DeepFold: enhancing protein structure prediction through optimized loss functions, improved template features, and re-optimized energy function.

MOTIVATION: Predicting protein structures with high accuracy is a critical challenge for the broad community of life sciences and industry. Despite progress made by deep neural networks like AlphaFold2, there is a need for further improvements in the quality of detailed structures, such as side-chains, along with protein backbone structures. RESULTS: Building upon the successes of AlphaFold2, the modifications we made include changing the losses of side-chain torsion angles and frame aligned point error, adding loss functions for side chain confidence and secondary structure prediction, and replacing template feature generation with a new alignment method based on conditional random fields. We also performed re-optimization by conformational space annealing using a molecular mechanics energy function which integrates the potential energies obtained from distogram and side-chain prediction. In the CASP15 blind test for single protein and domain modeling (109 domains), DeepFold ranked fourth among 132 groups with improvements in the details of the structure in terms of backbone, side-chain, and Molprobity. In terms of protein backbone accuracy, DeepFold achieved a median GDT-TS score of 88.64 compared with 85.88 of AlphaFold2. For TBM-easy/hard targets, DeepFold ranked at the top based on Z-scores for GDT-TS. This shows its practical value to the structural biology community, which demands highly accurate structures. In addition, a thorough analysis of 55 domains from 39 targets with publicly available structures indicates that DeepFold shows superior side-chain accuracy and Molprobity scores among the top-performing groups. AVAILABILITY AND IMPLEMENTATION: DeepFold tools are open-source software available at https://github.com/newtonjoo/deepfold.

Application of conformational space annealing to the protein structure modeling using cryo-EM maps.

Conformational space annealing (CSA), a global optimization method, has been applied to various protein structure modeling tasks. In this paper, we applied CSA to the cryo-EM structure modeling task by combining the python subroutine of CSA (PyCSA) and the fast relax (FastRelax) protocol of PyRosetta. Refinement of initial structures generated from two methods, rigid fitting of predicted structures to the Cryo-EM map and de novo protein modeling by tracing the Cryo-EM map, was performed by CSA. In the refinement of the rigid-fitted structures, the final models showed that CSA can generate reliable atomic structures of proteins, even when large movements of protein domains were required. In the de novo modeling case, although the overall structural qualities of the final models were rather dependent on the initial models, the final models generated by CSA showed improved MolProbity scores and cross-correlation coefficients to the maps. These results suggest that CSA can accomplish flexible fitting and refinement together by sampling diverse conformations effectively and thus can be utilized for cryo-EM structure modeling.

Industrializing AI/ML during the end-to-end drug discovery process.

Drug discovery aims to select proper targets and drug candidates to address unmet clinical needs. The end-to-end drug discovery process includes all stages of drug discovery from target identification to drug candidate selection. Recently, several artificial intelligence and machine learning (AI/ML)-based drug discovery companies have attempted to build data-driven platforms spanning the end-to-end drug discovery process. The ability to identify elusive targets essentially leads to the diversification of discovery pipelines, thereby increasing the ability to address unmet needs. Modern ML technologies are complementing traditional computer-aided drug discovery by accelerating candidate optimization in innovative ways. This review summarizes recent developments in AI/ML methods from target identification to molecule optimization, and concludes with an overview of current industrial trends in end-to-end AI/ML platforms.

Exploring the Folding Mechanism of Small Proteins GB1 and LB1.

The computational atomistic description of the folding reactions of the B1 domains, GB1 and LB1, of protein G and protein L, respectively, is an important challenge in current protein folding studies. Although the two proteins have overall very similar backbone structures (ß-hairpin-α-helix-ß-hairpin), their apparent folding behaviors observed experimentally were remarkably different. LB1 folds in a two-state manner with the single-exponential kinetics, whereas GB1 folds in a more complex manner with an early stage intermediate that may exist on the folding pathway. Here, we used a new method of all-atom molecular dynamics simulations to investigate the folding mechanisms of GB1 and LB1. With the Lorentzian energy term derived from the native structure, we successfully observed frequent folding and unfolding events in the simulations at a high temperature (414 K for GB1 or 393 K for LB1) for both the proteins. Three and two transition-state structures were predicted for the GB1 and LB1 folding, respectively, at the high temperature. Two of the three transition-state structures of GB1 have a better formed second ß-hairpin. One of the LB1 transition states has a better formed first hairpin, and the other has both hairpins equally formed. The structural features of these transition states are in good agreement with experimental transition-state analysis. At 300 K, more complex folding processes were observed in the simulations for both the proteins. Several intermediate structures were predicted for the two proteins, which led to the conclusion that both the proteins folded through similar mechanisms. However, the intermediate state accumulated in a sufficient amount only in the GB1 folding, which led to the double-exponential feature of its folding kinetics. On the other hand, the LB1 folding kinetics were well fitted by a single-exponential function. These results are fully consistent with those previously observed experimentally.

A general method for the derivation of the functional forms of the effective energy terms in coarse-grained energy functions of polymers. III. Determination of scale-consistent backbone-local and correlation potentials in the UNRES force field and force-field calibration and validation.

The general theory of the construction of scale-consistent energy terms in the coarse-grained force fields presented in Paper I of this series has been applied to the revision of the UNRES force field for physics-based simulations of proteins. The potentials of mean force corresponding to backbone-local and backbone-correlation energy terms were calculated from the ab initio energy surfaces of terminally blocked glycine, alanine, and proline, and the respective analytical expressions, derived by using the scale-consistent formalism, were fitted to them. The parameters of all these potentials depend on single-residue types, thus reducing their number and preventing over-fitting. The UNRES force field with the revised backbone-local and backbone-correlation terms was calibrated with a set of four small proteins with basic folds: tryptophan cage variant (TRP1; α), Full Sequence Design (FSD; α + ß), villin headpiece (villin; α), and a truncated FBP-28 WW-domain variant (2MWD; ß) (the NEWCT-4P force field) and, subsequently, with an enhanced set of 9 proteins composed of TRP1, FSD, villin, 1BDC (α), 2I18 (α), 1QHK (α + ß), 2N9L (α + ß), 1E0L (ß), and 2LX7 (ß) (the NEWCT-9P force field). The NEWCT-9P force field performed better than NEWCT-4P in a blind-prediction-like test with a set of 26 proteins not used in calibration and outperformed, in a test with 76 proteins, the most advanced OPT-WTFSA-2 version of UNRES with former backbone-local and backbone-correlation terms that contained more energy terms and more optimizable parameters. The NEWCT-9P force field reproduced the bimodal distribution of backbone-virtual-bond angles in the simulated structures, as observed in experimental protein structures.

Non-sequential protein structure alignment by conformational space annealing and local refinement.

Protein structure alignment is an important tool for studying evolutionary biology and protein modeling. A tool which intensively searches for the globally optimal non-sequential alignments is rarely found. We propose ALIGN-CSA which shows improvement in scores, such as DALI-score, SP-score, SO-score and TM-score over the benchmark set including 286 cases. We performed benchmarking of existing popular alignment scoring functions, where the dependence of the search algorithm was effectively eliminated by using ALIGN-CSA. For the benchmarking, we set the minimum block size to 4 to prevent much fragmented alignments where the biological relevance of small alignment blocks is hard to interpret. With this condition, globally optimal alignments were searched by ALIGN-CSA using the four scoring functions listed above, and TM-score is found to be the most effective in generating alignments with longer match lengths and smaller RMSD values. However, DALI-score is the most effective in generating alignments similar to the manually curated reference alignments, which implies that DALI-score is more biologically relevant score. Due to the high demand on computational resources of ALIGN-CSA, we also propose a relatively fast local refinement method, which can control the minimum block size and whether to allow the reverse alignment. ALIGN-CSA can be used to obtain much improved alignment at the cost of relatively more extensive computation. For faster alignment, we propose a refinement protocol that improves the score of a given alignment obtained by various external tools. All programs are available from http://lee.kias.re.kr.

CHARMM-GUI Glycan Modeler for modeling and simulation of carbohydrates and glycoconjugates.

Characterizing glycans and glycoconjugates in the context of three-dimensional structures is important in understanding their biological roles and developing efficient therapeutic agents. Computational modeling and molecular simulation have become an essential tool complementary to experimental methods. Here, we present a computational tool, Glycan Modeler for in silico N-/O-glycosylation of the target protein and generation of carbohydrate-only systems. In our previous study, we developed Glycan Reader, a web-based tool for detecting carbohydrate molecules from a PDB structure and generation of simulation system and input files. As integrated into Glycan Reader in CHARMM-GUI, Glycan Modeler (Glycan Reader & Modeler) enables to generate the structures of glycans and glycoconjugates for given glycan sequences and glycosylation sites using PDB glycan template structures from Glycan Fragment Database (http://glycanstructure.org/fragment-db). Our benchmark tests demonstrate the universal applicability of Glycan Reader & Modeler to various glycan sequences and target proteins. We also investigated the structural properties of modeled glycan structures by running 2-µs molecular dynamics simulations of HIV envelope protein. The simulations show that the modeled glycan structures built by Glycan Reader & Modeler have the similar structural features compared to the ones solved by X-ray crystallography. We also describe the representative examples of glycoconjugate modeling with video demos to illustrate the practical applications of Glycan Reader & Modeler. Glycan Reader & Modeler is freely available at http://charmm-gui.org/input/glycan.

Data-assisted protein structure modeling by global optimization in CASP12.

In CASP12, 2 types of data-assisted protein structure modeling were experimented. Either SAXS experimental data or cross-linking experimental data was provided for a selected number of CASP12 targets that the CASP12 predictor could utilize for better protein structure modeling. We devised 2 separate energy terms for SAXS data and cross-linking data to drive the model structures into more native-like structures that satisfied the given experimental data as much as possible. In CASP11, we successfully performed protein structure modeling using simulated sparse and ambiguously assigned NOE data and/or correct residue-residue contact information, where the only energy term that folded the protein into its native structure was the term which was originated from the given experimental data. However, the 2 types of experimental data provided in CASP12 were far from being sufficient enough to fold the target protein into its native structure because SAXS data provides only the overall shape of the molecule and the cross-linking contact information provides only very low-resolution distance information. For this reason, we combined the SAXS or cross-linking energy term with our regular modeling energy function that includes both the template energy term and the de novo energy terms. By optimizing the newly formulated energy function, we obtained protein models that fit better with provided SAXS data than the X-ray structure of the target. However, the improvement of the model relative to the 1 modeled without the SAXS data, was not significant. Consistent structural improvement was achieved by incorporating cross-linking data into the protein structure modeling.

Protein structure modeling and refinement by global optimization in CASP12.

For protein structure modeling in the CASP12 experiment, we have developed a new protocol based on our previous CASP11 approach. The global optimization method of conformational space annealing (CSA) was applied to 3 stages of modeling: multiple sequence-structure alignment, three-dimensional (3D) chain building, and side-chain re-modeling. For better template selection and model selection, we updated our model quality assessment (QA) method with the newly developed SVMQA (support vector machine for quality assessment). For 3D chain building, we updated our energy function by including restraints generated from predicted residue-residue contacts. New energy terms for the predicted secondary structure and predicted solvent accessible surface area were also introduced. For difficult targets, we proposed a new method, LEEab, where the template term played a less significant role than it did in LEE, complemented by increased contributions from other terms such as the predicted contact term. For TBM (template-based modeling) targets, LEE performed better than LEEab, but for FM targets, LEEab was better. For model refinement, we modified our CASP11 molecular dynamics (MD) based protocol by using explicit solvents and tuning down restraint weights. Refinement results from MD simulations that used a new augmented statistical energy term in the force field were quite promising. Finally, when using inaccurate information (such as the predicted contacts), it was important to use the Lorentzian function for which the maximal penalty arising from wrong information is always bounded.

A Simple and Efficient Protein Structure Refinement Method.

Improving the quality of a given protein structure can serve as the ultimate solution for accurate protein structure prediction, and seeking such a method is currently a challenge in computational structural biology. In order to promote and encourage much needed such efforts, CASP (Critical Assessment of Structure Prediction) has been providing an ideal computational experimental platform, where it was reported only recently (since CASP10) that systematic protein structure refinement is possible by carrying out extensive (approximately millisecond) MD simulations with proper restraints generated from the given structure. Using an explicit solvent model and much reduced positional and distance restraints than previously exercised, we propose a refinement protocol that combines a series of short (5 ns) MD simulations with energy minimization procedures. Testing and benchmarking on 54 CASP8-10 refinement targets and 34 CASP11 refinement targets shows quite promising results. Using only a small fraction of MD simulation steps (nanosecond versus millisecond), systematic protein structure refinement was demonstrated in this work, indicating that refinement of a given model can be achieved using a few hours of desktop computing.

Finding multiple reaction pathways via global optimization of action.

Global searching for reaction pathways is a long-standing challenge in computational chemistry and biology. Most existing approaches perform only local searches due to computational complexity. Here we present a computational approach, Action-CSA, to find multiple diverse reaction pathways connecting fixed initial and final states through global optimization of the Onsager-Machlup action using the conformational space annealing (CSA) method. Action-CSA successfully overcomes large energy barriers via crossovers and mutations of pathways and finds all possible pathways of small systems without initial guesses on pathways. The rank order and the transition time distribution of multiple pathways are in good agreement with those of long Langevin dynamics simulations. The lowest action folding pathway of FSD-1 is consistent with recent experiments. The results show that Action-CSA is an efficient and robust computational approach to study the multiple pathways of complex reactions and large-scale conformational changes.

Contact-assisted protein structure modeling by global optimization in CASP11.

We have applied the conformational space annealing method to the contact-assisted protein structure modeling in CASP11. For Tp targets, where predicted residue-residue contact information was provided, the contact energy term in the form of the Lorentzian function was implemented together with the physical energy terms used in our template-free modeling of proteins. Although we observed some structural improvement of Tp models over the models predicted without the Tp information, the improvement was not substantial on average. This is partly due to the inaccuracy of the provided contact information, where only about 18% of it was correct. For Ts targets, where the information of ambiguous NOE (Nuclear Overhauser Effect) restraints was provided, we formulated the modeling in terms of the two-tier optimization problem, which covers: (1) the assignment of NOE peaks and (2) the three-dimensional (3D) model generation based on the assigned NOEs. Although solving the problem in a direct manner appears to be intractable at first glance, we demonstrate through CASP11 that remarkably accurate protein 3D modeling is possible by brute force optimization of a relevant energy function. For 19 Ts targets of the average size of 224 residues, generated protein models were of about 3.6 Å Cα atom accuracy. Even greater structural improvement was observed when additional Tc contact information was provided. For 20 out of the total 24 Tc targets, we were able to generate protein structures which were better than the best model from the rest of the CASP11 groups in terms of GDT-TS. Proteins 2016; 84(Suppl 1):189-199. © 2015 Wiley Periodicals, Inc.

Template based protein structure modeling by global optimization in CASP11.

For the template-based modeling (TBM) of CASP11 targets, we have developed three new protein modeling protocols (nns for server prediction and LEE and LEER for human prediction) by improving upon our previous CASP protocols (CASP7 through CASP10). We applied the powerful global optimization method of conformational space annealing to three stages of optimization, including multiple sequence-structure alignment, three-dimensional (3D) chain building, and side-chain remodeling. For more successful fold recognition, a new alignment method called CRFalign was developed. It can incorporate sensitive positional and environmental dependence in alignment scores as well as strong nonlinear correlations among various features. Modifications and adjustments were made to the form of the energy function and weight parameters pertaining to the chain building procedure. For the side-chain remodeling step, residue-type dependence was introduced to the cutoff value that determines the entry of a rotamer to the side-chain modeling library. The improved performance of the nns server method is attributed to successful fold recognition achieved by combining several methods including CRFalign and to the current modeling formulation that can incorporate native-like structural aspects present in multiple templates. The LEE protocol is identical to the nns one except that CASP11-released server models are used as templates. The success of LEE in utilizing CASP11 server models indicates that proper template screening and template clustering assisted by appropriate cluster ranking promises a new direction to enhance protein 3D modeling. Proteins 2016; 84(Suppl 1):221-232. © 2015 Wiley Periodicals, Inc.

Template-free modeling by LEE and LEER in CASP11.

For the template-free modeling of human targets of CASP11, we utilized two of our modeling protocols, LEE and LEER. The LEE protocol took CASP11-released server models as the input and used some of them as templates for 3D (three-dimensional) modeling. The template selection procedure was based on the clustering of the server models aided by a community detection method of a server-model network. Restraining energy terms generated from the selected templates together with physical and statistical energy terms were used to build 3D models. Side-chains of the 3D models were rebuilt using target-specific consensus side-chain library along with the SCWRL4 rotamer library, which completed the LEE protocol. The first success factor of the LEE protocol was due to efficient server model screening. The average backbone accuracy of selected server models was similar to that of top 30% server models. The second factor was that a proper energy function along with our optimization method guided us, so that we successfully generated better quality models than the input template models. In 10 out of 24 cases, better backbone structures than the best of input template structures were generated. LEE models were further refined by performing restrained molecular dynamics simulations to generate LEER models. CASP11 results indicate that LEE models were better than the average template models in terms of both backbone structures and side-chain orientations. LEER models were of improved physical realism and stereo-chemistry compared to LEE models, and they were comparable to LEE models in the backbone accuracy. Proteins 2016; 84(Suppl 1):118-130. © 2015 Wiley Periodicals, Inc.

Protein structure determination by conformational space annealing using NMR geometric restraints.

We have carried out numerical experiments to investigate the applicability of the global optimization method of conformational space annealing (CSA) to the enhanced NMR protein structure determination over existing PDB structures. The NMR protein structure determination is driven by the optimization of collective multiple restraints arising from experimental data and the basic stereochemical properties of a protein-like molecule. By rigorous and straightforward application of CSA to the identical NMR experimental data used to generate existing PDB structures, we redetermined 56 recent PDB protein structures starting from fully randomized structures. The quality of CSA-generated structures and existing PDB structures were assessed by multiobjective functions in terms of their consistencies with experimental data and the requirements of protein-like stereochemistry. In 54 out of 56 cases, CSA-generated structures were better than existing PDB structures in the Pareto-dominant manner, while in the remaining two cases, it was a tie with mixed results. As a whole, all structural features tested improved in a statistically meaningful manner. The most improved feature was the Ramachandran favored portion of backbone torsion angles with about 8.6% improvement from 88.9% to 97.5% (P-value <10(-17)). We show that by straightforward application of CSA to the efficient global optimization of an energy function, NMR structures will be of better quality than existing PDB structures.

Sigma-RF: prediction of the variability of spatial restraints in template-based modeling by random forest.

BACKGROUND: In template-based modeling when using a single template, inter-atomic distances of an unknown protein structure are assumed to be distributed by Gaussian probability density functions, whose center peaks are located at the distances between corresponding atoms in the template structure. The width of the Gaussian distribution, the variability of a spatial restraint, is closely related to the reliability of the restraint information extracted from a template, and it should be accurately estimated for successful template-based protein structure modeling. RESULTS: To predict the variability of the spatial restraints in template-based modeling, we have devised a prediction model, Sigma-RF, by using the random forest (RF) algorithm. The benchmark results on 22 CASP9 targets show that the variability values from Sigma-RF are of higher correlations with the true distance deviation than those from Modeller. We assessed the effect of new sigma values by performing the single-domain homology modeling of 22 CASP9 targets and 24 CASP10 targets. For most of the targets tested, we could obtain more accurate 3D models from the identical alignments by using the Sigma-RF results than by using Modeller ones. CONCLUSIONS: We find that the average alignment quality of residues located between and at two aligned residues, quasi-local information, is the most contributing factor, by investigating the importance of input features used in the RF machine learning. This average alignment quality is shown to be more important than the previously identified quantity of a local information: the product of alignment qualities at two aligned residues.

Soft-core potentials in thermodynamic integration: comparing one- and two-step transformations.

Molecular dynamics-based free energy calculations allow the determination of a variety of thermodynamic quantities from computer simulations of small molecules. Thermodynamic integration (TI) calculations can suffer from instabilities during the creation or annihilation of particles. This "singularity" problem can be addressed with "soft-core" potential functions which keep pairwise interaction energies finite for all configurations and provide smooth free energy curves. "One-step" transformations, in which electrostatic and van der Waals forces are simultaneously modified, can be simpler and less expensive than "two-step" transformations in which these properties are changed in separate calculations. Here, we study solvation free energies for molecules of different hydrophobicity using both models. We provide recommended values for the two parameters α(LJ) and ß(C) controlling the behavior of the soft-core Lennard-Jones and Coulomb potentials and compare one- and two-step transformations with regard to their suitability for numerical integration. For many types of transformations, the one-step procedure offers a convenient and accurate approach to free energy estimates.