A comparison of classical and machine learning-based phenotype prediction methods on simulated data and three plant species.

Genomic selection is an integral tool for breeders to accurately select plants directly from genotype data leading to faster and more resource-efficient breeding programs. Several prediction methods have been established in the last few years. These range from classical linear mixed models to complex non-linear machine learning approaches, such as Support Vector Regression, and modern deep learning-based architectures. Many of these methods have been extensively evaluated on different crop species with varying outcomes. In this work, our aim is to systematically compare 12 different phenotype prediction models, including basic genomic selection methods to more advanced deep learning-based techniques. More importantly, we assess the performance of these models on simulated phenotype data as well as on real-world data from Arabidopsis thaliana and two breeding datasets from soy and corn. The synthetic phenotypic data allow us to analyze all prediction models and especially the selected markers under controlled and predefined settings. We show that Bayes B and linear regression models with sparsity constraints perform best under different simulation settings with respect to explained variance. Further, we can confirm results from other studies that there is no superiority of more complex neural network-based architectures for phenotype prediction compared to well-established methods. However, on real-world data, for which several prediction models yield comparable results with slight advantages for Elastic Net, this picture is less clear, suggesting that there is a lot of room for future research.

Molecular engineering of organophosphate hydrolysis activity from a weak promiscuous lactonase template.

Rapid evolution of enzymes provides unique molecular insights into the remarkable adaptability of proteins and helps to elucidate the relationship between amino acid sequence, structure, and function. We interrogated the evolution of the phosphotriesterase from Pseudomonas diminuta (PdPTE), which hydrolyzes synthetic organophosphates with remarkable catalytic efficiency. PTE is thought to be an evolutionarily "young" enzyme, and it has been postulated that it has evolved from members of the phosphotriesterase-like lactonase (PLL) family that show promiscuous organophosphate-degrading activity. Starting from a weakly promiscuous PLL scaffold (Dr0930 from Deinococcus radiodurans ), we designed an extremely efficient organophosphate hydrolase (OPH) with broad substrate specificity using rational and random mutagenesis in combination with in vitro activity screening. The OPH activity for seven organophosphate substrates was simultaneously enhanced by up to 5 orders of magnitude, achieving absolute values of catalytic efficiencies up to 10(6) M(-1) s(-1). Structural and computational analyses identified the molecular basis for the enhanced OPH activity of the engineered PLL variants and demonstrated that OPH catalysis in PdPTE and the engineered PLL differ significantly in the mode of substrate binding.

Binding pocket optimization by computational protein design.

Engineering specific interactions between proteins and small molecules is extremely useful for biological studies, as these interactions are essential for molecular recognition. Furthermore, many biotechnological applications are made possible by such an engineering approach, ranging from biosensors to the design of custom enzyme catalysts. Here, we present a novel method for the computational design of protein-small ligand binding named PocketOptimizer. The program can be used to modify protein binding pocket residues to improve or establish binding of a small molecule. It is a modular pipeline based on a number of customizable molecular modeling tools to predict mutations that alter the affinity of a target protein to its ligand. At its heart it uses a receptor-ligand scoring function to estimate the binding free energy between protein and ligand. We compiled a benchmark set that we used to systematically assess the performance of our method. It consists of proteins for which mutational variants with different binding affinities for their ligands and experimentally determined structures exist. Within this test set PocketOptimizer correctly predicts the mutant with the higher affinity in about 69% of the cases. A detailed analysis of the results reveals that the strengths of PocketOptimizer lie in the correct introduction of stabilizing hydrogen bonds to the ligand, as well as in the improved geometric complemetarity between ligand and binding pocket. Apart from the novel method for binding pocket design we also introduce a much needed benchmark data set for the comparison of affinities of mutant binding pockets, and that we use to asses programs for in silico design of ligand binding.

Automated scaffold selection for enzyme design.

A major goal of computational protein design is the construction of novel functions on existing protein scaffolds. There the first question is which scaffold is suitable for a specific reaction. Given a set of catalytic residues and their spatial arrangement, one wants to identify a protein scaffold that can host this active site. Here, we present an algorithm called ScaffoldSelection that is able to rapidly search large sets of protein structures for potential attachment sites of an enzymatic motif. The method consists of two steps; it first identifies pairs of backbone positions in pocket-like regions. Then, it combines these to complete attachment sites using a graph theoretical approach. Identified matches are assessed for their ability to accommodate the substrate or transition state. A representative set of structures from the Protein Data Bank ( approximately 3500) was searched for backbone geometries that support the catalytic residues for 12 chemical reactions. Recapitulation of native active site geometries is used as a benchmark for the performance of the program. The native motif is identified in all 12 test cases, ranking it in the top percentile in 5 out of 12. The algorithm is fast and efficient, although dependent on the complexity of the motif. Comparisons to other methods show that ScaffoldSelection performs equally well in terms of accuracy and far better in terms of speed. Thus, ScaffoldSelection will aid future computational protein design experiments by preselecting protein scaffolds that are suitable for a specific reaction type and the introduction of a predefined amino acid motif.

Establishing wild-type levels of catalytic activity on natural and artificial (beta alpha)8-barrel protein scaffolds.

The generation of high levels of new catalytic activities on natural and artificial protein scaffolds is a major goal of enzyme engineering. Here, we used random mutagenesis and selection in vivo to establish a sugar isomerisation reaction on both a natural (beta alpha)(8)-barrel enzyme and a catalytically inert chimeric (beta alpha)(8)-barrel scaffold, which was generated by the recombination of 2 (beta alpha)(4)-half barrels. The best evolved variants show turnover numbers and substrate affinities that are similar to those of wild-type enzymes catalyzing the same reaction. The determination of the crystal structure of the most proficient variant allowed us to model the substrate sugar in the novel active site and to elucidate the mechanistic basis of the newly established activity. The results demonstrate that natural and inert artificial protein scaffolds can be converted into highly proficient enzymes in the laboratory, and provide insights into the mechanisms of enzyme evolution.