Jianyang Zeng

Predicting drug-target interactions using restricted Boltzmann machines

Motivation: In silico prediction of drug-target interactions plays an important role toward identifying and developing new uses of existing or abandoned drugs. Network-based approaches have recently become a popular tool for discovering new drug-target interactions (DTIs). Unfortunately, most of these network-based approaches can only predict binary interactions between drugs and targets, and information about different types of interactions has not been well exploited for DTI prediction in previous studies. On the other hand, incorporating additional information about drug-target relationships or drug modes of action can improve prediction of DTIs. Furthermore, the predicted types of DTIs can broaden our understanding about the molecular basis of drug action. Results: We propose a first machine learning approach to integrate multiple types of DTIs and predict unknown drug-target relationships or drug modes of action. We cast the new DTI prediction problem into a two-layer graphical model, called restricted Boltzmann machine, and apply a practical learning algorithm to train our model and make predictions. Tests on two public databases show that our restricted Boltzmann machine model can effectively capture the latent features of a DTI network and achieve excellent performance on predicting different types of DTIs, with the area under precision-recall curve up to 89.6. In addition, we demonstrate that integrating multiple types of DTIs can significantly outperform other predictions either by simply mixing multiple types of interactions without distinction or using only a single interaction type. Further tests show that our approach can infer a high fraction of novel DTIs that has been validated by known experiments in the literature or other databases. These results indicate that our approach can have highly practical relevance to DTI prediction and drug repositioning, and hence advance the drug discovery process. Availability: Software and datasets are available on request. Contact: zengjy321@tsinghua.edu.cn Supplementary information: Supplementary data are available at Bioinformatics online.

Integrative data analysis of multi-platform cancer data with a multimodal deep learning approach

Identification of cancer subtypes plays an important role in revealing useful insights into disease pathogenesis and advancing personalized therapy. The recent development of high-throughput sequencing technologies has enabled the rapid collection of multi-platform genomic data (e.g., gene expression, miRNA expression, and DNA methylation) for the same set of tumor samples. Although numerous integrative clustering approaches have been developed to analyze cancer data, few of them are particularly designed to exploit both deep intrinsic statistical properties of each input modality and complex cross-modality correlations among multi-platform input data. In this paper, we propose a new machine learning model, called multimodal deep belief network (DBN), to cluster cancer patients from multi-platform observation data. In our integrative clustering framework, relationships among inherent features of each single modality are first encoded into multiple layers of hidden variables, and then a joint latent model is employed to fuse common features derived from multiple input modalities. A practical learning algorithm, called contrastive divergence (CD), is applied to infer the parameters of our multimodal DBN model in an unsupervised manner. Tests on two available cancer datasets show that our integrative data analysis approach can effectively extract a unified representation of latent features to capture both intra- and cross-modality correlations, and identify meaningful disease subtypes from multi-platform cancer data. In addition, our approach can identify key genes and miRNAs that may play distinct roles in the pathogenesis of different cancer subtypes. Among those key miRNAs, we found that the expression level of miR-29a is highly correlated with survival time in ovarian cancer patients. These results indicate that our multimodal DBN based data analysis approach may have practical applications in cancer pathogenesis studies and provide useful guidelines for personalized cancer therapy.

Allelic reprogramming of 3D chromatin architecture during early mammalian development

In mammals, chromatin organization undergoes drastic reprogramming after fertilization. However, the three-dimensional structure of chromatin and its reprogramming in preimplantation development remain poorly understood. Here, by developing a low-input Hi-C (genome-wide chromosome conformation capture) approach, we examined the reprogramming of chromatin organization during early development in mice. We found that oocytes in metaphase II show homogeneous chromatin folding that lacks detectable topologically associating domains (TADs) and chromatin compartments. Strikingly, chromatin shows greatly diminished higher-order structure after fertilization. Unexpectedly, the subsequent establishment of chromatin organization is a prolonged process that extends through preimplantation development, as characterized by slow consolidation of TADs and segregation of chromatin compartments. The two sets of parental chromosomes are spatially separated from each other and display distinct compartmentalization in zygotes. Such allele separation and allelic compartmentalization can be found as late as the 8-cell stage. Finally, we show that chromatin compaction in preimplantation embryos can partially proceed in the absence of zygotic transcription and is a multi-level hierarchical process. Taken together, our data suggest that chromatin may exist in a markedly relaxed state after fertilization, followed by progressive maturation of higher-order chromatin architecture during early development.

A deep learning framework for modeling structural features of RNA-binding protein targets

RNA-binding proteins (RBPs) play important roles in the post-transcriptional control of RNAs. Identifying RBP binding sites and characterizing RBP binding preferences are key steps toward understanding the basic mechanisms of the post-transcriptional gene regulation. Though numerous computational methods have been developed for modeling RBP binding preferences, discovering a complete structural representation of the RBP targets by integrating their available structural features in all three dimensions is still a challenging task. In this paper, we develop a general and flexible deep learning framework for modeling structural binding preferences and predicting binding sites of RBPs, which takes (predicted) RNA tertiary structural information into account for the first time. Our framework constructs a unified representation that characterizes the structural specificities of RBP targets in all three dimensions, which can be further used to predict novel candidate binding sites and discover potential binding motifs. Through testing on the real CLIP-seq datasets, we have demonstrated that our deep learning framework can automatically extract effective hidden structural features from the encoded raw sequence and structural profiles, and predict accurate RBP binding sites. In addition, we have conducted the first study to show that integrating the additional RNA tertiary structural features can improve the model performance in predicting RBP binding sites, especially for the polypyrimidine tract-binding protein (PTB), which also provides a new evidence to support the view that RBPs may own specific tertiary structural binding preferences. In particular, the tests on the internal ribosome entry site (IRES) segments yield satisfiable results with experimental support from the literature and further demonstrate the necessity of incorporating RNA tertiary structural information into the prediction model. The source code of our approach can be found in https://github.com/thucombio/deepnet-rbp.

High-resolution protein structure determination starting with a global fold calculated from exact solutions to the RDC equations.

We present a novel structure determination approach that exploits the global orientational restraints from RDCs to resolve ambiguous NOE assignments. Unlike traditional approaches that bootstrap the initial fold from ambiguous NOE assignments, we start by using RDCs to compute accurate secondary structure element (SSE) backbones at the beginning of structure calculation. Our structure determination package, called rdc-Panda (RDC-based SSE PAcking with NOEs for Structure Determination and NOE Assignment), consists of three modules: (1) rdc-exact; (2) Packer; and (3) hana (HAusdorff-based NOE Assignment). rdc-exact computes the global optimal solution of backbone dihedral angles for each secondary structure element by exactly solving a system of quartic RDC equations derived by Wang and Donald (Proceedings of the IEEE computational systems bioinformatics conference (CSB), Stanford, CA, 2004a; J Biomol NMR 29(3):223–242, 2004b), and systematically searching over the roots, each of which is a backbone dihedral ϕ- or ψ-angle consistent with the RDC data. Using a small number of unambiguous inter-SSE NOEs extracted using only chemical shift information, Packer performs a systematic search for the core structure, including all SSE backbone conformations. hana uses a Hausdorff-based scoring function to measure the similarity between the experimental spectra and the back-computed NOE pattern for each side-chain from a statistically-diverse rotamer library, and drives the selection of optimal position-specific rotamers for filtering ambiguous NOE assignments. Finally, a local minimization approach is used to compute the loops and refine side-chain conformations by fixing the core structure as a rigid body while allowing movement of loops and side-chains. rdc-Panda was applied to NMR data for the FF Domain 2 of human transcription elongation factor CA150 (RNA polymerase II C-terminal domain interacting protein), human ubiquitin, the ubiquitin-binding zinc finger domain of the human Y-family DNA polymerase Eta (pol η UBZ), and the human Set2-Rpb1 interacting domain (hSRI). These results demonstrated the efficiency and accuracy of our algorithm, and show that rdc-Panda can be successfully applied for high-resolution protein structure determination using only a limited set of NMR data by first computing RDC-defined backbones.

Inferential modeling of 3D chromatin structure

For eukaryotic cells, the biological processes involving regulatory DNA elements play an important role in cell cycle. Understanding 3D spatial arrangements of chromosomes and revealing long-range chromatin interactions are critical to decipher these biological processes. In recent years, chromosome conformation capture (3C) related techniques have been developed to measure the interaction frequencies between long-range genome loci, which have provided a great opportunity to decode the 3D organization of the genome. In this paper, we develop a new Bayesian framework to derive the 3D architecture of a chromosome from 3C-based data. By modeling each chromosome as a polymer chain, we define the conformational energy based on our current knowledge on polymer physics and use it as prior information in the Bayesian framework. We also propose an expectation-maximization (EM) based algorithm to estimate the unknown parameters of the Bayesian model and infer an ensemble of chromatin structures based on interaction frequency data. We have validated our Bayesian inference approach through cross-validation and verified the computed chromatin conformations using the geometric constraints derived from fluorescence in situ hybridization (FISH) experiments. We have further confirmed the inferred chromatin structures using the known genetic interactions derived from other studies in the literature. Our test results have indicated that our Bayesian framework can compute an accurate ensemble of 3D chromatin conformations that best interpret the distance constraints derived from 3C-based data and also agree with other sources of geometric constraints derived from experimental evidence in the previous studies. The source code of our approach can be found in https://github.com/wangsy11/InfMod3DGen.

An efficient parallel algorithm for accelerating computational protein design

Motivation: Structure-based computational protein design (SCPR) is an important topic in protein engineering. Under the assumption of a rigid backbone and a finite set of discrete conformations of side-chains, various methods have been proposed to address this problem. A popular method is to combine the dead-end elimination (DEE) and A* tree search algorithms, which provably finds the global minimum energy conformation (GMEC) solution. Results: In this article, we improve the efficiency of computing A* heuristic functions for protein design and propose a variant of A* algorithm in which the search process can be performed on a single GPU in a massively parallel fashion. In addition, we make some efforts to address the memory exceeding problem in A* search. As a result, our enhancements can achieve a significant speedup of the A*-based protein design algorithm by four orders of magnitude on large-scale test data through pre-computation and parallelization, while still maintaining an acceptable memory overhead. We also show that our parallel A* search algorithm could be successfully combined with iMinDEE, a state-of-the-art DEE criterion, for rotamer pruning to further improve SCPR with the consideration of continuous side-chain flexibility. Availability: Our software is available and distributed open-source under the GNU Lesser General License Version 2.1 (GNU, February 1999). The source code can be downloaded from http://www.cs.duke.edu/donaldlab/osprey.php or http://iiis.tsinghua.edu.cn/∼compbio/software.html. Contact: zengjy321@tsinghua.edu.cn Supplementary information: Supplementary data are available at Bioinformatics online.

Protein Loop Closure Using Orientational Restraints from NMR Data

Protein loops often play important roles in biological functions. Modeling loops accurately is crucial to determining the functional specificity of a protein. Despite the recent progress in loop prediction approaches, which led to a number of algorithms over the past decade, few rigorous algorithmic approaches exist to model protein loops using global orientational restraints, such as those obtained from residual dipolar coupling (RDC) data in solution nuclear magnetic resonance (NMR) spectroscopy. In this article, we present a novel, sparse data, RDC‐based algorithm, which exploits the mathematical interplay between RDC‐derived sphero‐conics and protein kinematics, and formulates the loop structure determination problem as a system of low‐degree polynomial equations that can be solved exactly, in closed‐form. The polynomial roots, which encode the candidate conformations, are searched systematically, using provable pruning strategies that triage the vast majority of conformations, to enumerate or prune all possible loop conformations consistent with the data; therefore, completeness is ensured. Results on experimental RDC datasets for four proteins, including human ubiquitin, FF2, DinI, and GB3, demonstrate that our algorithm can compute loops with higher accuracy, a three‐ to six‐fold improvement in backbone RMSD, versus those obtained by traditional structure determination protocols on the same data. Excellent results were also obtained on synthetic RDC datasets for protein loops of length 4, 8, and 12 used in previous studies. These results suggest that our algorithm can be successfully applied to determine protein loop conformations, and hence, will be useful in high‐resolution protein backbone structure determination, including loops, from sparse NMR data. Proteins 2012.

A markov random field framework for protein side-chain resonance assignment

Nuclear magnetic resonance (NMR) spectroscopy plays a critical role in structural genomics, and serves as a primary tool for determining protein structures, dynamics and interactions in physiologically-relevant solution conditions The current speed of protein structure determination via NMR is limited by the lengthy time required in resonance assignment, which maps spectral peaks to specific atoms and residues in the primary sequence Although numerous algorithms have been developed to address the backbone resonance assignment problem [68,2,10,37,14,64,1,31,60], little work has been done to automate side-chain resonance assignment [43, 48, 5] Most previous attempts in assigning side-chain resonances depend on a set of NMR experiments that record through-bond interactions with side-chain protons for each residue Unfortunately, these NMR experiments have low sensitivity and limited performance on large proteins, which makes it difficult to obtain enough side-chain resonance assignments On the other hand, it is essential to obtain almost all of the side-chain resonance assignments as a prerequisite for high-resolution structure determination To overcome this deficiency, we present a novel side-chain resonance assignment algorithm based on alternative NMR experiments measuring through-space interactions between protons in the protein, which also provide crucial distance restraints and are normally required in high-resolution structure determination We cast the side-chain resonance assignment problem into a Markov Random Field (MRF) framework, and extend and apply combinatorial protein design algorithms to compute the optimal solution that best interprets the NMR data Our MRF framework captures the contact map information of the protein derived from NMR spectra, and exploits the structural information available from the backbone conformations determined by orientational restraints and a set of discretized side-chain conformations (i.e., rotamers) A Hausdorff-based computation is employed in the scoring function to evaluate the probability of side-chain resonance assignments to generate the observed NMR spectra The complexity of the assignment problem is first reduced by using a dead-end elimination (DEE) algorithm, which prunes side-chain resonance assignments that are provably not part of the optimal solution Then an A* search algorithm is used to find a set of optimal side-chain resonance assignments that best fit the NMR data We have tested our algorithm on NMR data for five proteins, including the FF Domain 2 of human transcription elongation factor CA150 (FF2), the B1 domain of Protein G (GB1), human ubiquitin, the ubiquitin-binding zinc finger domain of the human Y-family DNA polymerase Eta (pol η UBZ), and the human Set2-Rpb1 interacting domain (hSRI) Our algorithm assigns resonances for more than 90% of the protons in the proteins, and achieves about 80% correct side-chain resonance assignments The final structures computed using distance restraints resulting from the set of assigned side-chain resonances have backbone RMSD 0.5−1.4 A and all-heavy-atom RMSD 1.0−2.2 A from the reference structures that were determined by X-ray crystallography or traditional NMR approaches These results demonstrate that our algorithm can be successfully applied to automate side-chain resonance assignment and high-quality protein structure determination Since our algorithm does not require any specific NMR experiments for measuring the through-bond interactions with side-chain protons, it can save a significant amount of both experimental cost and spectrometer time, and hence accelerate the NMR structure determination process.

A HAUSDORFF-BASED NOE ASSIGNMENT ALGORITHM USING PROTEIN BACKBONE DETERMINED FROM RESIDUAL DIPOLAR COUPLINGS AND ROTAMER PATTERNS

High-throughput structure determination based on solution Nuclear Magnetic Resonance (NMR) spectroscopy plays an important role in structural genomics. One of the main bottlenecks in NMR structure determination is the interpretation of NMR data to obtain a sufficient number of accurate distance restraints by assigning nuclear Overhauser effect (NOE) spectral peaks to pairs of protons. The difficulty in automated NOE assignment mainly lies in the ambiguities arising both from the resonance degeneracy of chemical shifts and from the uncertainty due to experimental errors in NOE peak positions. In this paper we present a novel NOE assignment algorithm, called HAusdorff-based NOE Assignment (HANA), that starts with a high-resolution protein backbone computed using only two residual dipolar couplings (RDCs) per residue, employs a Hausdorff-based pattern matching technique to deduce similarity between experimental and back-computed NOE spectra for each rotamer from a statistically diverse library, and drives the selection of optimal position-specific rotamers for filtering ambiguous NOE assignments. Our algorithm runs in time O(tn3 + tn log t), where t is the maximum number of rotamers per residue and n is the size of the protein. Application of our algorithm on biological NMR data for three proteins, namely, human ubiquitin, the zinc finger domain of the human DNA Y-polymerase Eta (pol eta) and the human Set2-Rpb1 interacting domain (hSRI) demonstrates that our algorithm overcomes spectral noise to achieve more than 90% assignment accuracy. Additionally, the final structures calculated using our automated NOE assignments have backbone RMSD < 1.7 A and all-heavy-atom RMSD < 2.5 A from reference structures that were determined either by X-ray crystallography or traditional NMR approaches. These results show that our NOE assignment algorithm can be successfully applied to protein NMR spectra to obtain high-quality structures.