Biology subjects

Jeliazkov, J. R.

Publications and source records attributed to Jeliazkov, J. R..

12 recordsLinked to original sources

bioRxiv2024

Humatch - fast, gene-specific joint humanisation of antibody heavy and light chains

1Antibodies are a popular and powerful class of therapeutic due to their ability to exhibit high affinity and specificity to target proteins. However, the majority of antibody therapeutics are not genetically human, with initial therapeutic designs typically obtained from animal models. Humanisation of these precursors is essential to reduce immunogenic risks when administered to humans. Here, we present Humatch, a computational tool designed to offer experimental-like joint humanisation of heavy and light chains in seconds. Humatch consists of three lightweight Convolutional Neural Networks (CNNs) trained to identify human heavy V-genes, light V-genes, and well-paired antibody sequences with near-perfect accuracy. We show that these CNNs, alongside germline similarity, can be used for fast humanisation that aligns well with known experimental data. Throughout the humanisation process, a sequence is guided towards a specific target gene and away from others via multiclass CNN outputs and gene-specific germline data. This guidance ensures final humanised designs do not sit between genes, a trait that is not naturally observed. Humatchs optimisation towards specific genes and good VH/VL pairing increases the chances that final designs will be stable and express well and reduces the chances of immunogenic epitopes forming between the two chains. Humatchs training data and source code are provided open-source. AvailabilitySource code is freely available at github.com/oxpig/Humatch. Data can be found at doi.org/10.5281/zenodo.13764770 Contactdeane@stats.ox.ac.uk Supplementary informationSupplementary data are available at bioRxiv online.

bioinformatics↗

bioRxiv2024

Baselining the Buzz. Trastuzumab-HER2 Affinity, and Beyond!

Strong antibody-antigen binding is the primary consideration when developing an efficacious therapeutic antibody. In recent years, much work has been devoted to applying complex machine learning models to this cause, yet simple baselines are often lacking. Here, we show that the widely used sequence alignment method, BLOSUM, can yield diverse, binder-enriched libraries from a single starting antibody. Using Trastuzumab-HER2 as a model system, we experimentally validated 720 novel designs generated with five different computational methods using surface plasmon resonance. The BLOSUM substitution matrix outperformed all four deep learning design approaches tested, achieving an estimated minimum binder enrichment of 12.5% and producing nine sub-nanomolar binders. These results underscore the importance of comparing against simple baselines and set a benchmark to guide future computational antibody library design. O_FIG O_LINKSMALLFIG WIDTH=200 HEIGHT=141 SRC="FIGDIR/small/586756v2_ufig1.gif" ALT="Figure 1"> View larger version (32K): org.highwire.dtl.DTLVardef@1597ee9org.highwire.dtl.DTLVardef@9af4b6org.highwire.dtl.DTLVardef@1380e61org.highwire.dtl.DTLVardef@1380d29_HPS_FORMAT_FIGEXP M_FIG C_FIG

bioinformatics↗

bioRxiv2023

Conformational sampling and interpolation using language-based protein folding neural networks

Protein language models (PLMs), such ESM2, learn a rich semantic grammar of the protein sequence space. When coupled to protein folding neural networks (e.g., ESMFold), they can facilitate the prediction of tertiary and quaternary protein structures at high accuracy. However, they are limited to modeling protein structures in single states. This manuscript demonstrates that ESMFold can predict alternate conformations of some proteins, including de novo designed proteins. Randomly masking the sequence prior to PLM input returned alternate embeddings that ESMFold sometimes mapped to distinct physiologically relevant conformations. From there, inversion of the ESMFold trunk facilitated the generation of high-confidence interconversion paths between the two states. These paths provide a deeper glimpse of how language-based protein folding neural networks derive structural information from high-dimensional sequence representations, while exposing limitations in their general understanding of protein structure and folding.

biophysics↗

bioRxiv2023

ESMFold Hallucinates Native-Like Protein Sequences

We describe attempts to design protein sequences by inverting the protein structure prediction algorithm ESMFold. State-of-the-art protein structure prediction methods achieve high accuracy by relying on evolutionary patterns derived from either multiple sequence alignments (AlphaFold, RosettaFold) or pretrained protein language models (PLMs; ESMFold, OmegaFold). In principle, by inverting these networks, protein sequences can be designed to fulfill one or more design objectives, such as high prediction confidence, predicted protein binding, or other geometric constraints that can be expressed with loss functions. In practice, sequences designed using an inverted AlphaFold model, termed AFDesign, contain unnatural sequence profiles shown to express poorly, whereas an inverted RosettaFold network has been shown to be sensitive to adversarial sequences. Here, we demonstrate that these limitations do not extend to neural networks that include PLMs, such as ESMFold. Using an inverted ESMFold model, termed ESM-Design, we generated sequences with profiles that are both more native-like and more likely to express than sequences generated using AFDesign, but less likely to express than sequences rescued by the structure-based design method ProteinMPNN. However, the safeguard offered by the PLM came with steep increases in memory consumption, preventing proteins greater than 150 residues from being modeled on a single GPU with 80GB VRAM. During this investigation, we also observed the role played by different sequence initialization schemes, with random sampling of discrete amino acids improving convergence and model quality over any continuous random initialization method. Finally, we showed how this approach can be used to introduce sequence and structure diversification in small proteins such as ubiquitin, while respecting the sequence conservation of active site residues. Our results highlight the effects of architectural differences between structure prediction networks on zero-shot protein design.

bioinformatics↗

bioRxiv2021

Accurately positioning functional residues with robotics-inspired computational protein design

Accurate positioning of functional residues is critical for the design of new protein functions, but has remained difficult because of the prevalence of irregular local geometries in active sites. Here we introduce two computational methods that build local protein geometries from sequence with atomic accuracy: fragment kinematic closure (FKIC) and loophash kinematic closure (LHKIC). FKIC and LHKIC integrate two approaches: robotics-inspired kinematics of protein backbones and insertion of peptide fragments, and show up to 140-fold improvements in native-like predictions over either approach alone. We then integrate these methods into a new design protocol, pull-into-place (PIP), to position functionally important sidechains via design of new structured loop conformations. We validate PIP by remodeling a sizeable active site region in an enzyme and confirming the engineered new conformations of two designs with crystal structures. The described methods can be applied broadly to the design of many new protein geometries and functions.

bioengineering↗

bioRxiv2020

Robustification of RosettaAntibody and Rosetta SnugDock

In recent years, the observed antibody sequence space has grown exponentially due to advances in high-throughput sequencing of immune receptors. The rise in sequences has not been mirrored by a rise in structures, as experimental structure determination techniques have remained low-throughput. Computational modeling, however, has the potential to close the sequence-structure gap. To achieve this goal, computational methods must be robust, fast, easy to use, and accurate. Here we report on the latest advances made in RosettaAntibody and Rosetta SnugDock--methods for antibody structure prediction and antibody-antigen docking. We simplified the user interface, expanded and automated the template database, generalized the kinematics of antibody-antigen docking (which enabled modeling of single-domain antibodies) and incorporated new loop modeling techniques. To evaluate the effects of our updates on modeling accuracy, we developed rigorous tests under a new scientific benchmarking framework within Rosetta. Benchmarking revealed that more structurally similar templates could be identified in the updated database and that SnugDock broadened its applicability without losing accuracy. However, there are further advances to be made, including increasing the accuracy and speed of CDR-H3 loop modeling, before computational approaches can accurately model any antibody.

biophysics↗

bioRxiv2019

Affinity-engineered human antibodies detect celiac disease gluten pMHC complexes and inhibit T-cell activation

Antibodies specific for antigenic peptides bound to major histocompatibility complex (MHC) molecules are valuable tools for studies of antigen presentation. Such T-cell receptor (TCR)-like antibodies may also have therapeutic potential in human disease due to their ability to target disease-associated antigens with high specificity. We previously generated celiac disease (CeD) relevant TCR-like antibodies that recognize the prevalent gluten epitope DQ2.5-glia-1a in complex with HLA-DQ2.5. Here, we report on second-generation high-affinity antibodies towards this epitope as well as a panel of novel TCR-like antibodies to another immunodominant gliadin epitope, DQ2.5-glia-2. The strategy for affinity engineering was based on Rosetta modeling combined with pIX phage display and is applicable to similar protein engineering efforts. We isolated picomolar affinity binders and validated them in Fab and IgG format. Flow cytometry experiments with CeD biopsy material confirm the unique disease specificity of these TCR-like antibodies and reinforce the notion that B cells and plasma cells have a dominant role in gluten antigen presentation in the inflamed CeD gut. Further, the lead candidate 3.C11 potently inhibited CD4+ T-cell activation and proliferation in vitro in an HLA and epitope specific manner, pointing to a potential for targeted disease interception without compromising systemic immunity. Significance StatementConsumption of gluten-containing food drives celiac disease in genetically predisposed individuals. The underlying disease mechanism is not fully understood, but it is strictly dependent on activation of pathogenic T cells. We have engineered high-affinity human antibodies recognizing the T-cell target HLA-DQ2.5 in complex with gluten epitopes and studied cell-specific antigen presentation in patients, which shows that plasma cells and not dendritic cells dominate the inflamed tissue. The only available treatment is lifelong adherence to a gluten-free diet, which is difficult and not effective in all cases. We show that at least one of our antibodies can specifically inhibit activation of pathogenic T-cells in vitro and therefore shows promise for therapy.

immunology↗

bioRxiv2019

A finite vocabulary of antibody-antigen interaction enables predictability of paratope-epitope binding

Antibody-antigen binding relies on the specific interaction of amino acids at the paratope-epitope interface. The predictability of antibody-antigen binding is a prerequisite for de novo antibody and (neo-)epitope design. A fundamental premise for the predictability of antibody-antigen binding is the existence of paratope-epitope interaction motifs that are universally shared among antibody-antigen structures. In the largest set of non-redundant antibody-antigen structures, we identified structural interaction motifs, which together compose a commonly shared structure-based vocabulary of paratope-epitope interactions. We show that this vocabulary enables the machine learnability of antibody-antigen binding on the paratope-epitope level using generative machine learning. The vocabulary (i) is compact, less than 104 motifs, (ii) distinct from non-immune protein-protein interactions, and (iii) mediates specific oligo- and polyreactive interactions between paratope-epitope pairs. Our work successfully leveraged combined structure- and sequence-based learning showing that machine-learning-driven predictive paratope and epitope engineering is feasible.

immunology↗

bioRxiv2019

Novel sampling strategies and a coarse-grained score function for docking homomers, flexible heteromers, and oligosaccharides using Rosetta in CAPRI Rounds 37-45

CAPRI Rounds 37 through 45 introduced larger complexes, new macromolecules, and multi-stage assemblies. For these rounds, we used and expanded docking methods in Rosetta to model 23 target complexes. We successfully predicted 14 target complexes and recognized and refined near-native models generated by other groups for two further targets. Notably, for targets T110 and T136, we achieved the closest prediction of any CAPRI participant. We created several innovative approaches during these rounds. Since Round 39 (target 122), we have used the new RosettaDock 4.0, which has a revamped coarse-grained energy function and the ability to perform conformer selection during docking with hundreds of pre-generated protein backbones. Ten of the complexes had some degree of symmetry in their interactions, so we tested Rosetta SymDock, realized its shortcomings, and developed the next-generation symmetric docking protocol, SymDock2, which includes docking of multiple backbones and induced-fit refinement. Since the last CAPRI assessment, we also developed methods for modeling and designing carbohydrates in Rosetta, and we used them to successfully model oligosaccharide-protein complexes in Round 41. While the results were broadly encouraging, they also highlighted the pressing need to invest in (1) flexible docking algorithms with the ability to model loop and linker motions and in (2) new sampling and scoring methods for oligosaccharide-protein interactions.

biophysics↗

bioRxiv2019

Toward computational design of protein crystals with improved resolution

Substantial advances have been made in the computational design of protein interfaces over the last 20 years. However, the interfaces targeted by design have typically been stable and high affinity. Here, we report the development of a generic computational design method to stabilize the weak interactions at crystallographic interfaces. Initially, we analyzed structures reported in the Protein Data Bank (PDB) to determine whether crystals with more stable interfaces result in higher resolution structures. We found that, for twenty-two variants of a single protein crystallized by a single individual, Rosetta score correlates with resolution. We next developed and tested a computational design protocol, seeking to identify point mutations that would improve resolution on a highly stable variant of staphylococcal nuclease (SNase {Delta}+PHS). Only one of eleven initial designs crystallized, forcing us to re-evaluate our strategy and base our designs on an ensemble of protein backbones. Using this strategy, four of the five designed proteins crystallized. Collecting diffraction data for multiple crystals per design and solving crystal structures, we found that designed crystals improved resolution modestly and in unpredictable ways, including altering crystal space group. Post-hoc, in silico analysis showed that crystal space groups could have been predicted for four of six variants (including WT), but that resolution did not correlate with interface stability, as it did in the preliminary results. Our results show that single point mutations can have significant effects on crystal resolution and space group, and that it is possible to computationally identify such mutations, suggesting a potential design strategy to generate high-resolution protein crystals from poorly diffracting ones.

biophysics↗

bioRxiv2017

Repertoire analysis of antibody CDR-H3 loops suggests affinity maturation does not typically result in rigidification

Antibodies can rapidly evolve in specific response to antigens. Affinity maturation drives this evolution through cycles of mutation and selection leading to enhanced antibody specificity and affinity. Elucidating the biophysical mechanisms that underlie affinity maturation is fundamental to understanding B-cell immunity. An emergent hypothesis is that affinity maturation reduces the conformational flexibility of the antibodys antigen-binding paratope to minimize entropic losses incurred upon binding. In recent years, computational and experimental approaches have tested this hypothesis on a small number of antibodies, often observing a decrease in the flexibility of the Complementarity Determining Region (CDR) loops that typically comprise the paratope and in particular the CDR-H3 loop, which contributes a plurality of antigen contacts. However, there were a few exceptions, and previous studies were limited to a small handful of cases. Here, we determined the structural flexibility of the CDR-H3 loop for thousands of recently-determined homology models of the human peripheral blood cell antibody repertoire using rigidity theory. We found no clear delineation in the flexibility of naive and antigen-experienced antibodies. To account for possible sources of error, we additionally analyzed hundreds of human and mouse antibodies in the Protein Data Bank through both rigidity theory and B-factor analysis. By both metrics, we observed only a slight decrease in the CDR-H3 loop flexibility when comparing affinity-matured antibodies to naive antibodies, and the decrease was not as drastic as previously reported. Further analysis, incorporating molecular dynamics (MD) simulations, revealed a spectrum of changes in flexibility. Our results suggest that rigidification may be just one of many biophysical mechanisms for increasing affinity.

biophysics↗

bioRxiv2017

The Rosetta all-atom energy function for macromolecular modeling and design

Over the past decade, the Rosetta biomolecular modeling suite has informed diverse biological questions and engineering challenges ranging from interpretation of low-resolution structural data to design of nanomaterials, protein therapeutics, and vaccines. Central to Rosettas success is the energy function: amodel parameterized from small molecule and X-ray crystal structure data used to approximate the energy associated with each biomolecule conformation. This paper describes the mathematical models and physical concepts that underlie the latest Rosetta energy function, beta_nov15. Applying these concepts,we explain how to use Rosetta energies to identify and analyze the features of biomolecular models.Finally, we discuss the latest advances in the energy function that extend capabilities from soluble proteins to also include membrane proteins, peptides containing non-canonical amino acids, carbohydrates, nucleic acids, and other macromolecules.

biophysics↗