A Review of Deep Learning Methods for Antibodies.

Driven by its successes across domains such as computer vision and natural language processing, deep learning has recently entered the field of biology by aiding in cellular image classification, finding genomic connections, and advancing drug discovery. In drug discovery and protein engineering, a major goal is to design a molecule that will perform a useful function as a therapeutic drug. Typically, the focus has been on small molecules, but new approaches have been developed to apply these same principles of deep learning to biologics, such as antibodies. Here we give a brief background of deep learning as it applies to antibody drug development, and an in-depth explanation of several deep learning algorithms that have been proposed to solve aspects of both protein design in general, and antibody design in particular.

Automated Aufbau of antibody structures from given sequences using Macromoltek's SmrtMolAntibody.

This study was a part of the second antibody modeling assessment. The assessment is a blind study of the performance of multiple software programs used for antibody homology modeling. In the study, research groups were given sequences for 11 antibodies and asked to predict their corresponding structures. The results were measured using root-mean-square deviation (rmsd) between the submitted models and X-ray crystal structures. In 10 of 11 cases, the results using SmrtMolAntibody show good agreement between the submitted models and X-ray crystal structures. In the first stage, the average rmsd was 1.4 Å. Average rmsd values for the framework was 1.2 Å and for the H3 loop was 3.0 Å. In stage two, there was a slight improvement with an rmsd for the H3 loop of 2.9 Å.

Structure-based design of supercharged, highly thermoresistant antibodies.

Mutation of surface residues to charged amino acids increases resistance to aggregation and can enable reversible unfolding. We have developed a protocol using the Rosetta computational design package that "supercharges" proteins while considering the energetic implications of each mutation. Using a homology model, a single-chain variable fragment antibody was designed that has a markedly enhanced resistance to thermal inactivation and displays an unanticipated ≈30-fold improvement in affinity. Such supercharged antibodies should prove useful for assays in resource-limited settings and for developing reagents with improved shelf lives.

Computed structures of point deletion mutants and their enzymatic activities.

Point deletions in enzymes can vary in effect from negligible to complete loss of activity; however, these effects are not generally predictable. Deletions are widely observed in nature and often result in diseases such as cancer, cystic fibrosis, or osteogenesis imperfecta. Here, we have developed an algorithm to model the perturbed structures of deletion mutants with the ultimate goal of predicting their activities. The algorithm works by deleting the specified residue from the wild-type structure, creating a gap that is closed using a combination of local and global moves that change the backbone torsion angles of the protein structure. On a set of five proteins for which both wild-type and deletion mutant x-ray crystal structures are available, the algorithm produces deep, narrow energy funnels within 1.5 Å of the crystal structure for the deletion mutants. To assess the ability of our algorithm to predict activity from the predicted structures, we tested the correlation of experimental activity with several measures of the predicted structure ensemble using a set of 45 point deletions from ricin. Estimates incorporating likely prevalence of active and inactive deletion sites suggest that activity can be predicted correctly over 60% of the time from the active site root-mean squared deviation of the lowest energy predicted structures. The predictions are stronger than simple sequence organization measures, but more fundamental work is required in structure prediction and enzyme activity determination to allow consistent prediction of activity.

Benchmarking and analysis of protein docking performance in Rosetta v3.2.

RosettaDock has been increasingly used in protein docking and design strategies in order to predict the structure of protein-protein interfaces. Here we test capabilities of RosettaDock 3.2, part of the newly developed Rosetta v3.2 modeling suite, against Docking Benchmark 3.0, and compare it with RosettaDock v2.3, the latest version of the previous Rosetta software package. The benchmark contains a diverse set of 116 docking targets including 22 antibody-antigen complexes, 33 enzyme-inhibitor complexes, and 60 'other' complexes. These targets were further classified by expected docking difficulty into 84 rigid-body targets, 17 medium targets, and 14 difficult targets. We carried out local docking perturbations for each target, using the unbound structures when available, in both RosettaDock v2.3 and v3.2. Overall the performances of RosettaDock v2.3 and v3.2 were similar. RosettaDock v3.2 achieved 56 docking funnels, compared to 49 in v2.3. A breakdown of docking performance by protein complex type shows that RosettaDock v3.2 achieved docking funnels for 63% of antibody-antigen targets, 62% of enzyme-inhibitor targets, and 35% of 'other' targets. In terms of docking difficulty, RosettaDock v3.2 achieved funnels for 58% of rigid-body targets, 30% of medium targets, and 14% of difficult targets. For targets that failed, we carry out additional analyses to identify the cause of failure, which showed that binding-induced backbone conformation changes account for a majority of failures. We also present a bootstrap statistical analysis that quantifies the reliability of the stochastic docking results. Finally, we demonstrate the additional functionality available in RosettaDock v3.2 by incorporating small-molecules and non-protein co-factors in docking of a smaller target set. This study marks the most extensive benchmarking of the RosettaDock module to date and establishes a baseline for future research in protein interface modeling and structure prediction.

ROSETTA3: an object-oriented software suite for the simulation and design of macromolecules.

We have recently completed a full re-architecturing of the ROSETTA molecular modeling program, generalizing and expanding its existing functionality. The new architecture enables the rapid prototyping of novel protocols by providing easy-to-use interfaces to powerful tools for molecular modeling. The source code of this rearchitecturing has been released as ROSETTA3 and is freely available for academic use. At the time of its release, it contained 470,000 lines of code. Counting currently unpublished protocols at the time of this writing, the source includes 1,285,000 lines. Its rapid growth is a testament to its ease of use. This chapter describes the requirements for our new architecture, justifies the design decisions, sketches out central classes, and highlights a few of the common tasks that the new software can perform.

A generalized approach to sampling backbone conformations with RosettaDock for CAPRI rounds 13-19.

In CAPRI rounds 13-19, the most native-like structure predicted by RosettaDock resulted in two high, one medium, and one acceptable accuracy model out of 13 targets. The current rounds of CAPRI were especially challenging with many unbound and homology modeled starting structures. Novel docking methods, including EnsembleDock and SnugDock, allowed backbone conformational sampling during docking and enabled the creation of more accurate models. For Target 32, α-amylase/subtilisin inhibitor-subtilisin savinase, we sampled different backbone conformations at an interfacial loop to produce five high-quality models including the most accurate structure submitted in the challenge (2.1 Å ligand rmsd, 0.52 Å interface rmsd). For Target 41, colicin-immunity protein, we used EnsembleDock to sample the ensemble of nuclear magnetic resonance (NMR) models of the immunity protein to generate a medium accuracy structure. Experimental data identifying the catalytic residues at the binding interface for Target 40 (trypsin-inhibitor) were used to filter RosettaDock global rigid body docking decoys to determine high accuracy predictions for the two distinct binding sites in which the inhibitor interacts with trypsin. We discuss our generalized approach to selecting appropriate methods for different types of docking problems. The current toolset provides some robustness to errors in homology models, but significant challenges remain in accommodating larger backbone uncertainties and in sampling adequately for global searches.

Computational predictions of the mutant behavior of AraC.

An algorithm implemented in Rosetta correctly predicts the folding capabilities of the 17-residue N-terminal arm of the AraC gene regulatory protein when arabinose is bound to the protein and the dramatically different structure of this arm when arabinose is absent. The transcriptional activity of 43 mutant AraC proteins with alterations in the arm sequences was measured in vivo and compared with their predicted folding properties. Seventeen of the mutants possessed regulatory properties that could be directly compared with folding predictions. Sixteen of the 17 mutants were correctly predicted. The algorithm predicts that the N-terminal arm sequences of AraC homologs fold to the Escherichia coli AraC arm structure. In contrast, it predicts that random sequences of the same length and many partially randomized E. coli arm sequences do not fold to the E. coli arm structure. The high level of success shows that relatively "simple" computational methods can in some cases predict the behavior of mutant proteins with good reliability.

Structure prediction of domain insertion proteins from structures of individual domains.

Multidomain proteins continue to be a major challenge in protein structure prediction. Here we present a Monte Carlo (MC) algorithm, implemented within Rosetta, to predict the structure of proteins in which one domain is inserted into another. Three MC moves combine rigid-body and loop movements to search the constrained conformation by structure disruption and subsequent repair of chain breaks. Local searches find that the algorithm samples and recovers near-native structures consistently. Further global searches produced top-ranked structures within 5 A in 31 of 50 cases in low-resolution mode, and refinement of top-ranked low-resolution structures produced models within 2 A in 21 of 50 cases. Rigid-body orientations were often correctly recovered despite errors in linker conformation. The algorithm is broadly applicable to de novo structure prediction of both naturally occurring and engineered domain insertion proteins.

Incorporating biochemical information and backbone flexibility in RosettaDock for CAPRI rounds 6-12.

In CAPRI rounds 6-12, RosettaDock successfully predicted 2 of 5 unbound-unbound targets to medium accuracy. Improvement over the previous method was achieved with computational mutagenesis to select decoys that match the energetics of experimentally determined hot spots. In the case of Target 21, Orc1/Sir1, this resulted in a successful docking prediction where RosettaDock alone or with simple site constraints failed. Experimental information also helped limit the interacting region of TolB/Pal, producing a successful prediction of Target 26. In addition, we docked multiple loop conformations for Target 20, and we developed a novel flexible docking algorithm to simultaneously optimize backbone conformation and rigid-body orientation to generate a wide diversity of conformations for Target 24. Continued challenges included docking of homology targets that differ substantially from their template (sequence identity <50%) and accounting for large conformational changes upon binding. Despite a larger number of unbound-unbound and homology model binding targets, Rounds 6-12 reinforced that RosettaDock is a powerful algorithm for predicting bound complex structures, especially when combined with experimental data.