Artificial intelligence-driven computational methods for antibody design and optimization

ABSTRACT

Antibodies play a crucial role in our immune system. Their ability to bind to and neutralize pathogens opens opportunities to develop antibodies for therapeutic and diagnostic use. Computational methods capable of designing antibodies for a target antigen can revolutionize drug discovery, reducing the time and cost required for drug development. Artificial intelligence (AI) methods have recently achieved remarkable advancements in the design of protein sequences and structures, including the ability to generate scaffolds for a given motif and binders for a specific target. These generative methods have been applied to antigen-conditioned antibody design, with experimental binding confirmed for de novo-designed antibodies. This review surveys current AI methods used in antibody development, focusing on those for antigen-conditioned antibody design. The results obtained by AI-based methodologies in antibody and protein research suggest a promising direction for generating de novo binders for various target antigens.

Introduction

Antibodies are essential for defending the body against pathogenic infection. B cells naturally produce them in vivo through adaptive immunity in response to bacterial or viral infections and vaccinations. Antibodies recognize and bind to specific targets, known as antigens, where the binding occurs between the antibody’s paratope and the antigen’s epitope (Figure 1). The diverse repertoire of antibodies and their antigen-specificity to numerous antigen molecules arise from somatic rearrangements and hypermutations in antibody-coding genes, followed by clonal selection and maturation of lymphocytes.¹

Figure 1.

Antibody structure. Antibodies, or immunoglobulins, are Y-shaped glycoproteins composed of two identical heavy and light chains, respectively. The variable regions of the heavy and light chains (V_H and V_L) in the antigen-binding domain (F_ab) contain the hypervariable antigen-recognition regions known as complementarity-determining regions (CDRs), with three CDRHs in V_H and three CDRLs in V_L. Antigen-antibody complex interactions occur between the epitope and paratope of the antigen and antibody, respectively. The biological activity mediating domain (F_C) is made up of the C_H2 and C_H3 regions of the two heavy chains, providing a binding site for endogenous F_C receptors on lymphocytes to facilitate immune responses. Additionally, dyes and enzymes can be covalently linked to F_C for experimental visualization.

Using monoclonal antibodies in biochemistry, molecular/cellular biology, and medical imaging has significantly advanced biological and medical research by unveiling complex phenomena in the life sciences. The ability of monoclonal antibodies to target specific, unique antigens has expanded their potential for clinical applications, including in diagnostics and therapeutics,² and resulted in the development of over 170 antibody therapeutics that are licensed globally.³ Monoclonal antibody discovery efforts have traditionally relied on experiments such as hybridoma screening after immunization and phage display techniques.^4,5 More recently, the development of humanized and fully human antibodies through techniques such as complementarity-determining region (CDR) grafting, use of transgenic animals harboring human immunoglobulin (Ig) genes, next-generation sequencing of human B-cells, and proteomic analysis of the Ig repertoire from human plasma has accelerated the development process of therapeutic antibodies with improved efficacy and fewer side effects.^6–8 However, these methods are still labor-intensive, costly, time-consuming, and often produce suboptimal leads. More importantly, it remains challenging to discover epitope-specific antibodies that can target therapeutically relevant regions of antigens through these experimental approaches.

Recent advances in artificial intelligence (AI) have revolutionized computational structural biology with remarkable results in tasks such as protein structure prediction, protein sequence design, and de novo design of protein scaffolds and binders.^9–15 These models are designed primarily for general proteins but have also been applied specifically to antibody design.¹² Additionally, antibody-specific AI-based methods have been proposed to improve antibody properties, such as binding affinity and developability, and design novel monoclonal antibodies specific to target antigens or epitopes of interest with desirable properties. This review covers recent AI-based methods relevant to in silico antibody design, including structure prediction, representation learning, sequence design, and unconditioned and antigen-conditioned antibody design (Figure 2). Particularly, we detail antigen-conditioned methods, where the antigen information is explicitly inputted into the antibody design pipeline, including methods that have been experimentally validated for the de novo design of single-domain and single-chain variable fragment antibodies. While none of the antibodies designed using AI-based methods have yet reached clinical application, AI-based antibody design is a rapidly emerging field and represents a promising direction for generating de novo antibodies for diagnostics and therapeutics.

Figure 2.

Subproblems in computational antibody design pipelines include a, antibody structure prediction, b, antibody representation learning, c, antibody sequence design, d, unconditioned antibody design, and e, antigen-conditioned antibody design. Full antibody design is illustrated in (d) and (e), but partial antibody design methods have also been developed. Here, seq-str stands for sequence and/or structure.

Antibody structure prediction

In the structure prediction problem, given a one-dimensional (1D) protein sequence, the model should predict its corresponding three-dimensional (3D) structure, as shown in Figure 2a. The development of AlphaFold2 (AF2) has significantly improved the accuracy of modeling various protein structures and their interactions without the need for experimental determination.⁹ As shown in Table 1, since AF2, several deep learning-based structure predictors have been developed.^{11,12,18–20} These methods have been mainly trained using datasets for general proteins, though methods like AlphaFold3 (AF3),¹² RoseTTAFoldNA (RFNA),²⁰ and RoseTTAFold All-Atom (RFAA)¹¹ have also been developed for predicting proteins in complexes with nucleic acids or other biomolecules. While these methods can be applied for antibody structure prediction, Polonsky et al.¹⁶ demonstrated the capabilities and limitations of AF2⁹ and ESMFold¹⁸ in modeling both antibody structures and antigen-antibody complexes, underscoring the need for further advancements in the area.

Table 1.

Selected AI-based protein (top) and antibody (bottom) structure prediction methods, each sorted by chronological order. Methods that support antibody structure prediction in complex with antigen are prefixed with an asterisk (*). The antibody-specific methods predict the structure of the entire F_V unless otherwise stated.

Method Name	Dataset	Description	Open Source
Protein Structure Prediction
*AF29	PDB, Uniclust30	Architecture contains evoformer and structure modules. Uses recycles, triangle axial attention, IPA mechanisms, and FAPE loss. Evaluated for antibodies by Polonsky et al.16 and Yin et al.17	https://github.com/deepmind/alphafold.
*ESMFold18	PDB, UniRef50	Removes the need for multiple sequence alignment (MSA) input in AF29 and accelerates inference by leveraging the ESM-218 protein language model (PLM).	https://github.com/facebookresearch/esm.
*RF219	PDB, UniRef50	Extends the three-track network in RF10 by using computationally efficient biaxial attention and SE(3)-equivariant transformer and is trained with a FAPE loss. Competitive performance with AF2.	https://github.com/uw-ipd/RoseTTAFold2.
RFNA20	PDB	Extends RF219 for nucleic acids and protein-nucleic acid complexes.	https://github.com/uw-ipd/RoseTTAFold2.NA
*RFAA11	PDB, Cambridge Structural Database	Generalizes RFNA20 for biomolecular assemblies including nucleic acids, small molecules, metal ions, and covalent modifications.	https://github.com/baker-laboratory/RoseTTAFold-All-Atom.
*AF312	PDB, MGnify, Rfam, JASPAR 9, synthetic negative data	Structure prediction of biomolecular assemblies using a pairformer block and substituting the structure module in AF29 by a diffusion model for all-atom structure generation.	https://github.com/google-deepmind/alphafold3.
Antibody Structure Prediction
DeepH321	SAbDab	A deep ResNet method coupled with Rosetta-based energy optimization to predict CDRH3 structures.	https://github.com/Graylab/deepH3-distances-orientations.
DeepAb22	OAS	A deep ResNet method using embeddings from an LSTM-based antibody language model (ALM) and output attention mechanisms coupled with Rosetta-based energy optimization to predict FV structures.	https://github.com/RosettaCommons/DeepAb.
ABlooper23	SAbDab	E(n)-equivariant graph neural network (EGNN) method to directly predict the backbone structure of CDR loop regions.	https://github.com/brennanaba/ABlooper.
DeepSCAb24	SAbDab	Extends the DeepH321 architecture for the entire FV. It shows accurate side-chain modeling by predicting side-chain dihedrals using a rotamer module.	https://github.com/Graylab/DeepSCAb.
EquiFold25	SAbDab	SE(3)-EGNN method with iterative refinement trained using a FAPE loss. Also available to general proteins.	https://github.com/Genentech/equifold.
tFold-Ab26	SAbDab	Combines a pretrained ProtXLNet PLM with an architecture similar to AF2 to predict the structure of antibodies and nanobodies.	https://drug.ai.tencent.com/en.
IgFold27	SAbDab, OAS	Combines a pretrained AntiBERTy ALM, graph neural networks (GNNs) with transformers, and IPA modules. Rosetta refinement is added to predict accurate antibody structures.	https://github.com/Graylab/IgFold.
xTrimoABFold28	PDB	Combines a pretrained transformer-based ALM and AF2 modules without using MSA.	N/A
AbFold29	SAbDab, OAS	Combines AF2 modules with point cloud embeddings from a preprocessed antibody backbone structure.	N/A
ABodyBuilder2 of ImmuneBuilder30	SAbDab	Leverages the structure module in AF2 to predict antibody structures using one-hot sequence representations.	https://github.com/oxpig/ImmuneBuilder.
*tFold31	UniRef50, PDB, PPI (custom), OAS, SAbDab	Replaces ProtXLNet with ESM-PPI in tFold-Ab26 and extends to tFold-Ag for predicting antigen-antibody and antigen-nanobody complex structures.	https://github.com/TencentAI4S/tfold.
*RF2 Antibody32	antibody structures	RF219 fine-tuned on native antibody structures.	N/A
ABodyBuilder333	SAbDab	Enhance the architecture in ABodyBuilder2 by adding LM embeddings, improved relaxation, and optimizations from OpenFold.	https://github.com/Exscientia/ABodyBuilder3.

The following structure predictors for general proteins are omitted: RF,¹⁰ OmegaFold,³⁴ Uni-Fold,³⁵ OpenFold,³⁶ and HelixFold-Multimer.³⁷

Accurate antibody structure prediction is essential for designing antibodies with optimal binding affinity and specificity to target antigens. As summarized in Table 1, various machine learning (ML) networks such as DeepH3,²¹ DeepAb,²² ABlooper,²³ DeepSCAb,²⁴ EquiFold,²⁵ tFold-Ab,²⁶ IgFold,²⁷ xTrimoABFold,²⁸ AbFold,²⁹ ImmunoBuilder,³⁰ tFold,³¹ RoseTTAFold2 (RF2) Antibody,³² and ABodyBuilder3³³ have been developed for this purpose. Aside from AF2⁹ and RoseTTAFold (RF)¹⁰ for structure prediction of general proteins, separate models are necessary due to the unique nature of antibodies, especially the hypervariability of the CDRs. Attempts have also been made to improve structure prediction models by fine-tuning them with antibody-specific data. For example, RF2 Antibody³² fine-tuned the RF2¹⁹ model on antibody structures, resulting in superior performance compared to other antibody structure prediction methods. Nevertheless, most of the aforementioned methods are designed to predict apo-antibody structures, particularly the variable domain (F_V), with only a few capable of predicting antigen–antibody complex structures.

While RF2 Antibody³² and AF3¹² substantially improved the structure prediction accuracy of the antibodies and antigen–antibody complexes, challenges persist due to the limited availability of antibody structures in current databases, such as the Structural Antibody Database (SAbDab),³⁸ the overwhelming diversity and flexibility of the six CDR loops, and the absence of evolutionary histories of CDR loops critical for structure prediction using multiple sequence alignment (MSA). Data augmentation can be considered to address the scarcity of antibody experimental data. This can involve leveraging loop-like regions from the Protein Data Bank (PDB)³⁹ or using synthetic antigen–antibody complexes generated by methods such as Absolut!⁴⁰ or diffusion-based models. These augmented datasets can be used to train prediction networks specifically for antigen–antibody complexes or fine-tune structure predictors initially trained on general proteins.

Antibody representation learning

Representation learning has achieved breakthroughs with the recent development of AI-based language models (LMs) applied to natural language.⁴¹ Given the similarity between natural language and protein sequences, various LMs have also been proposed for sequence-based protein and antibody representation learning^18,42–58 as summarized in Table 2 and detailed by Ferruz et al.⁷⁷ Additionally, a recent trend includes the development of structure-based and multimodal representation learning methods,^{59–61,72–76} with notable works including GearNet⁶¹ and ESM3.⁵⁷ For proteins, representation learning involves transforming complex protein data, such as sequences, 3D structures, or both, into a simplified, low-dimensional representation with minimum information loss (Figure 2b). This latent space can be used for various downstream tasks, such as binding affinity prediction⁶⁶ or protein engineering.⁸⁰ Sequence and structure antibody datasets used to train these methods are reviewed by Akbar et al.⁸¹ and Kim et al.⁸²

Table 2.

Selected AI-based protein (top) and antibody (bottom) representation learning methods, each sorted by chronological order. Methods that can learn the representation of antigen–antibody complexes are prefixed with an asterisk (*).

Method Name	Dataset	Description	Open Source
Protein Representation Learning
ESM18,42,43	UniRef50/90, Uniclust30	Protein language models (PLMs) trained based on the Transformer architecture. Three models are ESM-MSA-1b, ESM-1 v, and ESM-2.	https://github.com/facebookresearch/esm.
GearNet59	AlphaFold DB (AFDB)	Protein structure encoder trained using a relational GNN.	https://github.com/DeepGraphLearning/GearNet.
Foldseek60	AFDB, SCOPe40	Trains a 3D interaction (3Di) alphabet to describe the structural features of proteins.	https://github.com/steineggerlab/foldseek.
ESM361	UniRef, MGnify90, JGI, OAS, PDB, AFDB, ESM Metagenomic Atlas, InterPro	Multimodal LM over sequence, structure, and function tokenized representations. The three representations are fused into a single latent space.	https://github.com/evolutionaryscale/esm.
Antibody Representation Learning
AntiBERTy51	OAS	ALM trained based on the BERT architecture.	https://github.com/jeffreyruffolo/AntiBERTy.
AntiBERTa52	OAS, SAbDab	ALM trained based on the RoBERTa58 architecture.	https://github.com/alchemab/antiberta.
AbLang53	OAS	ALM trained based on the RoBERTa58 architecture. Two independent models are trained for the heavy and light chains, respectively.	https://github.com/oxpig/AbLang.
Ig-VAE63	AbDb	VAE trained to reconstruct antibody and nanobody structures.	https://github.com/ProteinDesignLab/IgVAE.
AbBERT54	OAS	Transformer-based ALM trained on unpaired heavy/light sequences.	N/A
IgLM55	OAS	ALM trained conditionally on the chains and species of the antibody sequence based on the Generative Pre-trained Transformer (GPT)-2 architecture with bidirectional context.	https://github.com/Graylab/IgLM.
dWJS64	OAS, hu4D5 antibody mutant dataset	Antibody sequence encoder using discrete energy-based models.	https://github.com/Genentech/walk-jump.
ReprogBert65	SAbDab, RAbD, Coronavirus Antibody Database (CoV-AbDab)	Repurposes an English LM for learning protein and antibody embeddings via transfer learning.	https://github.com/IBM/ReprogBERT.
*GearBind66	SKEMPI 2.0, CATH 4.3	Antibody structure encoder based on the architecture from GearNet55 trained to predict binding free energy.	N/A
AntiBERTa2, AntiBERTa2- CSSP52	OAS, custom sequence dataset, SAbDab, paired B cell receptor sequence dataset, HER2 binding dataset (HER2)	RoFormer63−based ALM architecture pretrained using masked language modeling and CSSP approaches, respectively.	https://huggingface.co/alchemab.
IgBert, IgT557	OAS	Two ALMs pretrained using BERT64 and T565 architectures from ProtBert and ProtT5 in ProtTrans.40	https://huggingface.co/Exscientia.
FAbCon58	OAS, custom sequence dataset, paired B cell receptor sequence dataset, HER2, SARS-CoV-2 binding dataset, IL-6 binding dataset	ALM based on the Falcon large LM (LLM)66 architecture trained using causal language modeling (CLM).	https://huggingface.co/alchemab.

Typical LMs and sequence encoders take the sequence as an input. Structure encoders take the structure as an input. Multimodal LMs and sequence-structure encoders take both sequence and structure as inputs. The following sequence-based representation learning methods for general proteins are omitted: ProGen,¹⁶¹ ProtTrans,⁴⁴ ProtGPT2,⁴⁵ CARP,⁴⁶ LM-Design⁴⁶, xTrimoPGLM,⁴⁸ DPLM,⁴⁵ and ESM All-Atom (ESM-AA).⁵⁰ The following structure-based and multimodal representation learning methods for general proteins are omitted: ProstT5,⁷² SaProt,⁷³ FoldToken2,⁷⁴ ProTokens,⁷⁵ and ProTrek.⁷⁶

Given the vast availability of antibody sequence data, especially in the Observed Antibody Space (OAS) database,⁸³ antibody-specific LMs have been vastly investigated^51–58 (Table 2). Like their natural language counterparts, these antibody language models (ALMs) are mostly based on Transformers^67,84 and BERT^62,68 architectures. These methods have attempted to learn the representations of the antibody heavy and light chains, both jointly and independently. In addition to general representation learning, recent efforts have explored conditional sequence generation for antibody design. For example, IgLM⁵⁵ is a representation learning method enabling the generation of human-like antibody sequences by prefixing two tokens, indicating species and chain type (heavy or light), to the input sequence during training. Models trained for general proteins have also been applied successfully to antibodies. In a work by Hie et al.,⁸⁵ ESM-2¹⁸ is leveraged to propose mutations for human antibodies, achieving enhanced binding affinity through two rounds of mutations verified in wet lab experiments. Finally, supervised fine-tuning of pretrained ALMs has been demonstrated to improve antigen-specificity prediction⁸⁶ and applied to train scoring networks to evaluate candidate mutants.^{54,69, 70, 71}

Structural encoders trained for antibodies include Ig-VAE⁶³ and GearBind⁶⁶; the latter proposes a representation that is shown to successfully improve the binding affinity of an antibody to the spike protein of the SARS-CoV-2 Omicron strain. As discussed by Cai et al.,⁶⁶ structural encoders offer low-dimensional representations that can be useful to narrow down the search for antibody optimization or to improve the search process in antigen-conditioned antibody design. Furthermore, multimodal representation learning methods that integrate sequence and structural contexts have been effectively adapted for antibody-related tasks. Among them, proseLM⁷⁸ and the method by Shanker et al.⁷⁹ have been applied to antibody sequence design for affinity maturation and potency enhancement and are discussed further in the Antibody Sequence Design section below. In the case of CSSP,⁵⁶ multimodal contrastive learning is used to represent antibody sequence and structure in a common latent space, as in CLIP,⁸⁷ leading to improved accuracy in downstream binding prediction tasks. However, structure-based representation learning for antibodies is challenging given the limited structural data available, mostly contained in SAbDab.³⁸ Recently, synthetic structural antibody datasets have been made available using structure prediction methods.²⁷ Synthetic data associated with high-confidence metrics can be used as an effective data augmentation methodology for training structural encoders, similarly to the methodology for training the structural decoder in RF2.¹⁹

Antibody sequence design

In the sequence design problem, the model should predict a sequence of amino acid types that fold into a given protein 3D structure, as shown in Figure 2c. Following the seminal work by Ingraham et al.⁸⁸ using graph neural networks (GNNs) and attention mechanisms for autoregressive sequence prediction, a diverse set of protein sequence design methods for a fixed backbone structure has been proposed^13,89,90 (Table 3). Among these, ProteinMPNNN¹³ uses a methodology similar to Ingraham et al.⁸⁸ with a different set of features and can sample the output sequence in random orders. ProteinMPNN can handle complex structures for binder design, keep parts of the sequence fixed, and has been validated in wet lab experiments in previous works.^13,100 Technical advances such as inferring the sequence in a one-shot manner,^98,99 adding knowledge-based structural features,¹⁰¹ and proposing graph-denoising diffusion models for sequence design⁹⁰ were also investigated for sequence design methods. As opposed to general proteins, the design of antibody sequences is typically directed to the hypervariable CDRs. Models trained on general proteins are leveraged for this purpose with proper adjustments to focus on designing the F_V sequence, particularly the CDRs, and by fine-tuning or retraining on antibody datasets. Additional methods have been proposed to train antibody-specific models for sequence design.^{78,79,92,94–97} These methods are mostly designed for monomer scenarios, in which the sequence of antibodies is designed without any information regarding their target antigen.

Table 3.

Selected AI-based protein (top) and antibody (bottom) sequence design methods, each sorted by chronological order. Methods that support antibody sequence design in complex with antigen are prefixed with an asterisk (*).

Method Name	Dataset	Description	Open Source
Protein Sequence Design
Ingraham et al.88	CATH 4.2	Autoregressive sequence design using GNNs and attention mechanisms with a novel structured transformer model.	https://github.com/jingraham/neurips19-graph-protein-design.
*ESM-IF189	CATH 4.3, UniRef50	Leverages a geometric vector perceptron (GVP) architecture combined with GNNs or transformers. The training data is augmented with AF29 predictions.	https://github.com/facebookresearch/esm.
*ProteinMPNN13	CATH 4.2, PDB	Extends Ingraham et al.81 by utilizing different edge features, adding multichain support, and implementing features for flexible design.	https://github.com/dauparas/ProteinMPNN.
GraDe_IF90	CATH 4.2	Diffusion models by roto-translation equivariant graph neural network and encoding prior knowledge-based information.	https://github.com/ykiiiiii/GraDe_IF.
Antibody Sequence Design
*DiffAb91	SAbDab	Diffusion-based antibody co-design method that can be used for sequence design by fixing the antibody structure.	https://github.com/luost26/diffab.
FVHallucinator92	–	Leverages pretrained DeepAb22 and the trDesign93 approach for FV sequence design.	https://github.com/RosettaCommons/FvHallucinator.
Ing_Ab94	SAbDab	Architecture of Ingraham et al.95 retrained on SAbDab.38	N/A
*AbMPNN95	SAbDab, OAS	ProteinMPNN13 fine-tuned for antibodies using SAbDab38 and the ImmuneBuilder30 dataset predicted from OAS.	https://zenodo.org/records/8164693.
*IgDesign96	PDB, SAbDab	Proposes IgMPNN, which is ProteinMPNN13 without pairwise distance including side chain atoms, and extends LM-Design46 for sequence-based antigen-conditioned antibody CDRH design.	N/A
*AntiFold90	SAbDab, OAS	ESM-IF189 fine-tuned for antibodies using SAbDAb34 and the ImmuneBuilder30 dataset predicted from OAS.	https://github.com/oxpig/AntiFold.
*proseLM78	CATH 4.2, PDB, SAbDab	Adapt a PLM to incorporate structural and functional information, including backbone coordinates and nearby context, for sequence design. Also applicable to general proteins.	https://github.com/Profluent-AI/proseLM-public.

The following sequence design methods for general proteins are omitted: PiFold,⁹⁸ KW-Design⁸⁹, Frame2seq,⁹⁹ and Shanker et al.⁷⁹

In the current drug discovery pipeline, a set of antibodies is discovered using traditional methods. The initial sequences of antibody candidates are then further humanized and optimized to improve their therapeutic efficacy and developability. When the structure of these antibodies in complex with the target antigen is determined experimentally, the optimization can be defined as a sequence design problem using the antigen–antibody complex as a fixed backbone structure. Methods that can be applied to antigen-conditioned methods for antibody sequence design are mostly the ones trained for or adapted from general proteins,^13,95,96 structure-informed LMs,^78,79 or those originally developed for sequence-structure co-design that can be adapted for sequence design by fixing the antibody structure.⁸⁴ As an alternative, various methods^91,100–104 have attempted to co-design antibody sequences and structures to tackle this need, which will be further discussed in the Antibody Design section below. Recently, in vitro results confirmed that, given a target antigen–antibody complex structure, sequence design methods can optimize the initial antibodies. In particular, proseLM⁷⁸ and a recent work by Shanker et al.⁷⁹ have demonstrated using structural context-informed LMs to improve the binding affinity of antibodies in wet lab experiments. In IgDesign,⁹⁶ the sequence and structure of the antigen and the framework regions (FRs) of an antibody template are given as context to design the CDRs of the V_H using an antibody-specific sequence design model, IgMPNN, applied in a methodology similar to LM-Design⁴⁶. The experimental results obtained in these methods^78,79,96 signify their potential in antibody design pipelines for cases in which the experimental complex structure is available.

Antibody design

The antibody design problem includes two cases: 1) unconditioned antibody design, in which there is no explicit antigen information given as input to the model (Figure 2d), and 2) antigen-conditioned antibody design, in which the antigen sequence and/or structure information are explicitly provided as input to the model (Figure 2e). Additionally, antibody design methods have three classes depending on the generated output type. This classification is shown in Figure 3, where: 1) sequence-based methods generate only the sequence, 2) structure-based methods generate only the structure and require a subsequent sequence design for the generated backbone structure, as discussed above, and 3) sequence-structure co-design methods jointly generate sequence and structure. The various antibody design approaches are summarized in Table 4 and discussed below.

Figure 3.

Classification of antibody design methods based on the generated output type. In sequence-based design (a), only the antibody amino acid sequence is generated. In structure-based design (b), the model generates only the antibody structure, which can be followed by sequence design. In sequence-structure co-design (c), the method jointly generates the antibody sequence and structure (seq-str). Partial antibody design is illustrated here, but this classification also applies to full antibody design. The antigen enclosed in the square brackets is explicitly inputted to the model only in antigen-conditioned design. When applicable, the antigen structure input can be a free entity or docked to the antibody, depending on the method. Generally, unconditioned design models are trained on antibody datasets and output antibody candidates. In contrast, antigen-conditioned design models are trained on antigen-antibody complex datasets and output antigen-antibody complex candidates.

Table 4.

Selected AI-based unconditioned (top) and antigen-conditioned (bottom) antibody design methods, each sorted by chronological order. Methods accompanied by in vitro evaluations are prefixed with an asterisk (*).

Method Name	Range	Dataset & PTM4	Input	Output	Description	Open Source
Unconditioned Antibody Design
RefineGNN105	single CDRH	SAbDab, CoV-AbDabT	$\mathcal{S}$A\D, $\mathcal{X}$A\D	$\mathcal{S}$A, $\mathcal{X}$A	Graph-based method for autoregressive sequence design and iterative structure refinement using MPNNs.	https://github.com/wengong-jin/RefineGNN.
*Li et al.106	full scFV	Pfam, OAS	–	$\mathcal{S}$D	Generates a high-affinity scFV library via an end-to-end Bayesian, LM-based method.	https://github.com/AIforGreatGood/biotransfer.
*NOS107	single or multiple CDRs	custom	random seed for $\mathcal{S}$A\D from paired OAS	$\mathcal{S}$A	Gradient-guided discrete diffusion for CDR(s) sequence design. Generalizes LaMBO108 into a Bayesian optimization-based sequence designer for multiple objectives.	https://github.com/ngruver/NOS.
*AbDiffuser104	full or all CDRs	OAS, HER2, SAbDab	– or ($\mathcal{S}$GA\D, $\mathcal{X}$GA\D)	($\mathcal{S}$D, $\mathcal{X}$D) or ($\mathcal{S}$GA, $\mathcal{X}$GA)	Full-atom diffusion-based method with antibody-specific characteristics as diffusion priors and physics-based constraints. Also supports antigen-conditioned CDR design.	N/A
*Angermueller et al.109	full VHH	OAS, custom	–	$\mathcal{S}$D	A method based on a VAE and a classifier/regressor for designing diverse VHH sequences with mutations focused on CDRs.	N/A
*IgDiff110	full or specific FV region(s)	FrameDiffM, OAS-ABB2T	– or $\mathcal{X}$A\D	$\mathcal{X}$D or $\mathcal{X}$A	Extends FrameDiff111 by handling multiple chains for de novo antibody design.	N/A
*PropEn112	partial (automatic)	custom	$\mathcal{S}$A	$\mathcal{S}$A*	General method for efficient property optimization using an encoder-decoder framework.	N/A
BetterBodies113	CDRH3	custom	–	$\mathcal{S}$D	Combines VAE with reinforcement learning-guided latent diffusion for CDRH3 sequence design.	N/A
Antigen-Conditioned Antibody Design
HERN114	paratope (CDRH3)	SAbDab	$\mathcal{S}$G, $\mathcal{X}$G, E	$\mathcal{S}$GD, $\mathcal{X}$GD	Graph-based method for autoregressive paratope dock-and-design. Contains encoder, docking, and decoder modules.	https://github.com/wengong-jin/abdockgen.
DiffAb91	single or all CDRs	SAbDab	$\mathcal{S}$GA\D, $\mathcal{X}$GA\D	$\mathcal{S}$GA, $\mathcal{X}$GA	Diffusion-based method for antigen-conditioned antibody design. Jointly diffuses residue type, Cα atom coordinates, and orientation joint diffusion.	https://github.com/luost26/diffab.
*Sculptor115	full monobody	MD	$\mathcal{S}$G, $\mathcal{X}$G, E	$\mathcal{S}$GD, $\mathcal{X}$GD	Generates an interaction field around the epitope (input) for guided structure generation using Ig-VAE.63	N/A
MEAN102	single CDRH	SAbDab	$\mathcal{S}$GA\D, $\mathcal{X}$GA\D	$\mathcal{S}$GA, $\mathcal{X}$GA	Graph-based method with progressive full-shot decoding and internal and external context encoder modules.	https://github.com/THUNLP-MT/MEAN.
DockGPT116	all CDRs	BC40P, DIPS, SAbDabT	$\mathcal{S}$A\D, $\mathcal{S}$G, $\mathcal{X}$A\D, $\mathcal{X}$G	$\mathcal{S}$GA, $\mathcal{X}$GA	Docking method based on AF2 modules. Dock-and-design is possible by masking CDR structural features.	https://github.com/MattMcPartlon/protein-docking.
dyMEAN103	CDRH3, all CDRs, or any regions	SAbDab	$\mathcal{S}$A\D, $\mathcal{S}$G, $\mathcal{X}$G, E	$\mathcal{S}$GA, $\mathcal{X}$GA	Extension of MEAN102 for end-to-end full atom dock and design method using an adaptive multi-channel encoder to generate full-atom structures.	https://github.com/THUNLP-MT/dyMEAN.
ABGNN117	single CDRH	OASP, SAbDabT	$\mathcal{S}$GA\D, $\mathcal{X}$GA\D, E	$\mathcal{S}$GA, $\mathcal{X}$GA	Combines a pretrained AbBERT49 LM and HERN114 for one-shot co-design of CDRH.	https://github.com/KyGao/ABGNN.
*AbDesign of AbOpt118	CDRH3	SAbDab	$\mathcal{S}$GA\D, $\mathcal{X}$GA\D	$\mathcal{S}$GA, $\mathcal{X}$GA	Diffusion-based framework using a multi-channel EGNN for CDRH3 co-design.	https://github.com/pengzhangzhi/ab_opt.
HTP119	all CDRHs and/or CDRLs	UniProt, OAS, DIPS, SAbDab	$\mathcal{S}$GA\D, $\mathcal{X}$GA\D	$\mathcal{S}$GA, $\mathcal{X}$GA	Hierarchical training paradigm in four levels leveraging ESM-218 and EGNNs for CDR co-design.	https://github.com/smiles724/HTP.
*MaskedDesign120	CDRH3 or all CDRHs	SAbDab, IgMPNNM	$\mathcal{S}$A\D, $\mathcal{S}$G, $\mathcal{X}$G, E	$\mathcal{S}$GA, $\mathcal{X}$GA	Similar method to DockGPT116 followed by IgMPNN96 to design CDR sequence candidates given the designed complex structure.	https://github.com/AbSciBio/unlocking-de-novo-antibody-design.
ADesigner121	single CDRH	SAbDab	$\mathcal{S}$GA\D, $\mathcal{X}$GA\D	$\mathcal{S}$GA, $\mathcal{X}$GA	One-shot design by decoupling the problem into structure generation and sequence-structure co-learning with cross-gate MLP.	https://github.com/A4Bio/ADesigner.
tFold-Ag from tFold31	masked residues (e.g., CDRs)	ESM-2M, UniRef50P, PDBP, PPIP, OASP, AF2M, SAbDab	$\mathcal{S}$A\D, $\mathcal{S}$G, E or E+P	$\mathcal{S}$GA, $\mathcal{X}$GA,confidence	Structure prediction method with a sequence recovery module to allow co-designing of masked input regions. Uses ESM-PPI, developed by further pretraining ESM-2.18	https://github.com/TencentAI4S/tfold.
*RFdiffusion Antibody32	all CDRs in VHH or scFV	RFdiffusionM, UCST	$\mathcal{S}$A\D, $\mathcal{S}$G, $\mathcal{X}$A\D, $\mathcal{X}$G, E	$\mathcal{X}$GA	RFdiffusion14 fine-tuned on native antigen-antibody complexes to design antigen-conditioned de novo variable heavy chains and scFVs.	N/A
AbDPO122	CDRH3	SAbDab, syntheticT	$\mathcal{S}$GA\D, $\mathcal{X}$GA\D	$\mathcal{S}$GA, $\mathcal{X}$GA	Extends DiffAb84 with direct energy-based preference optimization.	N/A
ASSD123	single CDRH	ESM-2M, SAbDabT	$\mathcal{S}$GA\D, $\mathcal{X}$GA\D	$\mathcal{S}$GA, $\mathcal{X}$GA	Decouples the co-design method into sequence design by fine-tuning ESM-218 followed by structure prediction with MEAN.95	https://github.com/lkny123/ASSD_public.
DiffForce124	single CDR	SAbDab	$\mathcal{S}$GA\D, $\mathcal{X}$GA\D	$\mathcal{S}$GA, $\mathcal{X}$GA	Extends DiffAb84 by using molecular dynamics-based force-guided sampling.	N/A
AbX125	all CDRs	ESM-2M, SAbDab	$\mathcal{S}$GA\D, $\mathcal{X}$GA\D	$\mathcal{S}$GA, $\mathcal{X}$GA	Score-based diffusion model guided by evolutionary, physical, and geometric constraints. Full-atom version is also available.	https://github.com/zhanghaicang/carbonmatrix_public.
GeoAB126	single CDR	SAbDab	$\mathcal{S}$GA\D, $\mathcal{X}$GA\D	$\mathcal{S}$GA, $\mathcal{X}$GA	Graph-based method with heterogeneous residue-level encoder for design and affinity maturation.	https://github.com/EDAPINENUT/GeoAB.
IgGM127	all CDRs	ESM-PPIM, SAbDab	$\mathcal{S}$A\D, $\mathcal{S}$G, $\mathcal{X}$G, E	$\mathcal{S}$GA, $\mathcal{X}$GA	Multi-level network architecture with a pretrained ESM-PPI31 and diffusion and consistency models.	https://github.com/TencentAI4S/IgGM.

Pretraining and fine-tuning datasets are suffixed by superscripted P and T, respectively. Pretrained models (PTM) are suffixed by superscripted M. Dataset Abbreviations: OAS structures predicted by ABodyBuilder2 (OAS-ABB2); customized molecular dynamics simulations (MD); Database of Interacting Protein Structures (DIPS); protein–protein interaction dataset from HINT, IntAct, HIPPIE, PrePPI, BioGRID, ComPPI, and hu.MAP (PPI); unspecified antigen-antibody complex structures (UCS).

The following conventions are applied for both input and output columns. Abbreviations: antibody (A), antigen (G), desired design range (D), paratope (P), epitope (E), not specified (–). Notations: Let A, B, and C be three entities, where C ⊂ A, then $\mathcal{S}$_A is the sequence of A, $\mathcal{X}$_A is the structure of A, A* is an optimized A, AB is a complex of A and B, A\C is entity A excluding subentity C.

Case 1. Unconditioned antibody design

Given a set of antibodies, unconditioned methods learn to generate antibodies that match the training data. Despite not being explicitly antigen-conditioned, the training data can be specific to an antigen, e.g., only containing binders for a specific target. In this case, the design problem can be perceived similarly to antigen-conditioned antibody design without explicit antigen information. For these methods, the in silico evaluation is usually performed based on how well the generated data matches the training and validation data. For in vitro evaluation, the evaluation metrics consist of the expression rate of the generated antibodies and, for implicitly antigen-specific designs, their binding affinity to the target antigen.

Sequence-based methods

Given the recent success of LMs in both representation learning and generative purposes, the methods presented here and those previously mentioned in the Antibody Representation Learning section above overlap. LMs learn interdependencies of antibody sequences and can generate antibody-like sequences by sampling in an autoregressive or random order manner. In IgLM,⁵⁵ the antibody sequence is designed autoregressively after giving conditions on the species and the corresponding chain, i.e., heavy or light. For AntiBERTa,⁵² AbLang,⁵³ and ReprogBert,⁶⁵ the network is trained as a masked LM, and by masking desired positions, amino acid mutations can be sampled from their predictions. As another paradigm for discrete generative modeling, dWJS⁶⁴ presents a discrete walk-jump sampling algorithm, in which new antibodies are proposed directly by walking in a noisy amino acid distribution. A separate denoising neural network is trained in which, from this noisy distribution, a denoised amino acid sequence can be decoded.⁶⁴ On the other hand, Li et al.¹⁰⁶ propose an end-to-end Bayesian, LM-based method to design libraries of high-affinity single-chain variable fragments (scF_Vs), which were experimentally validated through binding assays, demonstrating that ML-generated scF_V libraries are highly diverse and outperform conventional directed evolution.

In NOS,¹⁰⁰ a method for gradient-guided sampling using discrete diffusion models is proposed. The method is tested for infilling CDRs and optimizing a value function when combined with Bayesian optimization and shows effectiveness in generating binders in wet lab experiments.¹⁰⁷ Recently, Angermueller et al.¹⁰⁹ proposed a method based on a variational autoencoder (VAE) and a classifier/regressor to design highly diverse single-domain antibodies (antibody variable heavy chain, V_HH) with high binding affinity to SARS-CoV-1 and SARS-CoV-2. The method trains the model in three rounds with a high-throughput experimental platform to measure binding affinity; the first trains the VAE, while the subsequent ones jointly train the classifier and regressor. Next, PropEn¹¹² creates a dataset with matching antibodies, i.e., each antibody sample is matched with a mutant with better properties and trains an encoder-decoder architecture that acts as a generative framework for property optimization. In this case, given an antibody sequence as input, PropEn generates a mutant candidate optimized for the property of interest. Finally, BetterBodies¹¹³ combines VAEs with offline reinforcement learning-guided diffusion in latent space to generate CDRH3 sequences with improved affinity to the SARS-CoV spike receptor-binding domain.

Structure-based and co-design methods

Recent methods also investigated unconditioned antibody design by generating antibody structures. Typically, given a dataset of antibody structures, these methods learn how to create similar antibody-like structures. For example, the latent space learned by Ig-VAE⁶³ can be used for generation purposes. In this case, given the feature vector obtained by the Ig-VAE encoder for a target antibody, similar antibodies can be generated by adding noise, followed by leveraging the Ig-VAE decoder for structure reconstruction. Ig-VAE has been used as a base for an antigen-conditioned generator,¹¹⁵ which is discussed below. RefineGNN,¹⁰⁵ a graph-based method, co-designs the sequence and structure of antibody CDRs given its FR characteristics. The iterative generation process is performed by two message-passing neural networks (MPNNs), each predicting the sequence and structure, respectively.¹⁰⁵ A diffusion-based unconditioned co-design method named AbDiffuser has also been proposed.¹⁰⁴ The method utilizes the AHo numbering system¹²⁸ to handle the variable length of antibodies and proposes a frame-averaging method to obtain the canonical pose of an antibody. Furthermore, it exhibits the ability to learn the distribution of a paired OAS database and generate novel antibodies when trained on a dataset of HER2 antibodies obtained by mutagenesis.¹⁰⁴ More recent methods include IgDiff,¹¹⁰ which fine-tunes FrameDiff¹¹¹ on synthetic antibodies to generate antibody-like structures.

Case 2. Antigen-conditioned antibody design

The ultimate objective in antibody design is to target a specific antigen and its epitope. Given this objective, explicitly incorporating antigen information in the design pipeline is desirable. However, this has been challenging for structure-based and co-design methods due to the low number of experimental antigen–antibody complex structures, currently amounting to approximately 9,600 structures in SAbDab.¹²⁹ Furthermore, given redundancies in sequence, structure, and antigen types, the effective number of distinct experimental examples is reduced further after preprocessing steps like clustering by sequence similarity. Traditionally, methods like Rosetta Antibody Design (RAbD)¹³⁰ used this database of known structures to sample CDR loops and graft them into the initial antigen–antibody complex structure. The proposed designs were then optimized and selected based on binding energy values using the Metropolis-Hastings algorithm. Recently, various AI-based antigen-conditioned antibody design methods have been proposed following the recent advances in AI architectures and data augmentation techniques.^{31,32,91,96,102,103,114–127}

While antibody design methods have initially focused on generating only the antibody CDRs,^84,95,96 recent methods have also proposed generating the entire F_V.^100,106 Interestingly, most of the proposed CDR generation methods^92,102,103 achieve this goal by masking the input; hence, theoretically, they can also be used to generate the entire F_V when no masking is applied. Given the more flexible design range of the more recent antibody design methods, they can be classified into partial and full antibody design methods. In partial antibody design, it is assumed that the antigen-binding antibody candidate(s) are known, with or without structural information, whose properties, such as binding affinity and stability, are to be enhanced by designing only a specific region(s). On the other hand, in full antibody design, it is assumed that there is no information on an antibody candidate, and the main objective is to generate a completely novel binder, i.e., the entire antibody F_V, with desirable properties.

Sequence-based methods

As described in Case 1 above, unconditioned methods can be used for a specific target when trained with data specific to an antigen. However, it is possible to explicitly condition protein language models (PLMs) for specific targets, as discussed by Ferruz et al.⁷⁷ In this case, the condition would be the antigen information, which must be chosen from the ones in the PLM training dataset. For example, the antigen sequence could be given as a fixed sequence input for the PLM, with only the antibody sequence region being masked for generation. Despite this possibility, the applicability of sequence-based generative methods, such as ESM-2¹⁸ and EvoDiff,¹³¹ for target-conditioned sequence-based design is incipient and needs additional research, especially for antigen-conditioned antibody design.

Structure-based and co-design methods

AI-based antigen-conditioned structure-based and co-design methods have been proposed recently.^{31,32,91,96,102,103,114–126} These works use different methodologies to condition the generation of the antibody given the antigen information and the design range. The methods leveraging GNN-based architectures, diffusion-based generative frameworks, and other approaches for antigen-conditioned design are detailed in the subsections below. This organization aims to accommodate the prevalent approaches for structure-based and co-design methods; it does not imply a mutual exclusivity between GNN-based and diffusion-based methods, e.g., a method may use a GNN architecture within a diffusion framework.

GNN-Based methods

In antigen-conditioned antibody design methods using GNN architectures,^{102,103,114,117,119,123,126} the antigen–antibody complex is represented as a graph with nodes representing residues or atoms and edges connecting surrounding nodes in sequence and structure levels. A GNN is then trained on this graph representation to learn features for sequence and structure generation, the former only for co-design methods. An illustration of such a process following the methodology of MEAN¹⁰² is shown in Figure 4a.

Figure 4.

Common approaches for antigen-conditioned antibody sequence-and-structure co-design. a, a GNN-based method following the methodology in MEAN¹⁰² where the antigen-antibody complex is represented by a graph with three subgraphs representing each chain, i.e., antibody heavy chain V_H, antibody light chain V_L, and antigen. Each residue in the chains is represented by a node whose embeddings contain its position and properties (e.g., amino acid type), while each inter-residue interaction is represented by an edge. Learning is performed by internal and external context encoders via message passing within and between the subgraphs, respectively, to update the node embeddings. After the final layer, the protein sequence of the chains is predicted from the node embeddings. b, a diffusion-based method. Here, we assume that only the interaction regions of the antibody are designed, while others are kept fixed. During the forward diffusion process, diffusion-based methods gradually add noise to a representation of the antibody sequence/structure (e.g., in DiffAb,⁹¹ joint diffusion of residue types, Cα atom coordinates, and residue orientations) from timestep 0 to timestep T following a Markov process based on q(x_t|x_t−1), where x_t is the representation at timestep t. During the backward diffusion process, a model with parameters θ is trained to recover the original representation by gradually denoising p_θ(x_t−1|x_t) from timestep T to timestep 0. Note that separate neural networks can be used for the backward diffusion process of each representation in the joint diffusion.

HERN¹¹⁴ proposes a method for de novo paratope dock-and-design, in which the CDRH3 is assumed to be the paratope. The method iteratively generates and refines the paratope sequence and structure given a target antigen/epitope. The epitope-paratope complex is represented as a hierarchical graph consisting of a residue-level graph with only Cα atoms (coarse-grained level) and an atom-level graph that includes the side chains (fine-grained level). The learned representations are used to model pairwise forces for predicting updated atom coordinates at each iteration, where the residue-level representation is used to generate the paratope sequence autoregressively. The method is evaluated for in silico metrics in CDRH3 generation with no framework information.

Following that, MEAN¹⁰² is a method for the co-design of antibody CDRs by formulating the generation as a conditional graph translation problem. The sequence and structure are generated via a multi-round progressive full-shot inference process, providing a higher computational efficiency when compared to autoregressive approaches. It assumes that the structure of the full antigen (or only the epitope) docked to an incomplete antibody is given. The task is then to design a masked region of the antibody, e.g., CDRH3. The MEAN architecture contains GNN layers that alternate between two modules: an internal context encoder and an external context encoder. Feature embedding vectors and atom coordinates are updated via message passing in these encoders. After the final layer, the network outputs coordinate predictions, followed by amino acid type predictions from the learned residue embeddings. MEAN is tested for in silico metrics, such as amino acid recovery (AAR) and root mean square deviation (RMSD), with respect to ground truth structures. Despite being tested only for the design of single CDRHs, extending MEAN for other regions or full antibody design is straightforward.

As an extension of MEAN,¹⁰² dynamic MEAN (dyMEAN)¹⁰³ is proposed for end-to-end full-atom antigen-conditioned antibody design. It assumes that the antigen/epitope is given and that only an incomplete antibody sequence is available, ultimately omitting the need for an antibody template structure as an input. This way, dyMEAN incorporates antibody structural initialization, epitope-antibody docking, and an adaptive multi-channel equivariant encoder for full-atom design in one framework. dyMEAN is tested in silico for recovery metrics, docking scores, i.e., DockQ,¹³² and binding free energy scores. Compared to previous graph methods, its ablation studies highlight the importance of iterative full-atom generation.

Next, HTP¹¹⁹ has been proposed to co-design the antibody CDRs by leveraging a transformer-based PLM, ESM-2,¹⁸ and equivariant GNNs (EGNNs) with a four-level hierarchical training paradigm. The first two training stages concern the PLM training and fine-tuning, while the third trains the EGNNs, and the last employs the PLM and EGNNs for CDR co-design with fixed PLM parameters. The method is evaluated in silico for AAR and RMSD to the ground truth antibodies. ASSD,¹²³ on the other hand, decouples the antigen-conditioned CDR design task into sequence design with a pretrained ESM-2¹⁸ and structure prediction steps with an architecture similar to MEAN.¹⁰² Recently, GeoAB¹²⁶ has been proposed to co-design CDRs with a focus on generating realistic geometries in two steps. The first step leverages a generative model based on NeRF¹³³ conditioned on antigen information to generate template structures. In contrast, the second step employs a GNN using a heterogeneous residue-level encoder and equivariant atom-level layers for sequence and structure refinement. GeoAB is tested with in silico metrics and is shown to generate more realistic structures when compared to existing methods.

Diffusion-based methods

Given the high success achieved by diffusion models for image generation^134,135 and their characteristics of generating designs with high diversity, various diffusion models have been proposed for unconditioned and conditioned protein structure design.^{14,15,111,136–142} For these methods, suitable representations and geometric properties are crucial, as exemplified in FrameDiff,¹¹¹ a methodological framework to generate protein structures using the frame representation from AF2.⁹ Recently, diffusion models that leverage the inductive bias learned by structure prediction networks have been proposed and achieved high experimental success. For instance, RFdiffusion,¹⁴ which focuses only on protein structure generation, fine-tunes a pretrained RF2¹⁹ model for the denoising diffusion process, interpolating noise with the structure prediction given the current structural state as a condition. For the joint generation of sequence and structure, i.e., co-design, diffusion models capable of handling discrete spaces, such as D3PMs,¹⁴³ should be incorporated into the method as attempted by DiffAb.⁹¹ Many recent works^{91,118,122,124,125} attempt to address diffusion framework-based antigen-conditioned antibody design methods, as illustrated in Figure 4b.

DiffAb⁸⁴ is a diffusion-based antigen-conditioned antibody co-design method for single or multiple CDRs. It assumes that the structure of an antigen docked to an incomplete antibody and their respective sequences are given, where the missing parts of the antibody are to be designed. The method represents a residue by its amino acid type in a one-hot representation, Cα coordinates, and residue orientation. It jointly diffuses these elements, where noise is gradually added based on a multinomial distribution for sequence, a Gaussian distribution for the coordinates, and an isotropic Gaussian distribution for the orientations, respectively, during the forward diffusion process. The method leverages a neural network architecture equivariant to rotation and translation for co-design; a multilayer perceptron (MLP) is used to calculate single and pairwise residue embeddings to be fed to invariant point attention (IPA) layers as proposed in AF2⁹ before performing the backward diffusion process on the three representations with three respective MLP layers. As the structural representation used by DiffAb only considers backbone atom generation, the full-atom structure is obtained by Rosetta side chain repacking. DiffAb is evaluated using in silico recovery metrics for antibody design and optimization.

AbDPO¹²² extends the architecture presented in DiffAb⁹¹ with direct preference optimization (DPO)¹⁴⁴ for generating antibodies with lower energy. It first trains the method on antigen–antibody complexes to generate synthetic antibodies that are further used for fine-tuning the model in another step for residue-level energy-based optimization with DPO. It is shown that the model can achieve lower energies after fine-tuning. DiffForce¹²⁴ proposes another extension of DiffAb.⁹¹ The method guides the sampling during the diffusion process using a differentiable force field, which acts similarly to potentials introduced in RFdiffusion¹⁴ and Chroma.¹⁵ A score-based diffusion model that incorporates evolutionary, physical, and geometric constraints is introduced in AbX.¹²⁵ A pretrained LM is used as a prior for sequence design, and two additional loss terms, based on a van der Waals loss and a frame-aligned point error (FAPE)⁹ loss, are used to train the diffusion model using continuous timesteps.

Finally, RFdiffusion Antibody³² is a structure-based model developed by fine-tuning RFdiffusion¹⁴ predominantly on antibody complex structures. Like RFdiffusion, during training, an antigen–antibody complex and a random timestep are sampled, followed by corrupting the antibody structure by gradually adding noise based on the sampled timestep while keeping the antibody framework and antigen structure fixed. The framework structure is given in a global frame-invariant manner to allow sampling on different docking positions. Given a framework template and a target antigen with the desired epitope region (i.e., hotspot residues) specified, the model designs multiple samples of the backbone structure of the CDRs in V_HH with various docking positions on the target epitope. The generated samples are then passed to ProteinMPNN¹³ to design the CDR loop sequences, then further validated and filtered with Rosetta’s change of Gibbs free energy (∆∆G) metrics and RF2 antibody, a fine-tuned version of RF2¹⁹ for antibody structure prediction. Compared to RFdiffusion, the method enables a more accurate framework recapitulation and efficient hotspot targeting, thus leading to favorable interactions with the target antigen at the specified epitope region. Additionally, the method incorporated a methodology for the design of scF_Vs by finding suitable framework templates for a target antigen. The work investigated the biochemical characterization and structure determination of V_HH and scF_V designs, demonstrating their ability to design novel CDR structures that specifically and accurately bind to the desired target epitope at atomic-level precision.

Other methods

In Sculptor,¹¹⁵ an epitope-specific generative algorithm is proposed to design nanobodies. An interaction field is created around the specified epitope of the target antigen to guide the generation of both the nanobody structure, leveraging the encoder-decoder from Ig-VAE,⁶³ and the docking position with the antigen structure.¹¹⁵ The method was validated for the design of venom toxin binders, where one binder was confirmed experimentally.¹¹⁵ DockGPT¹¹⁶ is proposed for antigen–antibody docking by using an architecture based on AF2 modules.⁹ Given unbound sequences and coordinates of the antigen and antibody, and optionally the inter-chain contacts and binding interface, it predicts the complex structure and confidence metrics similarly to structure prediction networks. Despite being proposed for antigen–antibody docking, DockGPT is adapted for CDR loop design by masking the CDR structural features.¹¹⁶ In a work by Shanehsazzadeh et al.,¹²⁰ given the sequence and structure of the antigen, the epitope region, the sequence of the antibody FRs, and the length of the CDRs in V_H to be designed, a structure design method similar to DockGPT,¹¹⁶ named MaskedDesign, is used to output an antigen–antibody complex structure candidate, followed by the design of CDR sequence candidates using IgMPNN.⁹⁶ In ADesigner,¹²¹ a neural network is proposed for the geometric modeling of structures combined with a layer that co-designs sequence and structure in a refinement step. For tFold-Ag,³¹ a sequence recovery head is added to an antibody-specific structure prediction network. The co-design of antibody CDRs is performed by fine-tuning the model for design and giving masked inputs.

Discussion

Numerous advances have been achieved in antibody design with AI-based methods. Some of these include powerful representation learners, such as ALMs^51–53,55 that can also be used for antibody generation, methods able to generate the structure of the entire variable domain F_V⁹¹, and antigen-conditioned antibody design methods capable of sequence-structure co-design.^{32,92,102,103,114} For some of these methods, in vitro experiments have verified their ability to effectively design novel antibodies for different target antigens.^{32,78,79,104,115} Altogether, these outcomes present promising research directions to explore in the next few years in AI-based antigen-conditioned antibody design. Despite the recent advances, various limitations of AI-based methods need to be addressed for their effective incorporation in antibody development pipelines for therapeutic and diagnostic applications.

Docking performance and novel AI architectures

Assuming an increased performance by structure prediction networks to predict loop regions, the development of antigen-conditioned antibody design methods will be determined by the capacity to predict the docking position efficiently. Various methods incorporate docking as a part of the design methodology,^103,114,116 while others consider docking a separate problem.¹¹⁸ Regardless, the docking performance is crucial for effective design when the antigen and antibody are given in their unbound conformations. Alternative methods that tackle the docking problem as a search over surface characteristics have been proposed and achieved experimental success when applied to protein–protein interactions^145,146; these methods can readily be applied to antibodies. Other means for docking evaluation are based on molecular dynamics simulations.¹⁴⁷

Recent works on antigen-conditioned antibody design have mainly focused on applying GNNs and diffusion-based generative algorithms. Consequently, novel methods considering the characteristics of antibodies should be investigated. For GNN-based methods, message passing to learn node features and interaction characteristics between long-distance amino acids with no direct edges remains difficult.^148,149 This issue limits the usability of GNN-based methods for designing full antibody heavy and light chains, although this effect is alleviated on partial design tasks. For GNNs, the deterministic nature of their sampling may also limit the diversity of the generated designs.

For diffusion-based methods, limitations include the high number of sampling steps, the dependency on hand-designed noise schedules, and the discrete time modeling. These issues make precise generation tasks, such as the design of CDR loop structures and exact matches of sequence and structure, rather challenging. Especially, diffusion-based methods often struggle to obtain a good match between sequence and structure when generating both simultaneously in a co-design manner. To this end, alternatives like flow matching are actively being investigated to achieve higher designability, faster sampling, and conditioned generation and to address the properties needed for 3D structure generation.^150–153

Multi-objective data beyond sequence and structure

Another challenging aspect of antibody design is data scarcity. Effective data augmentation techniques are needed to improve the training of these methods. For example, unconditioned methods like AbDiffuser¹⁰⁴ utilize synthetic data from an antibody-specific structure prediction network²⁷ to train their diffusion process. Methodologies to create synthetic antigen–antibody complexes have been investigated,⁴⁰ but this remains an open challenge. Additionally, negative examples from failed experiments are meaningful when training AI-based methods. For example, in the RF2¹⁹ training, negative PPI pairs are generated to train confidence metrics and improve the robustness of the model.

Efforts from researchers to save and organize this negative data can improve the validation tools for generated designs. For instance, in AbDiffuser,¹⁰⁴ a dataset of known HER2-binding and non-binding mutants of trastuzumab was used to train a classifier to distinguish binders from non-binders. Advances in this direction include the release of the Protein–Protein Negative Interaction Dataset.¹⁵⁴ Moving forward, we envision a community-driven platform for biologists to report confirmed negative examples from their experiments, along with relevant remarks, e.g., low expression/stability measures, non-binding examples, dissimilar docking positions between design and results, which can be used to improve the robustness and generalization of AI-based models.

For most of the generative methods presented, the design pipeline generates many candidates, which are further filtered based on desired properties, e.g., naturalness, stability, expressibility, viscosity, immunogenicity, pharmacokinetics, potency, selectivity, cross-species reactivity, and binding affinity. The best candidates should optimize these multiple properties in a multi-objective optimization scenario. Several AI-based methods have been developed for the docking pose ranking and binding affinity/energy prediction of antigen–antibody complexes, such as DLAB,¹⁵⁵ CSM-AB,¹⁵⁶ the method proposed in Yang et al.,¹⁵⁷ and DSMBind.¹⁵⁸ Moreover, methods like TAP¹⁵⁹ and PropertyDAG¹⁶⁰ have been developed to predict CDR-related developability properties and rank design candidates, respectively. Prediction methods for antibody-related developability properties are further discussed in existing literature,^82,161,162 although the development of these prediction methods is largely hindered by limited data availability and the inconsistency of experimental protocols and measurements used to define ground-truth values for such metrics.

Recently, lab-in-the-loop approaches have been investigated to incorporate multiple wet lab data directly as feedback in the antibody design framework. These approaches are usually sequence-based and applied to improve the fitness of the candidates obtained by traditional methods. In Frey et al.,¹⁶³ a sequence-based antibody design method, dWJS,⁶⁴ is incorporated in a semi-autonomous, iterative optimization loop to design antibodies for four different antigens, resulting in better binding candidates. The framework consists of antibody design, property prediction, ranking, filtering for running experimental assays, and using the obtained in vitro data for subsequent optimization rounds. The method for sequence design proposed by Shanker et al.⁷⁹ also incorporates wet lab results into its iterative design framework, implementing an evolutionary-like optimization approach in a lab-in-the-loop framework. To achieve promising de novo antibody designs with desirable potency, developability, and manufacturability for rapid advancement toward clinical applications, lab-in-the-loop frameworks should be integrated with multi-objective optimization strategies.

In silico filtering and high-scale automated experimental evaluations

Various in silico evaluation metrics have been applied to filter candidates in AI-based protein antibody design. For de novo mini binder design, recent progress has been achieved by leveraging confidence metrics, such as pLDDT and pAE, from structure prediction networks, as these metrics have demonstrated strong correlations with experimental success rates.⁹¹ For antibodies, similar evaluation metrics, like the i-pTM of AF3,¹² have been shown to correlate with the experimental success of de novo designed scF_Vs³¹. Different in silico filtering strategies are performed in experimentally validated methods. RFdiffusion Antibody³² applies the confidence metrics from a fine-tuned version of RF2,¹⁹ while AbDiffuser¹⁰⁴ trains an ML-based classifier from a dataset of binders and non-binders for the target antigen. Both proseLM⁷⁸ and Shanker et al.⁷⁹ use the likelihoods of their structure-informed LMs to filter generated candidates for experiments. Other widely used in silico evaluation metrics include ∆∆G, AAR, and RMSD relative to ground truth structures. However, it is difficult to reliably assess design success based on these metrics alone, as even minor structural changes, like a single mutation, can substantially alter binding affinity.

The success rate of different methods also varies based on their assumptions and experimental evaluation methodologies. Higher success rates are observed in methods that assume an initial antigen–antibody complex structure for optimization.^78,79 For instance, proseLM⁷⁸ achieves success rates of 25.4% and 92.5% when the model proposes mutations to the CDR and the FR of the antibody, respectively. The success rate, however, falls to approximately 2% when the problem setting is more challenging: proposing mutations for the entire F_V given the complex structures. Similar results were found in MaskedDesign,¹²⁰ where over 1 million variants were screened with success rates of 10.6% and 1.8% when designing only the CDRH3 and all CDRHs together, respectively. On the other hand, in AbDiffuser,¹⁰⁴ a dataset of thousands of known binders is assumed, and only 16 candidates are chosen for validation. In this setting, all candidates were expressed successfully, and 57.1% were tight binders. Recently, for the complex problem of antibody design with no known initial binder, Bennett et al.³² mention that low success rates were obtained after performing high-throughput (9,000 designs) and lower-throughput (96 designs) campaigns for different antigens. Overall, the lack of consistency in experimental evaluation methodologies across studies poses additional challenges for the establishment of robust filtering criteria, leading to a pressing need to establish consensus on evaluation metrics and develop standardized benchmark datasets for reliable and effective candidate filtering.

Finally, AI-based computational methods will be able to generate a high number of candidates to be evaluated in wet lab experiments in the near future. To validate these candidates, wet lab experiments should keep pace by being able to test more proteins simultaneously in lab-in-the-loop scenarios. For this, research in high-throughput automation in robotics for drug discovery is essential. Collaborations among interdisciplinary teams spanning AI, robotics, computer science, and life sciences will be essential for the effective experimental validation of the expanding repertoire of AI-based antibody design methods.

Conclusion

Here, we have presented the current status of antigen-conditioned antibody design research using artificial intelligence, covering general protein and antibody-specific methods for structure prediction, representation learning, and sequence design, noting a rise in antibody-specific approaches. These advancements suggest a promising future for generating de novo antibodies for various targets, potentially transforming the development of new therapeutics. AI has the potential to expedite drug development and continued research is crucial for achieving a faster and more cost-effective drug discovery process for the common good.

Funding Statement

This work was supported by grants from Institute for Basic Science (IBS) [IBS-R029-C2] and [IBS-R030-C1], the National Research Foundation of Korea [RS-2024-00397681] to H.M.K., and the KAIST Convergence Research Institute Operation Program (to H.M.K.).

Disclosure statement

No potential conflict of interest was reported by the author(s).

Additional information

Correspondence and requests for materials should be addressed to Meeyoung Cha or Ho Min Kim.