Evaluation of AlphaFold antibody-antigen modeling with implications for improving predictive accuracy

A comparison of antibody-antigen complex sequence-to-structure prediction methods and their systematic biases

The ability to accurately predict antibody-antigen complex structures from their sequences could greatly advance our understanding of the immune system and would aid in the development of novel antibody therapeutics. There have been considerable recent advancements in predicting protein-protein interactions (PPIs) fueled by progress in machine learning (ML). To understand the current state of the field, we compare six representative methods for predicting antibody-antigen complexes from sequence, including two deep learning approaches trained to predict PPIs in general (AlphaFold-Multimer and RoseTTAFold), two composite methods that initially predict antibody and antigen structures separately and dock them (using antibody-mode ClusPro), local refinement in Rosetta (SnugDock) of globally docked poses from ClusPro, and a pipeline combining homology modeling with rigid-body docking informed by ML-based epitope and paratope prediction (AbAdapt). We find that AlphaFold-Multimer outperformed other methods, although the absolute performance leaves considerable room for improvement. AlphaFold-Multimer models of lower quality display significant structural biases at the level of tertiary motifs (TERMs) toward having fewer structural matches in non-antibody-containing structures from the Protein Data Bank (PDB). Specifically, better models exhibit more common PDB-like TERMs at the antibody-antigen interface than worse ones. Importantly, the clear relationship between performance and the commonness of interfacial TERMs suggests that the scarcity of interfacial geometry data in the structural database may currently limit the application of ML to the prediction of antibody-antigen interactions.

Characterization of a Novel Monoclonal Antibody with High Affinity and Specificity against Aflatoxins: A Discovery from Rosetta Antibody-Ligand Computational Simulation

Aflatoxin B1 (AFB1) accumulates in crops, where it poses a threat to human health. To detect AFB1, anti-AFB1 monoclonal antibodies have been developed and are widely used. While the sensitivity and specificity of these antibodies have been extensively studied, information regarding the atomic-level docking of AFB1 (and its derivatives) with these antibodies is limited. Such information is crucial for understanding the key interactions that are required for high affinity and specificity in aflatoxin binding. First, a 3D comparative model of anti-AFB1 antibody (Ab-4B5G6) was predicted from the sequence using RosettaAntibody. We then utilized RosettaLigand to dock AFB1 onto ten homology models, producing a total of 10,000 binding modes. Interestingly, the best-scoring mode predicted strong interactions involving four sites within the heavy chain: ALA33, ASN52, HIS95, and TRP99. Importantly, these strong binding interactions exclusively involve the variable domain of the heavy chain. The best-scoring mode with AFB1 was also obtained through AF multimer combined with RosettaLigand, and two interactions at TRP and HIS were consistent with those found by Rosetta antibody-ligand computational simulation. The role of tryptophan in π interactions in antibodies was confirmed through mutation experiments, and the resulting mutant (W99A) exhibited a >1000-fold reduction in binding affinity for AFB1 and analogs, indicating the effect of tryptophan on the stability of CDR-H3 region. Additionally, we evaluated the binding of two glycolic acid-derived molecular derivatives (with impaired hydrogen bonding potential), and these derivatives (AFB2-GA and AFG2-GA) demonstrated a very weak binding affinity for Ab-4B5G6. The heavy chain was successfully isolated, and its sensitivity and specificity were consistent with those of the intact antibody. The homology models of variable heavy (VH) single-domain antibodies were established by RosettaAntibody, and the docking analysis revealed the same residues, including Ala, His, and Trp. Compared to the potential binding mode of fragment variable (FV) region, the results from a model of VH indicated that there are seven models involved in hydrophobic interaction with TYR32, which is usually referred to as polar amino acid and has both hydrophobic and hydrophilic features depending on the circumstances. Our work encompasses the entire process of Rosetta antibody-ligand computational simulation, highlighting the significance of variable heavy domain structural design in enhancing molecular interactions.

TEMPRO: nanobody melting temperature estimation model using protein embeddings

Single-domain antibodies (sdAbs) or nanobodies have received widespread attention due to their small size (~ 15 kDa) and diverse applications in bio-derived therapeutics. As many modern biotechnology breakthroughs are applied to antibody engineering and design, nanobody thermostability or melting temperature (Tm) is crucial for their successful utilization. In this study, we present TEMPRO which is a predictive modeling approach for estimating the Tm of nanobodies using computational methods. Our methodology integrates various nanobody biophysical features to include Evolutionary Scale Modeling (ESM) embeddings, NetSurfP3 structural predictions, pLDDT scores per sdAb region from AlphaFold2, and each sequence's physicochemical characteristics. This approach is validated with our combined dataset containing 567 unique sequences with corresponding experimental Tm values from a manually curated internal data and a recently published nanobody database, NbThermo. Our results indicate the efficacy of protein embeddings in reliably predicting the Tm of sdAbs with mean absolute error (MAE) of 4.03 °C and root mean squared error (RMSE) of 5.66 °C, thus offering a valuable tool for the optimization of nanobodies for various biomedical and therapeutic applications. Moreover, we have validated the models' performance using experimentally determined Tms from nanobodies not found in NbThermo. This predictive model not only enhances nanobody thermostability prediction, but also provides a useful perspective of using embeddings as a tool for facilitating a broader applicability of downstream protein analyses.

Integrative modeling meets deep learning: Recent advances in modeling protein assemblies

Recent progress in protein structure prediction based on deep learning revolutionized the field of Structural Biology. Beyond single proteins, it also enabled high-throughput prediction of structures of protein-protein interactions. Despite the success in predicting complex structures, large macromolecular assemblies still require specialized approaches. Here we describe recent advances in modeling macromolecular assemblies using integrative and hierarchical approaches. We highlight applications that predict protein-protein interactions and challenges in modeling complexes based on the interaction networks, including the prediction of complex stoichiometry and heterogeneity.

AlphaFold2 in biomedical research: facilitating the development of diagnostic strategies for disease

Proteins, as the primary executors of physiological activity, serve as a key factor in disease diagnosis and treatment. Research into their structures, functions, and interactions is essential to better understand disease mechanisms and potential therapies. DeepMind's AlphaFold2, a deep-learning protein structure prediction model, has proven to be remarkably accurate, and it is widely employed in various aspects of diagnostic research, such as the study of disease biomarkers, microorganism pathogenicity, antigen-antibody structures, and missense mutations. Thus, AlphaFold2 serves as an exceptional tool to bridge fundamental protein research with breakthroughs in disease diagnosis, developments in diagnostic strategies, and the design of novel therapeutic approaches and enhancements in precision medicine. This review outlines the architecture, highlights, and limitations of AlphaFold2, placing particular emphasis on its applications within diagnostic research grounded in disciplines such as immunology, biochemistry, molecular biology, and microbiology.

Geometric epitope and paratope prediction

MOTIVATION

Identifying the binding sites of antibodies is essential for developing vaccines and synthetic antibodies. In this article, we investigate the optimal representation for predicting the binding sites in the two molecules and emphasize the importance of geometric information.

RESULTS

Specifically, we compare different geometric deep learning methods applied to proteins' inner (I-GEP) and outer (O-GEP) structures. We incorporate 3D coordinates and spectral geometric descriptors as input features to fully leverage the geometric information. Our research suggests that different geometrical representation information is useful for different tasks. Surface-based models are more efficient in predicting the binding of the epitope, while graph models are better in paratope prediction, both achieving significant performance improvements. Moreover, we analyze the impact of structural changes in antibodies and antigens resulting from conformational rearrangements or reconstruction errors. Through this investigation, we showcase the robustness of geometric deep learning methods and spectral geometric descriptors to such perturbations.

AVAILABILITY AND IMPLEMENTATION

The python code for the models, together with the data and the processing pipeline, is open-source and available at https://github.com/Marco-Peg/GEP.

A Comparison of Antibody-Antigen Complex Sequence-to-Structure Prediction Methods and their Systematic Biases

The ability to accurately predict antibody-antigen complex structures from their sequences could greatly advance our understanding of the immune system and would aid in the development of novel antibody therapeutics. There have been considerable recent advancements in predicting protein-protein interactions (PPIs) fueled by progress in machine learning (ML). To understand the current state of the field, we compare six representative methods for predicting antibody-antigen complexes from sequence, including two deep learning approaches trained to predict PPIs in general (AlphaFold-Multimer, RoseTTAFold), two composite methods that initially predict antibody and antigen structures separately and dock them (using antibody-mode ClusPro), local refinement in Rosetta (SnugDock) of globally docked poses from ClusPro, and a pipeline combining homology modeling with rigid-body docking informed by ML-based epitope and paratope prediction (AbAdapt). We find that AlphaFold-Multimer outperformed other methods, although the absolute performance leaves considerable room for improvement. AlphaFold-Multimer models of lower-quality display significant structural biases at the level of tertiary motifs (TERMs) towards having fewer structural matches in non-antibody containing structures from the Protein Data Bank (PDB). Specifically, better models exhibit more common PDB-like TERMs at the antibody-antigen interface than worse ones. Importantly, the clear relationship between performance and the commonness of interfacial TERMs suggests that scarcity of interfacial geometry data in the structural database may currently limit application of machine learning to the prediction of antibody-antigen interactions.

The power and pitfalls of AlphaFold2 for structure prediction beyond rigid globular proteins

Artificial intelligence-driven advances in protein structure prediction in recent years have raised the question: has the protein structure-prediction problem been solved? Here, with a focus on nonglobular proteins, we highlight the many strengths and potential weaknesses of DeepMind's AlphaFold2 in the context of its biological and therapeutic applications. We summarize the subtleties associated with evaluation of AlphaFold2 model quality and reliability using the predicted local distance difference test (pLDDT) and predicted aligned error (PAE) values. We highlight various classes of proteins that AlphaFold2 can be applied to and the caveats involved. Concrete examples of how AlphaFold2 models can be integrated with experimental data in the form of small-angle X-ray scattering (SAXS), solution NMR, cryo-electron microscopy (cryo-EM) and X-ray diffraction are discussed. Finally, we highlight the need to move beyond structure prediction of rigid, static structural snapshots toward conformational ensembles and alternate biologically relevant states. The overarching theme is that careful consideration is due when using AlphaFold2-generated models to generate testable hypotheses and structural models, rather than treating predicted models as de facto ground truth structures.

Nanobody engineering: computational modelling and design for biomedical and therapeutic applications

Nanobodies, the smallest functional antibody fragment derived from camelid heavy-chain-only antibodies, have emerged as powerful tools for diverse biomedical applications. In this comprehensive review, we discuss the structural characteristics, functional properties, and computational approaches driving the design and optimisation of synthetic nanobodies. We explore their unique antigen-binding domains, highlighting the critical role of complementarity-determining regions in target recognition and specificity. This review further underscores the advantages of nanobodies over conventional antibodies from a biosynthesis perspective, including their small size, stability, and solubility, which make them ideal candidates for economical antigen capture in diagnostics, therapeutics, and biosensing. We discuss the recent advancements in computational methods for nanobody modelling, epitope prediction, and affinity maturation, shedding light on their intricate antigen-binding mechanisms and conformational dynamics. Finally, we examine a direct example of how computational design strategies were implemented for improving a nanobody-based immunosensor, known as a Quenchbody. Through combining experimental findings and computational insights, this review elucidates the transformative impact of nanobodies in biotechnology and biomedical research, offering a roadmap for future advancements and applications in healthcare and diagnostics.

Overview of AlphaFold2 and breakthroughs in overcoming its limitations

Predicting three-dimensional (3D) protein structures has been challenging for decades. The emergence of AlphaFold2 (AF2), a deep learning-based machine learning method developed by DeepMind, became a game changer in the protein folding community. AF2 can predict a protein's three-dimensional structure with high confidence based on its amino acid sequence. Accurate prediction of protein structures can dramatically accelerate our understanding of biological mechanisms and provide a solid foundation for reliable drug design. Although AF2 breaks through the barriers in predicting protein structures, many rooms remain to be further studied. This review provides a brief historical overview of the development of protein structure prediction, covering template-based, template-free, and machine learning-based methods. In addition to reviewing the potential benefits (Pros) and considerations (Cons) of using AF2, this review summarizes the diverse applications, including protein structure predictions, dynamic changes, point mutation, integration of language model and experimental data, protein complex, and protein-peptide interaction. It underscores recent advancements in efficiency, reliability, and broad application of AF2. This comprehensive review offers valuable insights into the applications of AF2 and AF2-inspired AI methods in structural biology and its potential for clinically significant drug target discovery.

Structural characterization and AlphaFold modeling of human T cell receptor recognition of NRAS cancer neoantigens

T cell receptors (TCRs) that recognize cancer neoantigens are important for anti-cancer immune responses and immunotherapy. Understanding the structural basis of TCR recognition of neoantigens provides insights into their exquisite specificity and can enable design of optimized TCRs. We determined crystal structures of a human TCR in complex with NRAS Q61K and Q61R neoantigen peptides and HLA-A1 MHC, revealing the molecular underpinnings for dual recognition and specificity versus wild-type NRAS peptide. We then used multiple versions of AlphaFold to model the corresponding complex structures, given the challenge of immune recognition for such methods. Interestingly, one implementation of AlphaFold2 (TCRmodel2) was able to generate accurate models of the complexes, while AlphaFold3 also showed strong performance, although success was lower for other complexes. This study provides insights into TCR recognition of a shared cancer neoantigen, as well as the utility and practical considerations for using AlphaFold to model TCR-peptide-MHC complexes.

Experimental Structures of Antibody/MHC-I Complexes Reveal Details of Epitopes Overlooked by Computational Prediction

mAbs to MHC class I (MHC-I) molecules have proved to be crucial reagents for tissue typing and fundamental studies of immune recognition. To augment our understanding of epitopic sites seen by a set of anti-MHC-I mAb, we determined X-ray crystal structures of four complexes of anti-MHC-I Fabs bound to peptide/MHC-I/β2-microglobulin (pMHC-I). An anti-H2-Dd mAb, two anti-MHC-I α3 domain mAbs, and an anti-β2-microglobulin mAb bind pMHC-I at sites consistent with earlier mutational and functional experiments, and the structures explain allelomorph specificity. Comparison of the experimentally determined structures with computationally derived models using AlphaFold Multimer showed that although predictions of the individual pMHC-I heterodimers were quite acceptable, the computational models failed to properly identify the docking sites of the mAb on pMHC-I. The experimental and predicted structures provide insight into strengths and weaknesses of purely computational approaches and suggest areas that merit additional attention.

Flexible protein-protein docking with a multitrack iterative transformer

Conventional protein-protein docking algorithms usually rely on heavy candidate sampling and reranking, but these steps are time-consuming and hinder applications that require high-throughput complex structure prediction, for example, structure-based virtual screening. Existing deep learning methods for protein-protein docking, despite being much faster, suffer from low docking success rates. In addition, they simplify the problem to assume no conformational changes within any protein upon binding (rigid docking). This assumption precludes applications when binding-induced conformational changes play a role, such as allosteric inhibition or docking from uncertain unbound model structures. To address these limitations, we present GeoDock, a multitrack iterative transformer network to predict a docked structure from separate docking partners. Unlike deep learning models for protein structure prediction that input multiple sequence alignments, GeoDock inputs just the sequences and structures of the docking partners, which suits the tasks when the individual structures are given. GeoDock is flexible at the protein residue level, allowing the prediction of conformational changes upon binding. On the Database of Interacting Protein Structures (DIPS) test set, GeoDock achieves a 43% top-1 success rate, outperforming all other tested methods. However, in the standard DIPS train/test splits, we discovered contamination of close homologs in the training set. After decontaminating the training set, the success rate is 31%. On the DB5.5 test set and a benchmark dataset of antibody-antigen complexes, GeoDock outperforms the deep learning models trained using the same dataset but falls behind most of the conventional methods and AlphaFold-Multimer. GeoDock attains an average inference speed of under 1 s on a single GPU, enabling its application in large-scale structure screening. Although binding-induced conformational changes are still a challenge owing to limited training and evaluation data, our architecture sets up the foundation to capture this backbone flexibility. Code and a demonstration Jupyter notebook are available at https://github.com/Graylab/GeoDock.

Applications and challenges in designing VHH-based bispecific antibodies: leveraging machine learning solutions

The development of bispecific antibodies that bind at least two different targets relies on bringing together multiple binding domains with different binding properties and biophysical characteristics to produce a drug-like therapeutic. These building blocks play an important role in the overall quality of the molecule and can influence many important aspects from potency and specificity to stability and half-life. Single-domain antibodies, particularly camelid-derived variable heavy domain of heavy chain (VHH) antibodies, are becoming an increasingly popular choice for bispecific construction due to their single-domain modularity, favorable biophysical properties, and potential to work in multiple antibody formats. Here, we review the use of VHH domains as building blocks in the construction of multispecific antibodies and the challenges in creating optimized molecules. In addition to exploring traditional approaches to VHH development, we review the integration of machine learning techniques at various stages of the process. Specifically, the utilization of machine learning for structural prediction, lead identification, lead optimization, and humanization of VHH antibodies.