MOTIVATION: Combinatorial drug therapy holds great promise for tackling complex diseases, but the vast number of possible drug combinations makes exhaustive experimental testing infeasible. Computational models have been developed to guide experimental screens by assigning synergy scores to drug pair-cell line combinations, where they take input structural and chemical information on drugs and molecular features of cell lines. The premise of these models is that they leverage this biological and chemical information to predict synergy measurements. RESULTS: In this study, we demonstrate that replacing drug and cell line representations with simple one-hot encodings results in comparable or even slightly improved performance across diverse published drug combination models. This unexpected finding suggests that current models use these representations primarily as identifiers and exploit covariation in the synergy labels. Our synthetic data experiments show that models can learn from the true features; however, when drugs and cell lines recur across drug-drug-cell triplets, this repeating structure impairs feature-based learning. While the current synergy prediction models can aid in prioritizing drug pairs within a panel of tested drugs and cell lines, our results highlight the need for better strategies to learn from intended features and to generalize to unseen drugs and cell lines. AVAILABILITY AND IMPLEMENTATION: The scripts to run the experiments are available at: https://github.com/tastanlab/ohe.

Many cancer monotherapies demonstrate limited clinical efficacy, making combination therapies a relevant treatment strategy. The extensive number of potential drug combinations and context-specific response profiles complicates the prediction of drug combination responses. Existing computational models are typically trained to predict a single aggregated synergy score, which summarises drug responses across different dosage combinations, such as Bliss or Loewe scores. This oversimplification of the drug-response surface leads to high prediction uncertainty and limited actionability, as these models fail to distinguish between potency and efficacy. We introduce DeepSynBa, an actionable model that predicts the complete dose-response matrix of drug pairs instead of relying on an aggregated synergy score. This is achieved by predicting parameters describing the response surface as an intermediate layer in the model. Evaluated on the NCI-ALMANAC and the O'Neil datasets, DeepSynBa outperforms the state-of-the-art methods in the dose-response matrix prediction task across most evaluation scenarios, including testing on novel drug combinations, cell lines, and drugs, across nine different tissue types. We also show that DeepSynBa yields reliable synergy score predictions. More importantly, DeepSynBa can predict drug combination responses across different dosages for untested combinations. The intermediate dose-response parameter layer enables the separation of efficacy from potency, informing the selection of dosage ranges that optimise efficacy while limiting off-target toxicity in experimental screens. The predictive capability and the downstream actionability make DeepSynBa a powerful tool for advancing drug combination research beyond the limitations of the current approaches. The code and the dataset for DeepSynBa are available at https://github.com/hikuru/DeepSynBa.

MOTIVATION: Protein language models (pLMs) have emerged as powerful tools for capturing the intricate information encoded in protein sequences, facilitating various downstream protein prediction tasks. With numerous pLMs available, there is a critical need for diverse benchmarks to systematically evaluate their performance across biologically relevant tasks. Here, we introduce DARKIN, a zero-shot classification benchmark designed to assign phosphosites to understudied kinases, termed dark kinases. Kinases, which catalyze phosphorylation, are central to cellular signaling pathways. While phosphoproteomics enables the large-scale identification of phosphosites, determining the cognate kinase responsible for the phosphorylation event remains an experimental challenge. RESULTS: In DARKIN, we prepared training, validation, and test folds that respect the zero-shot nature of this classification problem, incorporating stratification based on kinase groups and sequence similarity. We evaluated multiple pLMs using two zero-shot classifiers: a novel, training-free k-NN-based method, and a bilinear classifier. Our findings indicate that ESM, ProtT5-XL, and SaProt exhibit superior performance on this task. DARKIN provides a challenging benchmark for assessing pLM efficacy and fosters deeper exploration of under-characterized (dark) kinases by offering a biologically relevant test bed. AVAILABILITY AND IMPLEMENTATION: The DARKIN benchmark data and the scripts for generating additional splits are publicly available at: https://github.com/tastanlab/darkin.

MOTIVATION: Combination drug therapies are effective treatments for cancer. However, the genetic heterogeneity of the patients and exponentially large space of drug pairings pose significant challenges for finding the right combination for a specific patient. Current in silico prediction methods can be instrumental in reducing the vast number of candidate drug combinations. However, existing powerful methods are trained with cancer cell line gene expression data, which limits their applicability in clinical settings. While synergy measurements on cell line models are available at large scale, patient-derived samples are too few to train a complex model. On the other hand, patient-specific single-drug response data are relatively more available. RESULTS: In this work, we propose a deep learning framework, Personalized Deep Synergy Predictor (PDSP), that enables us to use the patient-specific single drug response data for customizing patient drug synergy predictions. PDSP is first trained to learn synergy scores of drug pairs and their single drug responses for a given cell line using drug structures and large scale cell line gene expression data. Then, the model is fine-tuned for patients with their patient gene expression data and associated single drug response measured on the patient ex vivo samples. In this study, we evaluate PDSP on data from three leukemia patients and observe that it improves the prediction accuracy by 27% compared to models trained on cancer cell line data. AVAILABILITY AND IMPLEMENTATION: PDSP is available at https://github.com/hikuru/PDSP.

Abstract Named entity recognition (NER) aims to identify mentions of named entities in an unstructured text and classify them into predefined named entity classes. While deep learning-based pre-trained language models help to achieve good predictive performances in NER, many domain-specific NER applications still call for a substantial amount of labeled data. Active learning (AL), a general framework for the label acquisition problem, has been used for NER tasks to minimize the annotation cost without sacrificing model performance. However, the heavily imbalanced class distribution of tokens introduces challenges in designing effective AL querying methods for NER. We propose several AL sentence query evaluation functions that pay more attention to potential positive tokens and evaluate these proposed functions with both sentence-based and token-based cost evaluation strategies. We also propose a better data-driven normalization approach to penalize sentences that are too long or too short. Our experiments on three datasets from different domains reveal that the proposed approach reduces the number of annotated tokens while achieving better or comparable prediction performance with conventional methods.

MOTIVATION: Recent experimental evidence has shown that some long non-coding RNAs (lncRNAs) contain small open reading frames (sORFs) that are translated into functional micropeptides, suggesting that these lncRNAs are misannotated as non-coding. Current methods to detect misannotated lncRNAs rely on ribosome-profiling (Ribo-Seq) and mass-spectrometry experiments, which are cell-type dependent and expensive. RESULTS: Here, we propose a computational method to identify possible misannotated lncRNAs from sequence information alone. Our approach first builds deep learning models to discriminate coding and non-coding transcripts and leverages these models' training dynamics to identify misannotated lncRNAs-i.e. lncRNAs with coding potential. The set of misannotated lncRNAs we identified significantly overlap with experimentally validated ones and closely resemble coding protein sequences as evidenced by significant BLAST hits. Our analysis on a subset of misannotated lncRNA candidates also shows that some ORFs they contain yield high confidence folded structures as predicted by AlphaFold2. This methodology offers promising potential for assisting experimental efforts in characterizing the hidden proteome encoded by misannotated lncRNAs and for curating better datasets for building coding potential predictors. AVAILABILITY AND IMPLEMENTATION: Source code is available at https://github.com/nabiafshan/DetectingMisannotatedLncRNAs. SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

MOTIVATION: With the wide availability of single-cell RNA-seq (scRNA-seq) technology, population-scale scRNA-seq datasets across multiple individuals and time points are emerging. While the initial investigations of these datasets tend to focus on standard analysis of clustering and differential expression, leveraging the power of scRNA-seq data at the personalized dynamic gene co-expression network level has the potential to unlock subject and/or time-specific network-level variation, which is critical for understanding phenotypic differences. Community detection from co-expression networks of multiple time points or conditions has been well-studied; however, none of the existing settings included networks from multiple subjects and multiple time points simultaneously. To address this, we develop Multi-subject Dynamic Community Detection (MuDCoD) for multi-subject community detection in personalized dynamic gene networks from scRNA-seq. MuDCoD builds on the spectral clustering framework and promotes information sharing among the networks of the subjects as well as networks at different time points. It clusters genes in the personalized dynamic gene networks and reveals gene communities that are variable or shared not only across time but also among subjects. RESULTS: Evaluation and benchmarking of MuDCoD against existing approaches reveal that MuDCoD effectively leverages apparent shared signals among networks of the subjects at individual time points, and performs robustly when there is no or little information sharing among the networks. Applications to population-scale scRNA-seq datasets of human-induced pluripotent stem cells during dopaminergic neuron differentiation and CD4+ T cell activation indicate that MuDCoD enables robust inference for identifying time-varying personalized gene modules. Our results illustrate how personalized dynamic community detection can aid in the exploration of subject-specific biological processes that vary across time. AVAILABILITY AND IMPLEMENTATION: MuDCoD is publicly available at https://github.com/bo1929/MuDCoD as a Python package. Implementation includes simulation and real-data experiments together with extensive documentation.

Drug failures due to unforeseen adverse effects at clinical trials pose health risks for the participants and lead to substantial financial losses. Side effect prediction algorithms have the potential to guide the drug design process. LINCS L1000 dataset provides a vast resource of cell line gene expression data perturbed by different drugs and creates a knowledge base for context specific features. The state-of-the-art approach that aims at using context specific information relies on only the high-quality experiments in LINCS L1000 and discards a large portion of the experiments. In this study, our goal is to boost the prediction performance by utilizing this data to its full extent. We experiment with 5 deep learning architectures. We find that a multi-modal architecture produces the best predictive performance among multi-layer perceptron-based architectures when drug chemical structure (CS), and the full set of drug perturbed gene expression profiles (GEX) are used as modalities. Overall, we observe that the CS is more informative than the GEX. A convolutional neural network-based model that uses only SMILES string representation of the drugs achieves the best results and provides 13.0% macro-AUC and 3.1% micro-AUC improvements over the state-of-the-art. We also show that the model is able to predict side effect-drug pairs that are reported in the literature but was missing in the ground truth side effect dataset. DeepSide is available at http://github.com/OnurUner/DeepSide.

MOTIVATION: Genome-wide association studies show that variants in individual genomic loci alone are not sufficient to explain the heritability of complex, quantitative phenotypes. Many computational methods have been developed to address this issue by considering subsets of loci that can collectively predict the phenotype. This problem can be considered a challenging instance of feature selection in which the number of dimensions (loci that are screened) is much larger than the number of samples. While currently available methods can achieve decent phenotype prediction performance, they either do not scale to large datasets or have parameters that require extensive tuning. RESULTS: We propose a fast and simple algorithm, Macarons, to select a small, complementary subset of variants by avoiding redundant pairs that are likely to be in linkage disequilibrium. Our method features two interpretable parameters that control the time/performance trade-off without requiring parameter tuning. In our computational experiments, we show that Macarons consistently achieves similar or better prediction performance than state-of-the-art selection methods while having a simpler premise and being at least two orders of magnitude faster. Overall, Macarons can seamlessly scale to the human genome with ∼107 variants in a matter of minutes while taking the dependencies between the variants into account. AVAILABILITYAND IMPLEMENTATION: Macarons is available in Matlab and Python at https://github.com/serhan-yilmaz/macarons. SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

Drug combination therapies have been a viable strategy for the treatment of complex diseases such as cancer due to increased efficacy and reduced side effects. However, experimentally validating all possible combinations for synergistic interaction even with high-throughout screens is intractable due to vast combinatorial search space. Computational techniques can reduce the number of combinations to be evaluated experimentally by prioritizing promising candidates. We present MatchMaker that predicts drug synergy scores using drug chemical structure information and gene expression profiles of cell lines in a deep learning framework. For the first time, our model utilizes the largest known drug combination dataset to date, DrugComb. We compare the performance of MatchMaker with the state-of-the-art models and observe up to ∼ 15% correlation and ∼ 33% mean squared error (MSE) improvements over the next best method. We investigate the cell types and drug pairs that are relatively harder to predict and present novel candidate pairs. MatchMaker is built and available at https://github.com/tastanlab/matchmaker.

Although miRNAs can cause widespread changes in expression programs, single miRNAs typically induce mild repression on their targets. Cooperativity among miRNAs is reported as one strategy to overcome this constraint. Expanding the catalog of synergistic miRNAs is critical for understanding gene regulation and for developing miRNA-based therapeutics. In this study, we develop miRCoop to identify synergistic miRNA pairs that have weak or no repression on the target mRNA individually, but when act together, induce strong repression. miRCoop uses kernel-based statistical interaction tests, together with miRNA and mRNA target information. We apply our approach to patient data of two different cancer types. In kidney cancer, we identify 66 putative triplets. For 64 of these triplets, there is at least one common transcription factor that potentially regulates all participating RNAs of the triplet, supporting a functional association among them. Furthermore, we find that identified triplets are enriched for certain biological processes that are relevant to kidney cancer. Some of the synergistic miRNAs are very closely encoded in the genome, hinting a functional association among them. In applying the method on tumor data with the primary liver site, we find 3105 potential triplet interactions. We believe miRCoop can aid our understanding of the complex regulatory interactions in different health and disease states of the cell and can help in designing miRNA-based therapies. Matlab code for the methodology is provided in https://github.com/guldenolgun/miRCoop.

MOTIVATION: Accurate classification of patients into molecular subgroups is critical for the development of effective therapeutics and for deciphering what drives these subgroups to cancer. The availability of multiomics data catalogs for large cohorts of cancer patients provides multiple views into the molecular biology of the tumors with unprecedented resolution. RESULTS: We develop Pathway-based MultiOmic Graph Kernel clustering (PAMOGK) that integrates multiomics patient data with existing biological knowledge on pathways. We develop a novel graph kernel that evaluates patient similarities based on a single molecular alteration type in the context of a pathway. To corroborate multiple views of patients evaluated by hundreds of pathways and molecular alteration combinations, we use multiview kernel clustering. Applying PAMOGK to kidney renal clear cell carcinoma (KIRC) patients results in four clusters with significantly different survival times (P-value =1.24e-11). When we compare PAMOGK to eight other state-of-the-art multiomics clustering methods, PAMOGK consistently outperforms these in terms of its ability to partition KIRC patients into groups with different survival distributions. The discovered patient subgroups also differ with respect to other clinical parameters such as tumor stage and grade, and primary tumor and metastasis tumor spreads. The pathways identified as important are highly relevant to KIRC. AVAILABILITY AND IMPLEMENTATION: github.com/tastanlab/pamogk. SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

BACKGROUND: While some non-coding RNAs (ncRNAs) are assigned critical regulatory roles, most remain functionally uncharacterized. This presents a challenge whenever an interesting set of ncRNAs needs to be analyzed in a functional context. Transcripts located close-by on the genome are often regulated together. This genomic proximity on the sequence can hint at a functional association. RESULTS: We present a tool, NoRCE, that performs cis enrichment analysis for a given set of ncRNAs. Enrichment is carried out using the functional annotations of the coding genes located proximal to the input ncRNAs. Other biologically relevant information such as topologically associating domain (TAD) boundaries, co-expression patterns, and miRNA target prediction information can be incorporated to conduct a richer enrichment analysis. To this end, NoRCE includes several relevant datasets as part of its data repository, including cell-line specific TAD boundaries, functional gene sets, and expression data for coding & ncRNAs specific to cancer. Additionally, the users can utilize custom data files in their investigation. Enrichment results can be retrieved in a tabular format or visualized in several different ways. NoRCE is currently available for the following species: human, mouse, rat, zebrafish, fruit fly, worm, and yeast. CONCLUSIONS: NoRCE is a platform-independent, user-friendly, comprehensive R package that can be used to gain insight into the functional importance of a list of ncRNAs of any type. The tool offers flexibility to conduct the users' preferred set of analyses by designing their own pipeline of analysis. NoRCE is available in Bioconductor and https://github.com/guldenolgun/NoRCE .

Changes in protein and gene expression levels are often used as features in predictive modeling such as survival prediction. A common strategy to aggregate information contained in individual proteins is to integrate the expression levels with the biological networks. In this work, we propose a novel patient representation where we integrate proteins' expression levels with the protein-protein interaction (PPI) networks: Patient representation with PRER (Pairwise Relative Expressions with Random walks). PRER captures the dysregulation patterns of proteins based on the neighborhood of a protein in the PPI network. Specifically, PRER computes a feature vector for a patient by comparing the source protein's expression level with other proteins' levels that are within its neighborhood. The neighborhood of the source protein is derived by biased random-walk strategy on the network. We test PRER's performance in survival prediction task in 10 different cancers using random forest survival models. PRER yields a statistically significant predictive performance in 9 out of 10 cancers when compared to the same model trained with features based on individual protein expressions. Furthermore, we identified the pairs of proteins that their interactions are predictive of patient survival but their individual expression levels are not. The set of identified relations provides a valuable collection of protein biomarkers with high prognostic value. PRER can be used for other complex diseases and prediction tasks that use molecular expression profiles as input. PRER is freely available at: https://github.com/hikuru/PRER.

Phenotypic heritability of complex traits and diseases is seldom explained by individual genetic variants identified in genome-wide association studies (GWAS). Many methods have been developed to select a subset of variant loci, which are associated with or predictive of the phenotype. Selecting connected SNPs on SNP-SNP networks have been proven successful in finding biologically interpretable and predictive SNPs. However, we argue that the connectedness constraint favors selecting redundant features that affect similar biological processes and therefore does not necessarily yield better predictive performance. In this paper, we propose a novel method called SPADIS that favors the selection of remotely located SNPs in order to account for their complementary effects in explaining a phenotype. SPADIS selects a diverse set of loci on a SNP-SNP network. This is achieved by maximizing a submodular set function with a greedy algorithm that ensures a constant factor approximation to the optimal solution. We compare SPADIS to the state-of-the-art method SConES, on a dataset of Arabidopsis Thaliana with continuous flowering time phenotypes. SPADIS has better average phenotype prediction performance in 15 out of 17 phenotypes when the same number of SNPs are selected and provides consistent improvements across multiple networks and settings on average. Moreover, it identifies more candidate genes and runs faster.

MOTIVATION: Protein phosphorylation is a key regulator of protein function in signal transduction pathways. Kinases are the enzymes that catalyze the phosphorylation of other proteins in a target-specific manner. The dysregulation of phosphorylation is associated with many diseases including cancer. Although the advances in phosphoproteomics enable the identification of phosphosites at the proteome level, most of the phosphoproteome is still in the dark: more than 95% of the reported human phosphosites have no known kinases. Determining which kinase is responsible for phosphorylating a site remains an experimental challenge. Existing computational methods require several examples of known targets of a kinase to make accurate kinase-specific predictions, yet for a large body of kinases, only a few or no target sites are reported. RESULTS: We present DeepKinZero, the first zero-shot learning approach to predict the kinase acting on a phosphosite for kinases with no known phosphosite information. DeepKinZero transfers knowledge from kinases with many known target phosphosites to those kinases with no known sites through a zero-shot learning model. The kinase-specific positional amino acid preferences are learned using a bidirectional recurrent neural network. We show that DeepKinZero achieves significant improvement in accuracy for kinases with no known phosphosites in comparison to the baseline model and other methods available. By expanding our knowledge on understudied kinases, DeepKinZero can help to chart the phosphoproteome atlas. AVAILABILITY AND IMPLEMENTATION: The source codes are available at https://github.com/Tastanlab/DeepKinZero. SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

MOTIVATION: Rapid and low cost sequencing of genomes enabled widespread use of genomic data in research studies and personalized customer applications, where genomic data is shared in public databases. Although the identities of the participants are anonymized in these databases, sensitive information about individuals can still be inferred. One such information is kinship. RESULTS: We define two routes kinship privacy can leak and propose a technique to protect kinship privacy against these risks while maximizing the utility of shared data. The method involves systematic identification of minimal portions of genomic data to mask as new participants are added to the database. Choosing the proper positions to hide is cast as an optimization problem in which the number of positions to mask is minimized subject to privacy constraints that ensure the familial relationships are not revealed. We evaluate the proposed technique on real genomic data. Results indicate that concurrent sharing of data pertaining to a parent and an offspring results in high risks of kinship privacy, whereas the sharing data from further relatives together is often safer. We also show arrival order of family members have a high impact on the level of privacy risks and on the utility of sharing data. AVAILABILITY AND IMPLEMENTATION: https://github.com/tastanlab/Kinship-Privacy. SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

Individuals (and their family members) share (partial) genomic data on public platforms. However, using special characteristics of genomic data, background knowledge that can be obtained from the Web, and family relationship between the individuals, it is possible to infer the hidden parts of shared (and unshared) genomes. Existing work in this field considers simple correlations in the genome (as well as Mendel's law and partial genomes of a victim and his family members). In this paper, we improve the existing work on inference attacks on genomic privacy. We mainly consider complex correlations in the genome by using an observable Markov model and recombination model between the haplotypes. We also utilize the phenotype information about the victims. We propose an efficient message passing algorithm to consider all aforementioned background information for the inference. We show that the proposed framework improves inference with significantly less information compared to existing work.

MOTIVATION: Genomic studies identify genomic loci representing genetic variations, transcription factor (TF) occupancy, or histone modification through next generation sequencing (NGS) technologies. Interpreting these loci requires evaluating them with known genomic and epigenomic annotations. RESULTS: We present GLANET as a comprehensive annotation and enrichment analysis tool which implements a sampling-based enrichment test that accounts for GC content and/or mappability biases, jointly or separately. GLANET annotates and performs enrichment analysis on these loci with a rich library. We introduce and perform novel data-driven computational experiments for assessing the power and Type-I error of its enrichment procedure which show that GLANET has attained high statistical power and well-controlled Type-I error rate. As a key feature, users can easily extend its library with new gene sets and genomic intervals. Other key features include assessment of impact of single nucleotide variants (SNPs) on TF binding sites and regulation based pathway enrichment analysis. AVAILABILITY AND IMPLEMENTATION: GLANET can be run using its GUI or on command line. GLANET's source code is available at https://github.com/burcakotlu/GLANET . Tutorials are provided at https://glanet.readthedocs.org . CONTACT: [email protected] or [email protected]. SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

MOTIVATION: Protein-protein interactions (PPIs) are critical for virtually every biological function. Recently, researchers suggested to use supervised learning for the task of classifying pairs of proteins as interacting or not. However, its performance is largely restricted by the availability of truly interacting proteins (labeled). Meanwhile, there exists a considerable amount of protein pairs where an association appears between two partners, but not enough experimental evidence to support it as a direct interaction (partially labeled). RESULTS: We propose a semi-supervised multi-task framework for predicting PPIs from not only labeled, but also partially labeled reference sets. The basic idea is to perform multi-task learning on a supervised classification task and a semi-supervised auxiliary task. The supervised classifier trains a multi-layer perceptron network for PPI predictions from labeled examples. The semi-supervised auxiliary task shares network layers of the supervised classifier and trains with partially labeled examples. Semi-supervision could be utilized in multiple ways. We tried three approaches in this article, (i) classification (to distinguish partial positives with negatives); (ii) ranking (to rate partial positive more likely than negatives); (iii) embedding (to make data clusters get similar labels). We applied this framework to improve the identification of interacting pairs between HIV-1 and human proteins. Our method improved upon the state-of-the-art method for this task indicating the benefits of semi-supervised multi-task learning using auxiliary information. AVAILABILITY: http://www.cs.cmu.edu/~qyj/HIVsemi.