotypes. We showed that the integration of various datasets increased classification accuracy, which suggests that our network is more complete than a network based solely on protein interactions. We also showed that our network contains significantly more known cancer-related genes than other feature selection algorithms. Through observations of some examples of cancer-specific subnetworks, we were able to predict more detailed and interpretable roles of oncogenes and other cancer candidate genes in the prostate cancer cells. AVAILABILITY: http://embio.yonsei.ac.kr/~Ahn/tc.php. CONTACT: sanghyun@cs.yonsei.ac.kr

Predicting chemical structure using reinforcement learning with a stack-augmented conditional variational autoencoder.

In this paper, a reinforcement learning model is proposed that can maximize the predicted binding affinity between a generated molecule and target proteins. The model used to generate molecules in the proposed model was the Stacked Conditional Variation AutoEncoder (Stack-CVAE), which acts as an agent in reinforcement learning so that the resulting chemical formulas have the desired chemical properties and show high binding affinity with specific target proteins. We generated 1000 chemical formulas using the chemical properties of sorafenib and the three target kinases of sorafenib. Then, we confirmed that Stack-CVAE generates more of the valid and unique chemical compounds that have the desired chemical properties and predicted binding affinity better than other generative models. More detailed analysis for 100 of the top scoring molecules show that they are novel ones not found in existing chemical databases. Moreover, they reveal significantly higher predicted binding affinity score for Raf kinases than for other kinases. Furthermore, they are highly druggable and synthesizable.

Blood-based biomarkers for immune-based therapy in advanced HCC: Promising but a long way to go.

The introduction of immune checkpoint inhibitors (ICIs) represents a key shift in the management strategy for patients with hepatocellular carcinoma (HCC). However, there is a paucity of predictive biomarkers that facilitate the identification of patients that would respond to ICI therapy. Although several researchers have attempted to resolve the issue, the data is insufficient to alter daily clinical practice. The use of minimally invasive procedures to obtain patient-derived specimen, such as using blood-based samples, is increasingly preferred. Circulating tumor DNA (ctDNA) can be isolated from the blood of cancer patients, and liquid biopsies can provide sufficient material to enable ongoing monitoring of HCC. This is particularly significant for patients for whom surgery is not indicated, including those with advanced HCC. In this review, we summarize the current state of understanding of blood-based biomarkers for ICI-based therapy in advanced HCC, which is promising despite there is still a long way to go.

Intrahepatic inflammatory IgA+PD-L1high monocytes in hepatocellular carcinoma development and immunotherapy.

BACKGROUND: IgA neutralizes pathogens to prevent infection at mucosal sites. However, emerging evidence shows that IgA contributes to aggravating inflammation or dismantling antitumor immunity in human diseased liver. The aim of this study was to elucidate the roles of inflammation-induced intrahepatic inflammatory IgA+ monocytes in the development of hepatocellular carcinoma (HCC). METHODS: Patient cohorts including steatohepatitis cohort (n=61) and HCC cohort (n=271) were established. Patients' surgical and biopsy specimens were analyzed using immunohistochemistry. Multicolor flow cytometry was performed with a subset of patient samples. Single-cell RNA-Seq analysis was performed using Gene Expression Omnibus (GEO) datasets. Additionally, we performed in vitro differentiation of macrophages, stimulation with coated IgA, and RNA sequencing. Hepa1-6 cells and C57BL/6N mice were used to obtain HCC syngeneic mouse models. RESULTS: Serum IgA levels were associated (p<0.001) with fibrosis progression and HCC development in patients with chronic liver diseases. Additionally, immunohistochemical staining of inflamed livers or HCC revealed IgA positivity in monocytes, with a correlation between IgA+ cell frequency and IgA serum levels. Compared with IgA- monocytes, intrahepatic IgA+ monocytes expressed higher levels of programmed death-ligand 1 (PD-L1) in inflamed livers and in HCC tumor microenvironment. Single-cell RNA sequencing using NCBI GEO database indicated an upregulation in inflammation-associated genes in the monocytes of patients whose plasma cell IGHA1 expression was greater than or equal to the median value. Bulk RNA sequencing demonstrated that in vitro stimulation of M2-polarized macrophages using coated IgA complex induced PD-L1 upregulation via YAP-mediated signaling. In vivo blockade of IgA signaling decreased the number of tumor-infiltrating IgA+PD-L1high macrophages and increased the number of CD69+CD8+ T cells to enhance antitumor effects in HCC mice models. CONCLUSIONS: Overall, the findings of this study showed that serum IgA levels was correlated with intrahepatic and intratumoral infiltration of inflammatory IgA+PD-L1high monocytes in chronic liver diseases and HCC, providing potential therapeutic targets.

GVES: machine learning model for identification of prognostic genes with a small dataset.

Machine learning may be a powerful approach to more accurate identification of genes that may serve as prognosticators of cancer outcomes using various types of omics data. However, to date, machine learning approaches have shown limited prediction accuracy for cancer outcomes, primarily owing to small sample numbers and relatively large number of features. In this paper, we provide a description of GVES (Gene Vector for Each Sample), a proposed machine learning model that can be efficiently leveraged even with a small sample size, to increase the accuracy of identification of genes with prognostic value. GVES, an adaptation of the continuous bag of words (CBOW) model, generates vector representations of all genes for all samples by leveraging gene expression and biological network data. GVES clusters samples using their gene vectors, and identifies genes that divide samples into good and poor outcome groups for the prediction of cancer outcomes. Because GVES generates gene vectors for each sample, the sample size effect is reduced. We applied GVES to six cancer types and demonstrated that GVES outperformed existing machine learning methods, particularly for cancer datasets with a small number of samples. Moreover, the genes identified as prognosticators were shown to reside within a number of significant prognostic genetic pathways associated with pancreatic cancer.

Increasing prediction accuracy of pathogenic staging by sample augmentation with a GAN.

Accurate prediction of cancer stage is important in that it enables more appropriate treatment for patients with cancer. Many measures or methods have been proposed for more accurate prediction of cancer stage, but recently, machine learning, especially deep learning-based methods have been receiving increasing attention, mostly owing to their good prediction accuracy in many applications. Machine learning methods can be applied to high throughput DNA mutation or RNA expression data to predict cancer stage. However, because the number of genes or markers generally exceeds 10,000, a considerable number of data samples is required to guarantee high prediction accuracy. To solve this problem of a small number of clinical samples, we used a Generative Adversarial Networks (GANs) to augment the samples. Because GANs are not effective with whole genes, we first selected significant genes using DNA mutation data and random forest feature ranking. Next, RNA expression data for selected genes were expanded using GANs. We compared the classification accuracies using original dataset and expanded datasets generated by proposed and existing methods, using random forest, Deep Neural Networks (DNNs), and 1-Dimensional Convolutional Neural Networks (1DCNN). When using the 1DCNN, the F1 score of GAN5 (a 5-fold increase in data) was improved by 39% in relation to the original data. Moreover, the results using only 30% of the data were better than those using all of the data. Our attempt is the first to use GAN for augmentation using numeric data for both DNA and RNA. The augmented datasets obtained using the proposed method demonstrated significantly increased classification accuracy for most cases. By using GAN and 1DCNN in the prediction of cancer stage, we confirmed that good results can be obtained even with small amounts of samples, and it is expected that a great deal of the cost and time required to obtain clinical samples will be reduced. The proposed sample augmentation method could also be applied for other purposes, such as prognostic prediction or cancer classification.

RefDNN: a reference drug based neural network for more accurate prediction of anticancer drug resistance.

Cancer is one of the most difficult diseases to treat owing to the drug resistance of tumour cells. Recent studies have revealed that drug responses are closely associated with genomic alterations in cancer cells. Numerous state-of-the-art machine learning models have been developed for prediction of drug responses using various genomic data and diverse drug molecular information, but those methods are ineffective to predict drug response to untrained drugs and gene expression patterns, which is known as the cold-start problem. In this study, we present a novel deep neural network model, termed RefDNN, for improved prediction of drug resistance and identification of biomarkers related to drug response. RefDNN exploits a collection of drugs, called reference drugs, to learn representations for a high-dimensional gene expression vector and a molecular structure vector of a drug and predicts drug response labels using the reference drug-based representations. These calculations come from the observation that similar chemicals have similar effects. The proposed model not only outperformed existing computational prediction models in most comparative experiments, but also showed more robust prediction for untrained drugs and cancer types than traditional machine learning models. RefDNN exploits the ElasticNet regularization to deal with high-dimensional gene expression data, which allows identification of gene markers associated with drug resistance. Lastly, we described an application of RefDNN in exploring a new candidate drug for liver cancer. As the proposed model can guarantee good prediction of drug responses to untrained drugs for given gene expression patterns, it may be of potential benefit in drug repositioning and personalized medicine.