The RSNA Pediatric Bone Age Machine Learning Challenge.

Purpose The Radiological Society of North America (RSNA) Pediatric Bone Age Machine Learning Challenge was created to show an application of machine learning (ML) and artificial intelligence (AI) in medical imaging, promote collaboration to catalyze AI model creation, and identify innovators in medical imaging. Materials and Methods The goal of this challenge was to solicit individuals and teams to create an algorithm or model using ML techniques that would accurately determine skeletal age in a curated data set of pediatric hand radiographs. The primary evaluation measure was the mean absolute distance (MAD) in months, which was calculated as the mean of the absolute values of the difference between the model estimates and those of the reference standard, bone age. Results A data set consisting of 14 236 hand radiographs (12 611 training set, 1425 validation set, 200 test set) was made available to registered challenge participants. A total of 260 individuals or teams registered on the Challenge website. A total of 105 submissions were uploaded from 48 unique users during the training, validation, and test phases. Almost all methods used deep neural network techniques based on one or more convolutional neural networks (CNNs). The best five results based on MAD were 4.2, 4.4, 4.4, 4.5, and 4.5 months, respectively. Conclusion The RSNA Pediatric Bone Age Machine Learning Challenge showed how a coordinated approach to solving a medical imaging problem can be successfully conducted. Future ML challenges will catalyze collaboration and development of ML tools and methods that can potentially improve diagnostic accuracy and patient care. © RSNA, 2018 Online supplemental material is available for this article. See also the editorial by Siegel in this issue.

Protein-ligand docking using FFT based sampling: D3R case study.

Fast Fourier transform (FFT) based approaches have been successful in application to modeling of relatively rigid protein-protein complexes. Recently, we have been able to adapt the FFT methodology to treatment of flexible protein-peptide interactions. Here, we report our latest attempt to expand the capabilities of the FFT approach to treatment of flexible protein-ligand interactions in application to the D3R PL-2016-1 challenge. Based on the D3R assessment, our FFT approach in conjunction with Monte Carlo minimization off-grid refinement was among the top performing methods in the challenge. The potential advantage of our method is its ability to globally sample the protein-ligand interaction landscape, which will be explored in further applications.

PROSTATEx Challenges for computerized classification of prostate lesions from multiparametric magnetic resonance images.

Grand challenges stimulate advances within the medical imaging research community; within a competitive yet friendly environment, they allow for a direct comparison of algorithms through a well-defined, centralized infrastructure. The tasks of the two-part PROSTATEx Challenges (the PROSTATEx Challenge and the PROSTATEx-2 Challenge) are (1) the computerized classification of clinically significant prostate lesions and (2) the computerized determination of Gleason Grade Group in prostate cancer, both based on multiparametric magnetic resonance images. The challenges incorporate well-vetted cases for training and testing, a centralized performance assessment process to evaluate results, and an established infrastructure for case dissemination, communication, and result submission. In the PROSTATEx Challenge, 32 groups apply their computerized methods (71 methods total) to 208 prostate lesions in the test set. The area under the receiver operating characteristic curve for these methods in the task of differentiating between lesions that are and are not clinically significant ranged from 0.45 to 0.87; statistically significant differences in performance among the top-performing methods, however, are not observed. In the PROSTATEx-2 Challenge, 21 groups apply their computerized methods (43 methods total) to 70 prostate lesions in the test set. When compared with the reference standard, the quadratic-weighted kappa values for these methods in the task of assigning a five-point Gleason Grade Group to each lesion range from - 0.24 to 0.27; superiority to random guessing can be established for only two methods. When approached with a sense of commitment and scientific rigor, challenges foster interest in the designated task and encourage innovation in the field.

Focused grid-based resampling for protein docking and mapping.

The fast Fourier transform (FFT) sampling algorithm has been used with success in application to protein-protein docking and for protein mapping, the latter docking a variety of small organic molecules for the identification of binding hot spots on the target protein. Here we explore the local rather than global usage of the FFT sampling approach in docking applications. If the global FFT based search yields a near-native cluster of docked structures for a protein complex, then focused resampling of the cluster generally leads to a substantial increase in the number of conformations close to the native structure. In protein mapping, focused resampling of the selected hot spot regions generally reveals further hot spots that, while not as strong as the primary hot spots, also contribute to ligand binding. The detection of additional ligand binding regions is shown by the improved overlap between hot spots and bound ligands.

Accounting for observed small angle X-ray scattering profile in the protein-protein docking server ClusPro.

The protein-protein docking server ClusPro is used by thousands of laboratories, and models built by the server have been reported in over 300 publications. Although the structures generated by the docking include near-native ones for many proteins, selecting the best model is difficult due to the uncertainty in scoring. Small angle X-ray scattering (SAXS) is an experimental technique for obtaining low resolution structural information in solution. While not sufficient on its own to uniquely predict complex structures, accounting for SAXS data improves the ranking of models and facilitates the identification of the most accurate structure. Although SAXS profiles are currently available only for a small number of complexes, due to its simplicity the method is becoming increasingly popular. Since combining docking with SAXS experiments will provide a viable strategy for fairly high-throughput determination of protein complex structures, the option of using SAXS restraints is added to the ClusPro server. © 2015 Wiley Periodicals, Inc.

The impact of side-chain packing on protein docking refinement.

We study the impact of optimizing the side-chain positions in the interface region between two proteins during the process of binding. Mathematically, the problem is similar to side-chain prediction, which has been extensively explored in the process of protein structure prediction. The protein-protein docking application, however, has a number of characteristics that necessitate different algorithmic and implementation choices. In this work, we implement a distributed approximate algorithm that can be implemented on multiprocessor architectures and enables a trade-off between accuracy and running speed. We report computational results on benchmarks of enzyme-inhibitor and other types of complexes, establishing that the side-chain flexibility our algorithm introduces substantially improves the performance of docking protocols. Furthermore, we establish that the inclusion of unbound side-chain conformers in the side-chain positioning problem is critical in these performance improvements. The code is available to the community under open source license.

Insights into the architecture of the eIF2Bα/ß/δ regulatory subcomplex.

Eukaryotic translation initiation factor 2B (eIF2B), the guanine nucleotide exchange factor for the G-protein eIF2, is one of the main targets for the regulation of protein synthesis. The eIF2B activity is inhibited in response to a wide range of stress factors and diseases, including viral infections, hypoxia, nutrient starvation, and heme deficiency, collectively known as the integrated stress response. eIF2B has five subunits (α-ε). The α, ß, and δ subunits are homologous to each other and form the eIF2B regulatory subcomplex, which is believed to be a trimer consisting of monomeric α, ß, and δ subunits. Here we use a combination of biophysical methods, site-directed mutagenesis, and bioinformatics to show that the human eIF2Bα subunit is in fact a homodimer, at odds with the current trimeric model for the eIF2Bα/ß/δ regulatory complex. eIF2Bα dimerizes using the same interface that is found in the homodimeric archaeal eIF2Bα/ß/δ homolog aIF2B and related metabolic enzymes. We also present evidence that the eIF2Bß/δ binding interface is similar to that in the eIF2Bα2 homodimer. Mutations at the predicted eIF2Bß/δ dimer interface cause genetic neurological disorders in humans. We propose that the eIF2B regulatory subcomplex is an α2ß2δ2 hexamer, composed of one α2 homodimer and two ßδ heterodimers. Our results offer novel insights into the architecture of eIF2B and its interactions with the G-protein eIF2.

Tunable, mixed-resolution modeling using library-based Monte Carlo and graphics processing units.

Building on our recently introduced library-based Monte Carlo (LBMC) approach, we describe a flexible protocol for mixed coarse-grained (CG)/all-atom (AA) simulation of proteins and ligands. In the present implementation of LBMC, protein side chain configurations are pre-calculated and stored in libraries, while bonded interactions along the backbone are treated explicitly. Because the AA side chain coordinates are maintained at minimal run-time cost, arbitrary sites and interaction terms can be turned on to create mixed-resolution models. For example, an AA region of interest such as a binding site can be coupled to a CG model for the rest of the protein. We have additionally developed a hybrid implementation of the generalized Born/surface area (GBSA) implicit solvent model suitable for mixed-resolution models, which in turn was ported to a graphics processing unit (GPU) for faster calculation. The new software was applied to study two systems: (i) the behavior of spin labels on the B1 domain of protein G (GB1) and (ii) docking of randomly initialized estradiol configurations to the ligand binding domain of the estrogen receptor (ERα). The performance of the GPU version of the code was also benchmarked in a number of additional systems.

A Message Passing Approach to Side Chain Positioning with Applications in Protein Docking Refinement.

We introduce a message-passing algorithm to solve the Side Chain Positioning (SCP) problem. SCP is a crucial component of protein docking refinement, which is a key step of an important class of problems in computational structural biology called protein docking. We model SCP as a combinatorial optimization problem and formulate it as a Maximum Weighted Independent Set (MWIS) problem. We then employ a modified and convergent belief-propagation algorithm to solve a relaxation of MWIS and develop randomized estimation heuristics that use the relaxed solution to obtain an effective MWIS feasible solution. Using a benchmark set of protein complexes we demonstrate that our approach leads to more accurate docking predictions compared to a baseline algorithm that does not solve the SCP.

Extending fragment-based free energy calculations with library Monte Carlo simulation: annealing in interaction space.

Pre-calculated libraries of molecular fragment configurations have previously been used as a basis for both equilibrium sampling (via library-based Monte Carlo) and for obtaining absolute free energies using a polymer-growth formalism. Here, we combine the two approaches to extend the size of systems for which free energies can be calculated. We study a series of all-atom poly-alanine systems in a simple dielectric solvent and find that precise free energies can be obtained rapidly. For instance, for 12 residues, less than an hour of single-processor time is required. The combined approach is formally equivalent to the annealed importance sampling algorithm; instead of annealing by decreasing temperature, however, interactions among fragments are gradually added as the molecule is grown. We discuss implications for future binding affinity calculations in which a ligand is grown into a binding site.

Rapid sampling of all-atom peptides using a library-based polymer-growth approach.

We adapted existing polymer growth strategies for equilibrium sampling of peptides described by modern atomistic forcefields with a simple uniform dielectric solvent. The main novel feature of our approach is the use of precalculated statistical libraries of molecular fragments. A molecule is sampled by combining fragment configurations-of single residues in this study-which are stored in the libraries. Ensembles generated from the independent libraries are reweighted to conform with the Boltzmann-factor distribution of the forcefield describing the full molecule. In this way, high-quality equilibrium sampling of small peptides (4-8 residues) typically requires less than one hour of single-processor wallclock time and can be significantly faster than Langevin simulations. Furthermore, approximate, clash-free ensembles can be generated for larger peptides (up to 32 residues in this study) in less than a minute of single-processor computing. We discuss possible applications of our growth procedure to free energy calculation, fragment assembly protein-structure prediction protocols, and to "multi-resolution" sampling.

Efficient equilibrium sampling of all-atom peptides using library-based Monte Carlo.

We applied our previously developed library-based Monte Carlo (LBMC) to equilibrium sampling of several implicitly solvated all-atom peptides. LBMC can perform equilibrium sampling of molecules using precalculated statistical libraries of molecular-fragment configurations and energies. For this study, we employed residue-based fragments distributed according to the Boltzmann factor of the optimized potential for liquid simulations all-atom (OPLS-AA) forcefield describing the individual fragments. Two solvent models were employed: a simple uniform dielectric and the generalized Born/surface area (GBSA) model. The efficiency of LBMC was compared to standard Langevin dynamics (LD) using three different statistical tools. The statistical analyses indicate that LBMC is more than 100 times faster than LD not only for the simple solvent model but also for GBSA.

General library-based Monte Carlo technique enables equilibrium sampling of semi-atomistic protein models.

We introduce "library-based Monte Carlo" (LBMC) simulation, which performs Boltzmann sampling of molecular systems based on precalculated statistical libraries of molecular-fragment configurations, energies, and interactions. The library for each fragment can be Boltzmann distributed and thus account for all correlations internal to the fragment. LBMC can be applied to both atomistic and coarse-grained models, as we demonstrate in this "proof-of-principle" report. We first verify the approach in a toy model and in implicitly solvated all-atom polyalanine systems. We next study five proteins, up to 309 residues in size. On the basis of atomistic equilibrium libraries of peptide-plane configurations, the proteins are modeled with fully atomistic backbones and simplified Go-like interactions among residues. We show that full equilibrium sampling can be obtained in days to weeks on a single processor, suggesting that more accurate models are well within reach. For the future, LBMC provides a convenient platform for constructing adjustable or mixed-resolution models: the configurations of all atoms can be stored at no run-time cost, while an arbitrary subset of interactions is "turned on".

Absolute free energies estimated by combining precalculated molecular fragment libraries.

The absolute free energy--or partition function, equivalently--of a molecule can be estimated computationally using a suitable reference system. Here, we demonstrate a practical method for staging such calculations by growing a molecule based on a series of fragments. Significant computer time is saved by precalculating fragment configurations and interactions for reuse in a variety of molecules. We use such fragment libraries and interaction tables for amino acids and capping groups to estimate free energies for small peptides. Equilibrium ensembles for the molecules are generated at no additional computational cost and are used to check our results by comparison to standard dynamics simulation. We explain how our work can be extended to estimate relative binding affinities.

Water and deuterium oxide permeability through aquaporin 1: MD predictions and experimental verification.

Determining the mechanisms of flux through protein channels requires a combination of structural data, permeability measurement, and molecular dynamics (MD) simulations. To further clarify the mechanism of flux through aquaporin 1 (AQP1), osmotic p(f) (cm(3)/s/pore) and diffusion p(d) (cm(3)/s/pore) permeability coefficients per pore of H(2)O and D(2)O in AQP1 were calculated using MD simulations. We then compared the simulation results with experimental measurements of the osmotic AQP1 permeabilities of H(2)O and D(2)O. In this manner we evaluated the ability of MD simulations to predict actual flux results. For the MD simulations, the force field parameters of the D(2)O model were reparameterized from the TIP3P water model to reproduce the experimentally observed difference in the bulk self diffusion constants of H(2)O vs. D(2)O. Two MD systems (one for each solvent) were constructed, each containing explicit palmitoyl-oleoyl-phosphatidyl-ethanolamine (POPE) phospholipid molecules, solvent, and AQP1. It was found that the calculated value of p(f) for D(2)O is approximately 15% smaller than for H(2)O. Bovine AQP1 was reconstituted into palmitoyl-oleoyl-phosphatidylcholine (POPC) liposomes, and it was found that the measured macroscopic osmotic permeability coefficient P(f) (cm/s) of D(2)O is approximately 21% lower than for H(2)O. The combined computational and experimental results suggest that deuterium oxide permeability through AQP1 is similar to that of water. The slightly lower observed osmotic permeability of D(2)O compared to H(2)O in AQP1 is most likely due to the lower self diffusion constant of D(2)O.

Modeling the fast gating mechanism in the ClC-0 chloride channel.

A simplified three-dimensional model ClC-0 chloride channel is constructed to couple the permeation of Cl- ions to the motion of a glutamate side chain that acts as the putative fast gate in the ClC-0 channel. The gate is treated as a single spherical particle attached by a rod to a pivot point. This particle moves in a one-dimensional arc under the influence of a bistable potential, which mimics the isomerization process by which the glutamate side chain moves from an open state (not blocking the channel pore) to a closed state (blocking the channel pore, at a position which also acts as a binding site for Cl- ions moving through the channel). A dynamic Monte Carlo (DMC) technique is utilized to perform Brownian dynamics simulations to investigate the dependence of the gate closing rate on both internal and external chloride concentration and the gate charge as well. To accelerate the simulation of gate closing to a time scale that can be accommodated with current methodology and computer power, namely, microseconds, parameters that govern the motion of the bare gate (i.e., in the absence of coupling to the permeating ions) are chosen appropriately. Our simulation results are in qualitative agreement with experimental observations and consistent with the "foot-in-the-door" mechanism (Chen et al. J. Gen. Physiol. 2003, 122, 641; Chen and Miller J. Gen. Physiol. 1996, 108, 237), although the absolute time scale of gate closing in the real channel is much longer (millisecond time scale). A simple model based on the fractional occupation probability of the Cl- binding site that is ultimately blocked by the fast gate suggests straightforward scalability of simulation results for the model channel considered herein to experimentally realistic time scales.

Diffusion constant of K+ inside Gramicidin A: a comparative study of four computational methods.

The local diffusion constant of K(+) inside the Gramicidin A (GA) channel has been calculated using four computational methods based on molecular dynamics (MD) simulations, specifically: Mean Square Displacement (MSD), Velocity Autocorrelation Function (VACF), Second Fluctuation Dissipation Theorem (SFDT) and analysis of the Generalized Langevin Equation for a Harmonic Oscillator (GLE-HO). All methods were first tested and compared for K(+) in bulk water--all predicted the correct diffusion constant. Inside GA, MSD and VACF methods were found to be unreliable because they are biased by the systematic force exerted by the membrane-channel system on the ion. SFDT and GLE-HO techniques properly unbias the influence of the systematic force on the diffusion properties and predicted a similar diffusion constant of K(+) inside GA, namely, ca. 10 times smaller than in the bulk. It was found that both SFDT and GLE-HO methods require extensive MD sampling on the order of tens of nanoseconds to predict a reliable diffusion constant of K(+) inside GA.

The role of the dielectric barrier in narrow biological channels: a novel composite approach to modeling single-channel currents.

A composite continuum theory for calculating ion current through a protein channel of known structure is proposed, which incorporates information about the channel dynamics. The approach is utilized to predict current through the Gramicidin A ion channel, a narrow pore in which the applicability of conventional continuum theories is questionable. The proposed approach utilizes a modified version of Poisson-Nernst-Planck (PNP) theory, termed Potential-of-Mean-Force-Poisson-Nernst-Planck theory (PMFPNP), to compute ion currents. As in standard PNP, ion permeation is modeled as a continuum drift-diffusion process in a self-consistent electrostatic potential. In PMFPNP, however, information about the dynamic relaxation of the protein and the surrounding medium is incorporated into the model of ion permeation by including the free energy of inserting a single ion into the channel, i.e., the potential of mean force along the permeation pathway. In this way the dynamic flexibility of the channel environment is approximately accounted for. The PMF profile of the ion along the Gramicidin A channel is obtained by combining an equilibrium molecular dynamics (MD) simulation that samples dynamic protein configurations when an ion resides at a particular location in the channel with a continuum electrostatics calculation of the free energy. The diffusion coefficient of a potassium ion within the channel is also calculated using the MD trajectory. Therefore, except for a reasonable choice of dielectric constants, no direct fitting parameters enter into this model. The results of our study reveal that the channel response to the permeating ion produces significant electrostatic stabilization of the ion inside the channel. The dielectric self-energy of the ion remains essentially unchanged in the course of the MD simulation, indicating that no substantial changes in the protein geometry occur as the ion passes through it. Also, the model accounts for the experimentally observed saturation of ion current with increase of the electrolyte concentration, in contrast to the predictions of standard PNP theory.