Learning dataset representation for automatic machine learning algorithm selection

The algorithm selection problem is defined as identifying the best-performing machine learning (ML) algorithm for a given combination of dataset, task, and evaluation measure. The human expertise required to evaluate the increasing number of ML algorithms available has resulted in the need to automate the algorithm selection task. Various approaches have emerged to handle the automatic algorithm selection challenge, including meta-learning. Meta-learning is a popular approach that leverages accumulated experience for future learning and typically involves dataset characterization. Existing meta-learning methods often represent a dataset using predefined features and thus cannot be generalized across different ML tasks, or alternatively, learn a dataset’s representation in a supervised manner and therefore are unable to deal with unsupervised tasks. In this study, we propose a novel learning-based task-agnostic method for producing dataset representations. Then, we introduce TRIO, a meta-learning approach, that utilizes the proposed dataset representations to accurately recommend top-performing algorithms for previously unseen datasets. TRIO first learns graphical representations for the datasets, using four tools to learn the latent interactions among dataset instances and then utilizes a graph convolutional neural network technique to extract embedding representations from the graphs obtained. We extensively evaluate the effectiveness of our approach on 337 datasets and 195 ML algorithms, demonstrating that TRIO significantly outperforms state-of-the-art methods for algorithm selection for both supervised (classification and regression) and unsupervised (clustering) tasks.

     

Translating the Foundational Model of Anatomy into OWL

The Foundational Model of Anatomy (FMA) represents the result of manual and disciplined modeling of the structural organization of the human body. It is a tremendous resource in bioinformatics that facilitates sharing of information among applications that use anatomy knowledge. The FMA was developed in Protégé and the Protégé frames language is the canonical representation language for the FMA. We present a translation of the original Protégé frame representation of the FMA into OWL. Our effort is complementary to the earlier efforts to represent FMA in OWL and is focused on two main goals: (1) representing only the information that is explicitly present in the frames representation of the FMA or that can be directly inferred from the semantics of Protégé frames; (2) representing all the information that is present in the frames representation of the FMA, thus producing an OWL representation for the complete FMA. Our complete representation of the FMA in OWL consists of two components: an OWL DL component that contains the FMA constructs that are compatible with OWL DL; and an OWL Full component that imports the OWL DL component and adds the FMA constructs that OWL DL does not allow.     

The PROMPT suite: interactive tools for ontology merging and mapping

Researchers in the ontology-design field have developed the content for ontologies in many domain areas. This distributed nature of ontology development has led to a large number of ontologies covering overlapping domains. In order for these ontologies to be reused, they first need to be merged or aligned to one another. We developed a suite of tools for managing multiple ontologies. These suite provides users with a uniform framework for comparing, aligning, and merging ontologies, maintaining versions, translating between different formalisms. Two of the tools in the suite support semi-automatic ontology merging: P is an interactive ontology-merging tool that guides the user through the merging process, presenting him with suggestions for next steps and identifying inconsistencies and potential problems. A P uses a graph structure of ontologies to find correlation between concepts and to provide additional information for P .     

Visualizing Objects with Mirrors

We present an automatic camera and mirrors placement method for visualizing 3D scenes such that complete and nonredundant visibility is guaranteed, as well as highest image resolution. We provide algorithms and geometries for optimal object‐independent mirrors and viewpoint placement for all objects in a specific class. We also briefly consider object‐dependent algorithmic problems, where mirrors and viewpoints are positioned as a function of the object.     

Cellular retinaldehyde-binding protein ligand interactions. Gln-210 and Lys-221 are in the retinoid binding pocket

Cellular retinaldehyde-binding protein (CRALBP) carries 11-cis-retinal and/or 11-cis-retinol as endogenous ligands in the retinal pigment epithelium (RPE) and Müller cells of the retina and has been linked with autosomal recessive retinitis pigmentosa. Ligand interactions determine the physiological role of CRALBP in the RPE where the protein is thought to function as a substrate carrier for 11-cis-retinol dehydrogenase in the synthesis of 11-cis-retinal for visual pigment regeneration. However, CRALBP is also present in optic nerve and brain where its natural ligand and function are not yet known. We have characterized the interactions of retinoids with native bovine CRALBP, human recombinant CRALBP (rCRALBP) and five mutant rCRALBPs. Efforts to trap and/or identify a Schiff base in the dark, under a variety of reducing, denaturing, and pH conditions were unsuccessful, suggesting the lack of covalent interactions between CRALBP and retinoid. Buried and solvent-exposed lysine residues were identified in bovine CRALBP by reductive methylation of the holoprotein followed by denaturation and reaction with [3H]acetic anhydride. Radioactive lysine residues were identified by Edman degradation and electrospray mass spectrometry following proteolysis and purification of modified peptides. Human rCRALBP mutants K152A, K221A, and K294A were prepared to investigate possible retinoid interactions with buried or partially buried lysines. Two other rCRALBP mutants, I162V and Q210R, were also prepared to identify substitutions altering the retinoid binding properties of a random mutant. The structures of all the mutants were verified by amino acid and mass spectral analyses and retinoid binding properties evaluated by UV-visible and fluorescence spectroscopy. All of the mutants bound 11-cis-retinal essentially like the wild type protein, indicating that the proteins were not grossly misfolded. Three of the mutants bound 9-cis-retinal like the wild type protein; however, Q210R and K221A bound less than stoichiometric amounts of the 9-cis-isomer and exhibited lower affinity for this retinoid relative to wild type rCRALBP. Residues Gln-210 and Lys-221 are located within a region of CRALBP exhibiting sequence homology with the ligand binding cavity of yeast phosphatidylinositol-transfer protein. The data implicate Gln-210 and Lys-221 as components of the CRALBP retinoid binding cavity and are discussed in the context of ligand interactions in structurally or functionally related proteins with known crystallographic structures.     

Amphiphilic four-helix bundle peptides designed for light-induced electron transfer across a soft interface

A family of four-helix bundle peptides were designed to be amphiphilic, possessing distinct hydrophilic and hydrophobic domains along the length of the bundle's exterior. This facilitates their vectorial insertion across a soft interface between polar and nonpolar media. Their design also now provides for selective incorporation of electron donor and acceptor cofactors within each domain. This allows translation of the designed intramolecular electron transfer along the bundle axis into a macroscopic charge separation across the interface.     

Fabrication of very high aspect ratio metal nanowires by a self-propulsion mechanism

A novel synthesis method of very high aspect ratio metal nanowires is described. The synthesis utilizes a nanoporous membrane as a template and self-electrophoresis as a directed force that continuously push formed nanowires out of the pores in a rate that is identical to the rate of their elongation. As a result, while the pores of membranes are only 6 microm long, the formed nanowires could be more than 100 microm long.     

Single functional group interactions with individual carbon nanotubes

Carbon nanotubes display a consummate blend of materials properties that affect applications ranging from nanoelectronic circuits and biosensors to field emitters and membranes. These applications use the non-covalent interactions between the nanotubes and chemical functionalities, often involving a few molecules at a time. Despite their wide use, we still lack a fundamental understanding and molecular-level control of these interactions. We have used chemical force microscopy to measure the strength of the interactions of single chemical functional groups with the sidewalls of vapour-grown individual single-walled carbon nanotubes. Surprisingly, the interaction strength does not follow conventional trends of increasing polarity or hydrophobicity, and instead reflects the complex electronic interactions between the nanotube and the functional group. Ab initio calculations confirm the observed trends and predict binding force distributions for a single molecular contact that match the experimental results. Our analysis also reveals the important role of molecular linkage dynamics in determining interaction strength at the single functional group level.     

High-resolution ab initio three-dimensional x-ray diffraction microscopy

Coherent x-ray diffraction microscopy is a method of imaging nonperiodic isolated objects at resolutions limited, in principle, by only the wavelength and largest scattering angles recorded. We demonstrate x-ray diffraction imaging with high resolution in all three dimensions, as determined by a quantitative analysis of the reconstructed volume images. These images are retrieved from the three-dimensional diffraction data using no a priori knowledge about the shape or composition of the object, which has never before been demonstrated on a nonperiodic object. We also construct two-dimensional images of thick objects with greatly increased depth of focus (without loss of transverse spatial resolution). These methods can be used to image biological and materials science samples at high resolution with x-ray undulator radiation and establishes the techniques to be used in atomic-resolution ultrafast imaging at x-ray free-electron laser sources.