pathDIP 4: an extended pathway annotations and enrichment analysis resource for human, model organisms and domesticated species.

PathDIP was introduced to increase proteome coverage of literature-curated human pathway databases. PathDIP 4 now integrates 24 major databases. To further reduce the number of proteins with no curated pathway annotation, pathDIP integrates pathways with physical protein-protein interactions (PPIs) to predict significant physical associations between proteins and curated pathways. For human, it provides pathway annotations for 5366 pathway orphans. Integrated pathway annotation now includes six model organisms and ten domesticated animals. A total of 6401 core and ortholog pathways have been curated from the literature or by annotating orthologs of human proteins in the literature-curated pathways. Extended pathways are the result of combining these pathways with protein-pathway associations that are predicted using organism-specific PPIs. Extended pathways expand proteome coverage from 81 088 to 120 621 proteins, making pathDIP 4 the largest publicly available pathway database for these organisms and providing a necessary platform for comprehensive pathway-enrichment analysis. PathDIP 4 users can customize their search and analysis by selecting organism, identifier and subset of pathways. Enrichment results and detailed annotations for input list can be obtained in different formats and views. To support automated bioinformatics workflows, Java, R and Python APIs are available for batch pathway annotation and enrichment analysis. PathDIP 4 is publicly available at http://ophid.utoronto.ca/pathDIP.

A stemness screen reveals C3orf54/INKA1 as a promoter of human leukemia stem cell latency.

There is a growing body of evidence that the molecular properties of leukemia stem cells (LSCs) are associated with clinical outcomes in acute myeloid leukemia (AML), and LSCs have been linked to therapy failure and relapse. Thus, a better understanding of the molecular mechanisms that contribute to the persistence and regenerative potential of LSCs is expected to result in the development of more effective therapies. We therefore interrogated functionally validated data sets of LSC-specific genes together with their known protein interactors and selected 64 candidates for a competitive in vivo gain-of-function screen to identify genes that enhanced stemness in human cord blood hematopoietic stem and progenitor cells. A consistent effect observed for the top hits was the ability to restrain early repopulation kinetics while preserving regenerative potential. Overexpression (OE) of the most promising candidate, the orphan gene C3orf54/INKA1, in a patient-derived AML model (8227) promoted the retention of LSCs in a primitive state manifested by relative expansion of CD34+ cells, accumulation of cells in G0, and reduced output of differentiated progeny. Despite delayed early repopulation, at later times, INKA1-OE resulted in the expansion of self-renewing LSCs. In contrast, INKA1 silencing in primary AML reduced regenerative potential. Mechanistically, our multidimensional confocal analysis found that INKA1 regulates G0 exit by interfering with nuclear localization of its target PAK4, with concomitant reduction of global H4K16ac levels. These data identify INKA1 as a novel regulator of LSC latency and reveal a link between the regulation of stem cell kinetics and pool size during regeneration.

In silico prediction of physical protein interactions and characterization of interactome orphans.

Protein-protein interactions (PPIs) are useful for understanding signaling cascades, predicting protein function, associating proteins with disease and fathoming drug mechanism of action. Currently, only ∼ 10% of human PPIs may be known, and about one-third of human proteins have no known interactions. We introduce FpClass, a data mining-based method for proteome-wide PPI prediction. At an estimated false discovery rate of 60%, we predicted 250,498 PPIs among 10,531 human proteins; 10,647 PPIs involved 1,089 proteins without known interactions. We experimentally tested 233 high- and medium-confidence predictions and validated 137 interactions, including seven novel putative interactors of the tumor suppressor p53. Compared to previous PPI prediction methods, FpClass achieved better agreement with experimentally detected PPIs. We provide an online database of annotated PPI predictions (http://ophid.utoronto.ca/fpclass/) and the prediction software (http://www.cs.utoronto.ca/~juris/data/fpclass/).

Protein crystallization analysis on the World Community Grid.

We have developed an image-analysis and classification system for automatically scoring images from high-throughput protein crystallization trials. Image analysis for this system is performed by the Help Conquer Cancer (HCC) project on the World Community Grid. HCC calculates 12,375 distinct image features on microbatch-under-oil images from the Hauptman-Woodward Medical Research Institute's High-Throughput Screening Laboratory. Using HCC-computed image features and a massive training set of 165,351 hand-scored images, we have trained multiple Random Forest classifiers that accurately recognize multiple crystallization outcomes, including crystals, clear drops, precipitate, and others. The system successfully recognizes 80% of crystal-bearing images, 89% of precipitate images, and 98% of clear drops.

Establishing a training set through the visual analysis of crystallization trials. Part I: approximately 150,000 images.

Structural crystallography aims to provide a three-dimensional representation of macromolecules. Many parts of the multistep process to produce the three-dimensional structural model have been automated, especially through various structural genomics projects. A key step is the production of crystals for diffraction. The target macromolecule is combined with a large and chemically diverse set of cocktails with some leading ideally, but infrequently, to crystallization. A variety of outcomes will be observed during these screening experiments that typically require human interpretation for classification. Human interpretation is neither scalable nor objective, highlighting the need to develop an automatic computer-based image classification. As a first step towards automated image classification, 147,456 images representing crystallization experiments from 96 different macromolecular samples were manually classified. Each image was classified by three experts into seven predefined categories or their combinations. The resulting data where all three observers are in agreement provides one component of a truth set for the development and rigorous testing of automated image-classification systems and provides information about the chemical cocktails used for crystallization. In this paper, the details of this study are presented.

Establishing a training set through the visual analysis of crystallization trials. Part II: crystal examples.

In the automated image analysis of crystallization experiments, representative examples of outcomes can be obtained rapidly. However, while the outcomes appear to be diverse, the number of crystalline outcomes can be small. To complement a training set from the visual observation of 147 456 crystallization outcomes, a set of crystal images was produced from 106 and 163 macromolecules under study for the North East Structural Genomics Consortium (NESG) and Structural Genomics of Pathogenic Protozoa (SGPP) groups, respectively. These crystal images have been combined with the initial training set. A description of the crystal-enriched data set and a preliminary analysis of outcomes from the data are described.

Automatic classification and pattern discovery in high-throughput protein crystallization trials.

Conceptually, protein crystallization can be divided into two phases search and optimization. Robotic protein crystallization screening can speed up the search phase, and has a potential to increase process quality. Automated image classification helps to increase throughput and consistently generate objective results. Although the classification accuracy can always be improved, our image analysis system can classify images from 1,536-well plates with high classification accuracy (85%) and ROC score (0.87), as evaluated on 127 human-classified protein screens containing 5,600 crystal images and 189,472 non-crystal images. Data mining can integrate results from high-throughput screens with information about crystallizing conditions, intrinsic protein properties, and results from crystallization optimization. We apply association mining, a data mining approach that identifies frequently occurring patterns among variables and their values. This approach segregates proteins into groups based on how they react in a broad range of conditions, and clusters cocktails to reflect their potential to achieve crystallization. These results may lead to crystallization screen optimization, and reveal associations between protein properties and crystallization conditions. We also postulate that past experience may lead us to the identification of initial conditions favorable to crystallization for novel proteins.

Automatic classification of sub-microlitre protein-crystallization trials in 1536-well plates.

A technique for automatically evaluating microbatch (400 nl) protein-crystallization trials is described. This method addresses analysis problems introduced at the sub-microlitre scale, including non-uniform lighting and irregular droplet boundaries. The droplet is segmented from the well using a loopy probabilistic graphical model with a two-layered grid topology. A vector of 23 features is extracted from the droplet image using the Radon transform for straight-edge features and a bank of correlation filters for microcrystalline features. Image classification is achieved by linear discriminant analysis of its feature vector. The results of the automatic method are compared with those of a human expert on 32 1536-well plates. Using the human-labeled images as ground truth, this method classifies images with 85% accuracy and a ROC score of 0.84. This result compares well with the experimental repeatability rate, assessed at 87%. Images falsely classified as crystal-positive variously contain speckled precipitate resembling microcrystals, skin effects or genuine crystals falsely labeled by the human expert. Many images falsely classified as crystal-negative variously contain very fine crystal features or dendrites lacking straight edges. Characterization of these misclassifications suggests directions for improving the method.