Motivation Implicit solvent models play an important role in describing the thermodynamics and the dynamics of biomolecular systems. Key to an efficient use of these models is the computation of generalized Born (GB) radii, which is accomplished by algorithms based on the electrostatics of inhomogeneous dielectric media. The speed and accuracy of such computations are still an issue especially for their intensive use in classical molecular dynamics. Here, we propose an alternative approach that encodes the physics of the phenomena and the chemical structure of the molecules in model parameters which are learned from examples. Results GB radii have been computed using (i) a linear model and (ii) a neural network. The input is the element, the histogram of counts of neighbouring atoms, divided by atom element, within 16 Å. Linear models are ca. 8 times faster than the most widely used reference method and the accuracy is higher with correlation coefficient with the inverse of ‘perfect’ GB radii of 0.94 versus 0.80 of the reference method. Neural networks further improve the accuracy of the predictions with correlation coefficient with ‘perfect’ GB radii of 0.97 and ca. 20% smaller root mean square error. Availability and implementation We provide a C program implementing the computation using the linear model, including the coefficients appropriate for the set of Bondi radii, as Supplementary Material. We also provide a Python implementation of the neural network model with parameter and example files in the Supplementary Material as well. Supplementary information Supplementary data are available at Bioinformatics online.

Background: Over the last several decades, predicting protein structures from amino acid sequences has been a core task in bioinformatics. Nowadays, the most successful methods employ multiple sequence alignments and can predict the structure with excellent performance. These predictions take advantage of all the amino acids at a given position and their frequencies. However, the effect of single amino acid substitutions in a specific protein tends to be hidden by the alignment profile. For this reason, single-sequence-based predictions attract interest even after accurate multiple-alignment methods have become available: the use of single sequences ensures that the effects of substitution are not confounded by homologous sequences. Objective: This work aims at understanding how the single-sequence secondary structure prediction of a residue is influenced by the surrounding ones. We aim at understanding how different prediction methods use single-sequence information to predict the structure. Methods: We compare mutual information, the coefficients of two linear models, and three deep learning networks. For the deep learning algorithms, we use the DeepLIFT analysis to assess the effect of each residue at each position in the prediction. Result: Mutual information and linear models quantify direct effects, whereas DeepLIFT applied on deep learning networks quantifies both direct and indirect effects Conclusion: Our analysis shows how different network architectures use the information of single protein sequences and highlights their differences with respect to linear models. In particular, the deep learning implementations take into account context and single position information differently, with the best results obtained using the BERT architecture.