PREFER: A New Predictive Modeling Framework for Molecular Discovery.

Machine-learning and deep-learning models have been extensively used in cheminformatics to predict molecular properties, to reduce the need for direct measurements, and to accelerate compound prioritization. However, different setups and frameworks and the large number of molecular representations make it difficult to properly evaluate, reproduce, and compare them. Here we present a new PREdictive modeling FramEwoRk for molecular discovery (PREFER), written in Python (version 3.7.7) and based on AutoSklearn (version 0.14.7), that allows comparison between different molecular representations and common machine-learning models. We provide an overview of the design of our framework and show exemplary use cases and results of several representation-model combinations on diverse data sets, both public and in-house. Finally, we discuss the use of PREFER on small data sets. The code of the framework is freely available on GitHub.

HistoNet: A Deep Learning-Based Model of Normal Histology.

We introduce HistoNet, a deep neural network trained on normal tissue. On 1690 slides with rat tissue samples from 6 preclinical toxicology studies, tissue regions were outlined and annotated by pathologists into 46 different tissue classes. From these annotated regions, we sampled small 224 × 224 pixels images (patches) at 6 different levels of magnification. Using 4 studies as training set and 2 studies as test set, we trained VGG-16, ResNet-50, and Inception-v3 networks separately at each magnification level. Among these model architectures, Inception-v3 and ResNet-50 outperformed VGG-16. Inception-v3 identified the tissue from query images, with an accuracy up to 83.4%. Most misclassifications occurred between histologically similar tissues. Investigation of the features learned by the model (embedding layer) using Uniform Manifold Approximation and Projection revealed not only coherent clusters associated with the individual tissues but also subclusters corresponding to histologically meaningful structures that had not been annotated or trained for. This suggests that the histological representation learned by HistoNet could be useful as the basis of other machine learning algorithms and data mining. Finally, we found that models trained on rat tissues can be used on non-human primate and minipig tissues with minimal retraining.

Free energy and electronic properties of water adsorption on the SnO2(110) surface.

A molecular understanding of the adsorption of water on SnO2 surfaces is crucial for several applications of this metal oxide, including catalysis and gas sensing. We have investigated water adsorption on the SnO2(110) surface using a combination of dynamic and static calculations to gain fundamental insight into the reaction mechanism at room temperature. The reaction dynamics are studied by following water adsorption and dissociation on the SnO2 surface with metadynamics calculations at low and high coverage. The electronic structure in the relevant isolated minima is investigated through Mulliken charge analysis and projected density of states analysis. Surface bridging oxygen (Obr) is found to play a decisive role in water adsorption forming rooted hydroxyl groups with the water H atoms. Bond formation with H significantly changes the electronic configuration of Obr and presumably leads to reduced band bending at the SnO2 surface. The free-energy estimation indicates that on a clean SnO2(110) surface at room temperature both associative and dissociative adsorption occur, with the latter being thermodynamically favored. Oxygen coverage strongly affects the ratio between associatively and dissociatively adsorbed H2O, favoring associative adsorption at high oxygen coverage (oxidized surface) and dissociative adsorption at low oxygen coverage (reduced surface). Electronic analyses of isolated surface minima show the existence of two different electron-transfer phenomena occurring at the surface, depending on the water adsorption mechanism. The relevance of these findings in explaining the changes in electric conductivity occurring in SnO2-based gas sensors upon water adsorption is discussed. Whereas associative adsorption leads to electron enrichment of the metal oxide surface, dissociative adsorption induces surface electron depletion. Both mechanisms are consistent with the electrical conductivity changes occurring upon interaction of SnO2 with water, causing cross sensitivity to the latter. The theoretical results form the basis for correlating the existing atomistic models with the experimental data and offer a coherent description of the reaction events on the surface at room temperature.

First principles analysis of H2O adsorption on the (110) surfaces of SnO2, TiO2 and their solid solutions.

Both associative and dissociative H(2)O adsorption on SnO(2)(110), TiO(2)(110), and Ti-enriched Sn(1-x)Ti(x)O(2)(110) surfaces have been investigated at low ((1)/(12) monolayer (ML)) and high coverage (1 ML) by density functional theory calculations using the Gaussian and plane waves formalism. The use of a large supercell allowed the simulation at low symmetry levels. On SnO(2)(110), dissociative adsorption was favored at all coverages and was accompanied by stable associative H(2)O configurations. Increasing the coverage from (1)/(12) to 1 ML stabilized the (associatively or dissociatively) adsorbed H(2)O on SnO(2)(110) because of the formation of intermolecular H bonds. In contrast, on TiO(2)(110), the adsorption of isolated H(2)O groups ((1)/(12) ML) was more stable than at high coverage, and the favored adsorption changed from dissociative to associative with increasing coverage. For dissociative H(2)O adsorption on Ti-enriched Sn(1-x)Ti(x)O(2)(110) surfaces with Ti atoms preferably located on 6-fold-coordinated surface sites, the analysis of the Wannier centers showed a polarization of electrons surrounding bridging O atoms that were bound simultaneously to 6-fold-coordinated Sn and Ti surface atoms. This polarization suggested the formation of an additional bond between the 6-fold-coordinated Ti(6c) and bridging O atoms that had to be broken upon H(2)O adsorption. As a result, the H(2)O adsorption energy initially decreased, with increasing surface Ti content reaching a minimum at 25% Ti for (1)/(12) ML. This behavior was even more accentuated at high H(2)O coverage (1 ML) with the adsorption energy decreasing rapidly from 145.2 to 101.6 kJ/mol with the surface Ti content increasing from 0 to 33%. A global minimum of binding energies at both low and high coverage was found between 25 and 33% surface Ti content, which may explain the minimal cross-sensitivity to humidity previously reported for Sn(1-x)Ti(x)O(2) gas sensors. Above 12.5% surface Ti content, the binding energy decreased with increasing coverage, suggesting that the partial desorption of H(2)O is facilitated at a high fractional coverage.

Adsorption of naphthalene and quinoline on Pt, Pd and Rh: a DFT study.

The adsorption of naphthalene and quinoline on Pt(111), Pd(111) and Rh(111) surfaces is studied using density functional theory. The metal surfaces are simulated by means of large confined clusters and for Pt by means of a slab with periodic boundary conditions (PBC). Calculation parameters such as basis set convergence, basis set superposition error and effects of cluster relaxation and size are analyzed in order to assess the aptness of the cluster model. For all the metals, the preferred sites of adsorption are analyzed, thus revealing their different behaviors concerning structure and stability of adsorption modes. On Pt, the molecules have the richest theoretical configurational variety. Naphthalene and quinoline are found to adsorb preferentially on di-bridge[7] sites on the three metals, and Rh exhibits higher adsorption energies than Pt and Pd. Structural features of the adsorbed molecules are correlated to the calculated adsorption energies. The di-bridge[7] adsorption modes are studied in deeper detail decomposing the adsorption energies in two terms arising from molecular distortion and binding interaction to the metal. Molecular distortion is correlated to the HOMO-LUMO energy gap. The larger adsorption energies found for interactions with Rh result from the lower contribution of the distortion term. Binding interactions are described by analyzing the wave functions of naphthalene and quinoline adsorbed on a subunit of the large clusters in order to reduce the complexity of the analysis. Molecular orbitals are studied using concepts of Frontier Molecular Orbitals theory. This approach reveals that in the adsorption of naphthalene and quinoline on Pt and Pd, an antibonding state lies below the Fermi energy, while on Rh all antibonding states are empty, in agreement with the larger interaction energies. In addition, further insight is gained by projecting the density of states on the d band of the clean surfaces and of the adsorbed systems. This results in the rationalization of the structural features in terms of the concepts of electronic structure theory. The distributions of electronic density are described by means of Hirshfeld charges and isosurfaces of differential electron density. The net electron transfer from the metals to the molecules for most of the sites correlates with the trends of the adsorption energies.

Modeling bulk and surface Pt using the "Gaussian and plane wave" density functional theory formalism: validation and comparison to k-point plane wave calculations.

We present a study on structural and electronic properties of bulk platinum and the two surfaces (111) and (100) comparing the Gaussian and plane wave method to standard plane wave schemes, normally employed for density functional theory calculations on metallic systems. The aim of this investigation is the assessment of methods based on the expansion of the Kohn-Sham orbitals into localized basis sets and on the supercell approach, in the description of the metallicity of Pt. Electronic structure calculations performed at Gamma-point only on supercells of different sizes, from 108 up to 864 atoms, are compared to the results obtained for the unit cell of four Pt atoms where the k-point expansion of the wave function over Monkhorst-Pack grids up to (10x10x10) has been employed. The evaluation of the two approaches with respect to bulk properties is done through the calculation of the equilibrium lattice constant, the bulk modulus, and the total and the d-projected density of states. For the Pt(111) and Pt(100) surfaces, we consider the relaxation of the first layers, the surface energies, the work function, the total density of states, as well as the center and filling of the d bands. Our results confirm that the accuracy of two approaches in the description of electronic and structural properties of Pt is equivalent, providing that consistent supercells and k-point meshes are used. Moreover, we estimate the supercell size that can be safely adopted in the Gaussian and plane wave method in order to obtain the same reliability of previous theoretical studies based on well converged plane wave calculations available in literature. The latter studies, in turn, set the level of agreement with experimental data. In particular, we obtain excellent agreement in the evaluation of the density of states for either bulk and surface systems, and our data are also in good agreement with previous works on Pt reported in literature. We conclude that Gaussian and plane wave calculations, with simulation cells of 400-800 atoms, can be safely used in the study of chemistry related problems involving transition metal surfaces.

Mutations in the "lid" region affect chain length specificity and thermostability of a Pseudomonas fragi lipase.

The cold-adapted Pseudomonas fragi lipase (PFL) displays highest activity on short-chain triglyceride substrates and is rapidly inactivated at moderate temperature. Sequence and structure comparison with homologous lipases endowed with different substrate specificity and stability, pointed to three polar residues in the lid region, that were replaced with the amino acids conserved at equivalent positions in the reference lipases. Substitutions at residues T137 and T138 modified the lipase chain-length preference profile, increasing the relative activity towards C8 substrates. Moreover, mutations conferred to PFL higher temperature stability. On the other hand, replacement of the serine at position 141 by glycine destabilized the protein.

The cold-active lipase of Pseudomonas fragi. Heterologous expression, biochemical characterization and molecular modeling.

A recombinant lipase cloned from Pseudomonas fragi strain IFO 3458 (PFL) was found to retain significant activity at low temperature. In an attempt to elucidate the structural basis of this behaviour, a model of its three-dimensional structure was built by homology and compared with homologous mesophilic lipases, i.e. the Pseudomonas aeruginosa lipase (45% sequence identity) and Burkholderia cepacia lipase (38%). In this model, features common to all known lipases have been identified, such as the catalytic triad (S83, D238 and H260) and the oxyanion hole (L17, Q84). Structural modifications recurrent in cold-adaptation, i.e. a large amount of charged residues exposed at the protein surface, have been detected. Noteworthy is the lack of a disulphide bridge conserved in homologous Pseudomonas lipases that may contribute to increased conformational flexibility of the cold-active enzyme.