OpenFold: retraining AlphaFold2 yields new insights into its learning mechanisms and capacity for generalization.

AlphaFold2 revolutionized structural biology with the ability to predict protein structures with exceptionally high accuracy. Its implementation, however, lacks the code and data required to train new models. These are necessary to (1) tackle new tasks, like protein-ligand complex structure prediction, (2) investigate the process by which the model learns and (3) assess the model's capacity to generalize to unseen regions of fold space. Here we report OpenFold, a fast, memory efficient and trainable implementation of AlphaFold2. We train OpenFold from scratch, matching the accuracy of AlphaFold2. Having established parity, we find that OpenFold is remarkably robust at generalizing even when the size and diversity of its training set is deliberately limited, including near-complete elisions of classes of secondary structure elements. By analyzing intermediate structures produced during training, we also gain insights into the hierarchical manner in which OpenFold learns to fold. In sum, our studies demonstrate the power and utility of OpenFold, which we believe will prove to be a crucial resource for the protein modeling community.

Recurrent signature patterns in HIV-1 B clade envelope glycoproteins associated with either early or chronic infections.

Here we have identified HIV-1 B clade Envelope (Env) amino acid signatures from early in infection that may be favored at transmission, as well as patterns of recurrent mutation in chronic infection that may reflect common pathways of immune evasion. To accomplish this, we compared thousands of sequences derived by single genome amplification from several hundred individuals that were sampled either early in infection or were chronically infected. Samples were divided at the outset into hypothesis-forming and validation sets, and we used phylogenetically corrected statistical strategies to identify signatures, systematically scanning all of Env. Signatures included single amino acids, glycosylation motifs, and multi-site patterns based on functional or structural groupings of amino acids. We identified signatures near the CCR5 co-receptor-binding region, near the CD4 binding site, and in the signal peptide and cytoplasmic domain, which may influence Env expression and processing. Two signatures patterns associated with transmission were particularly interesting. The first was the most statistically robust signature, located in position 12 in the signal peptide. The second was the loss of an N-linked glycosylation site at positions 413-415; the presence of this site has been recently found to be associated with escape from potent and broad neutralizing antibodies, consistent with enabling a common pathway for immune escape during chronic infection. Its recurrent loss in early infection suggests it may impact fitness at the time of transmission or during early viral expansion. The signature patterns we identified implicate Env expression levels in selection at viral transmission or in early expansion, and suggest that immune evasion patterns that recur in many individuals during chronic infection when antibodies are present can be selected against when the infection is being established prior to the adaptive immune response.

Structure of HIV-1 gp120 with gp41-interactive region reveals layered envelope architecture and basis of conformational mobility.

The viral spike of HIV-1 is composed of three gp120 envelope glycoproteins attached noncovalently to three gp41 transmembrane molecules. Viral entry is initiated by binding to the CD4 receptor on the cell surface, which induces large conformational changes in gp120. These changes not only provide a model for receptor-triggered entry, but affect spike sensitivity to drug- and antibody-mediated neutralization. Although some of the details of the CD4-induced conformational change have been visualized by crystal structures and cryoelectron tomograms, the critical gp41-interactive region of gp120 was missing from previous atomic-level characterizations. Here we determine the crystal structure of an HIV-1 gp120 core with intact gp41-interactive region in its CD4-bound state, compare this structure to unliganded and antibody-bound forms to identify structurally invariant and plastic components, and use ligand-oriented cryoelectron tomograms to define component mobility in the viral spike context. Newly defined gp120 elements proximal to the gp41 interface complete a 7-stranded beta-sandwich, which appeared invariant in conformation. Loop excursions emanating from the sandwich form three topologically separate--and structurally plastic--layers, topped off by the highly glycosylated gp120 outer domain. Crystal structures, cryoelectron tomograms, and interlayer chemistry were consistent with a mechanism in which the layers act as a shape-changing spacer, facilitating movement between outer domain and gp41-associated beta-sandwich and providing for conformational diversity used in immune evasion. A "layered" gp120 architecture thus allows movement among alternative glycoprotein conformations required for virus entry and immune evasion, whereas a beta-sandwich clamp maintains gp120-gp41 interaction and regulates gp41 transitions.

Modulation of integrin activation by an entropic spring in the {beta}-knee.

We show that the length of a loop in the ß-knee, between the first and second cysteines (C1-C2) in integrin EGF-like (I-EGF) domain 2, modulates integrin activation. Three independent sets of mutants, including swaps among different integrin ß-subunits, show that C1-C2 loop lengths of 12 and longer favor the low affinity state and masking of ligand-induced binding site (LIBS) epitopes. Shortening length from 12 to 4 residues progressively increases ligand binding and LIBS epitope exposure. Compared with length, the loop sequence had a smaller effect, which was ascribable to stabilizing loop conformation, and not interactions with the α-subunit. The data together with structural calculations support the concept that the C1-C2 loop is an entropic spring and an emerging theme that disordered regions can regulate allostery. Diversity in the length of this loop may have evolved among integrin ß-subunits to adjust the equilibrium between the bent and extended conformations at different set points.

Role of complex carbohydrates in human immunodeficiency virus type 1 infection and resistance to antibody neutralization.

Complex N-glycans flank the receptor binding sites of the outer domain of HIV-1 gp120, ostensibly forming a protective "fence" against antibodies. Here, we investigated the effects of rebuilding this fence with smaller glycoforms by expressing HIV-1 pseudovirions from a primary isolate in a human cell line lacking N-acetylglucosamine transferase I (GnTI), the enzyme that initiates the conversion of oligomannose N-glycans into complex N-glycans. Thus, complex glycans, including those that surround the receptor binding sites, are replaced by fully trimmed oligomannose stumps. Conversely, the untrimmed oligomannoses of the silent domain of gp120 are likely to remain unchanged. For comparison, we produced a mutant virus lacking a complex N-glycan of the V3 loop (N301Q). Both variants exhibited increased sensitivities to V3 loop-specific monoclonal antibodies (MAbs) and soluble CD4. The N301Q virus was also sensitive to "nonneutralizing" MAbs targeting the primary and secondary receptor binding sites. Endoglycosidase H treatment resulted in the removal of outer domain glycans from the GnTI- but not the parent Env trimers, and this was associated with a rapid and complete loss in infectivity. Nevertheless, the glycan-depleted trimers could still bind to soluble receptor and coreceptor analogs, suggesting a block in post-receptor binding conformational changes necessary for fusion. Collectively, our data show that the antennae of complex N-glycans serve to protect the V3 loop and CD4 binding site, while N-glycan stems regulate native trimer conformation, such that their removal can lead to global changes in neutralization sensitivity and, in extreme cases, an inability to complete the conformational rearrangements necessary for infection.

Evolution of graded refractive index in squid lenses.

A lens with a graded refractive index is required for vision in aquatic animals with camera-type eyes. This optical design entails a radial gradient of protein density, with low density in external layers and high density in internal layers. To maintain the optical stability of the eye, different material properties are required for proteins in different regions of the lens. In low-density regions of the lens where slight protein aggregation causes significant light scattering, aggregation must be minimized. Squid lens S-crystallin proteins are evolutionarily derived from the glutathione S-transferase protein family. We used biochemistry, optical modelling and phylogenetics to study the evolution and material properties of S-crystallins. S-crystallins are differentially expressed in a radial gradient, suggesting a role in refractive index. This gradient in S-crystallin expression is correlated with their evolutionary history and biochemistry. S-crystallins have been under positive selection. This selection appears to have resulted in stabilization of derived S-crystallins via mutations in the dimer interface and extended electrostatic fields. These derived S-crystallins probably cause the glassy organization and stability of low refractive index lens layers. Our work elucidates the molecular and evolutionary mechanisms underlying the production and maintenance of camera-like optics in squid lenses.

Evaluating the quality of NMR structures by local density of protons.

Evaluating the quality of experimentally determined protein structural models is an essential step toward identifying potential errors and guiding further structural refinement. Herein, we report the use of proton local density as a sensitive measure to assess the quality of nuclear magnetic resonance (NMR) structures. Using 256 high-resolution crystal structures with protons added and optimized, we show that the local density of different proton types display distinct distributions. These distributions can be characterized by statistical moments and are used to establish local density Z-scores for evaluating both global and local packing for individual protons. Analysis of 546 crystal structures at various resolutions shows that the local density Z-scores increase as the structural resolution decreases and correlate well with the ClashScore (Word et al. J Mol Biol 1999;285(4):1711-1733) generated by all atom contact analysis. Local density Z-scores for NMR structures exhibit a significantly wider range of values than for X-ray structures and demonstrate a combination of potentially problematic inflation and compression. Water-refined NMR structures show improved packing quality. Our analysis of a high-quality structural ensemble of ubiquitin refined against order parameters shows proton density distributions that correlate nearly perfectly with our standards derived from crystal structures, further validating our approach. We present an automated analysis and visualization tool for proton packing to evaluate the quality of NMR structures.

Computational design of high-affinity epitope scaffolds by backbone grafting of a linear epitope.

Computational grafting of functional motifs onto scaffold proteins is a promising way to engineer novel proteins with pre-specified functionalities. Typically, protein grafting involves the transplantation of protein side chains from a functional motif onto structurally homologous regions of scaffold proteins. Using this approach, we previously transplanted the human immunodeficiency virus 2F5 and 4E10 epitopes onto heterologous proteins to design novel "epitope-scaffold" antigens. However, side-chain grafting is limited by the availability of scaffolds with compatible backbone for a given epitope structure and offers no route to modify backbone structure to improve mimicry or binding affinity. To address this, we report here a new and more aggressive computational method-backbone grafting of linear motifs-that transplants the backbone and side chains of linear functional motifs onto scaffold proteins. To test this method, we first used side-chain grafting to design new 2F5 epitope scaffolds with improved biophysical characteristics. We then independently transplanted the 2F5 epitope onto three of the same parent scaffolds using the newly developed backbone grafting procedure. Crystal structures of side-chain and backbone grafting designs showed close agreement with both the computational models and the desired epitope structure. In two cases, backbone grafting scaffolds bound antibody 2F5 with 30- and 9-fold higher affinity than corresponding side-chain grafting designs. These results demonstrate that flexible backbone methods for epitope grafting can significantly improve binding affinities over those achieved by fixed backbone methods alone. Backbone grafting of linear motifs is a general method to transplant functional motifs when backbone remodeling of the target scaffold is necessary.

RosettaRemodel: a generalized framework for flexible backbone protein design.

We describe RosettaRemodel, a generalized framework for flexible protein design that provides a versatile and convenient interface to the Rosetta modeling suite. RosettaRemodel employs a unified interface, called a blueprint, which allows detailed control over many aspects of flexible backbone protein design calculations. RosettaRemodel allows the construction and elaboration of customized protocols for a wide range of design problems ranging from loop insertion and deletion, disulfide engineering, domain assembly, loop remodeling, motif grafting, symmetrical units, to de novo structure modeling.

Computational protein design using flexible backbone remodeling and resurfacing: case studies in structure-based antigen design.

Computational protein design has promise for vaccine design and other applications. We previously transplanted the HIV 4E10 epitope onto non-HIV protein scaffolds for structural stabilization and immune presentation. Here, we developed two methods to optimize the structure of an antigen, flexible backbone remodeling and resurfacing, and we applied these methods to a 4E10 scaffold. In flexible-backbone remodeling, an existing backbone segment is replaced by a de novo designed segment of prespecified length and secondary structure. With remodeling, we replaced a potentially immunodominant domain on the scaffold with a helix-loop segment that made intimate contact to the protein core. All three domain trim designs tested experimentally had improved thermal stability and similar binding affinity for the 4E10 antibody compared to the parent scaffold. A crystal structure of one design had a 0.8 Å backbone RMSD to the computational model in the rebuilt region. Comparison of parent and trimmed scaffold reactivity to anti-parent sera confirmed the deletion of an immunodominant domain. In resurfacing, the surface of an antigen outside a target epitope is redesigned to obtain variants that maintain only the target epitope. Resurfaced variants of two scaffolds were designed in which 50 positions amounting to 40% of the protein sequences were mutated. Surface-patch analyses indicated that most potential antibody footprints outside the 4E10 epitope were altered. The resurfaced variants maintained thermal stability and binding affinity. These results indicate that flexible-backbone remodeling and resurfacing are useful tools for antigen optimization and protein engineering generally.

Computation-guided backbone grafting of a discontinuous motif onto a protein scaffold.

The manipulation of protein backbone structure to control interaction and function is a challenge for protein engineering. We integrated computational design with experimental selection for grafting the backbone and side chains of a two-segment HIV gp120 epitope, targeted by the cross-neutralizing antibody b12, onto an unrelated scaffold protein. The final scaffolds bound b12 with high specificity and with affinity similar to that of gp120, and crystallographic analysis of a scaffold bound to b12 revealed high structural mimicry of the gp120-b12 complex structure. The method can be generalized to design other functional proteins through backbone grafting.

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.

Computational design of epitope-scaffolds allows induction of antibodies specific for a poorly immunogenic HIV vaccine epitope.

Broadly cross-reactive monoclonal antibodies define epitopes for vaccine development against HIV and other highly mutable viruses. Crystal structures are available for several such antibody-epitope complexes, but methods are needed to translate that structural information into immunogens that re-elicit similar antibodies. We describe a general computational method to design epitope-scaffolds in which contiguous structural epitopes are transplanted to scaffold proteins for conformational stabilization and immune presentation. Epitope-scaffolds designed for the poorly immunogenic but conserved HIV epitope 4E10 exhibited high epitope structural mimicry, bound with higher affinities to monoclonal antibody (mAb) 4E10 than the cognate peptide, and inhibited HIV neutralization by HIV+ sera. Rabbit immunization with an epitope-scaffold induced antibodies with structural specificity highly similar to mAb 4E10, an important advance toward elicitation of neutralizing activity. The results demonstrate that computationally designed epitope-scaffolds are valuable as structure-specific serological reagents and as immunogens to elicit antibodies with predetermined structural specificity.

Challenges for structure-based HIV vaccine design.

PURPOSE OF REVIEW: We review structural information on the native HIV envelope trimer and the known epitopes for broadly neutralizing antibodies and discuss how this structural information should guide the design of more effective immunogens. RECENT FINDINGS: Recent epitope mapping of HIV-positive sera demonstrates that the immune system is able to mount a potent and broadly neutralizing antibody response against conserved elements of the HIV envelope. The structure of trimeric envelope spikes on intact HIV-1 virions (the target of neutralizing antibodies) was determined at low resolution using cryo-electron tomography. Fitting high-resolution crystal structures of monomeric gp120 complexed with different neutralizing ligands into the cryo-electron density maps provides useful models for the native virion trimer and for mechanisms of neutralization. SUMMARY: So far, all attempts to elicit broadly neutralizing antibodies against HIV by immunization have failed. Recent structural information on the virion-associated HIV envelope spike and of the precise interaction of broadly neutralizing mAbs with their epitopes clarifies the steric and geometric constraints faced by antibodies targeting conserved HIV epitopes. Implications for vaccine design are discussed.