Beyond affinity: AI-supported developability assessment and multi-objective optimization in antibody development

Abstract

Therapeutic antibodies are a major class of medicines, but high target affinity alone does not ensure manufacturability, stability, safety, or clinical success. Developability has therefore become a central constraint in antibody engineering and has pushed the field beyond affinity-first screening toward multi-objective decision-making. Machine-learning-based models increasingly integrate antibody sequence, structure, interaction, and assay data to estimate properties such as affinity, specificity, aggregation, viscosity, solubility, stability, immunogenicity, and pharmacokinetics before experimental testing. In practice, these models are most useful when they help prioritize experiments rather than replace empirical evaluation. Here, we review the data resources used for antibody developability modeling, the main classes of property predictors, and optimization frameworks that balance competing design objectives. We cover Pareto optimization, Bayesian optimization, active learning, and conditional generative modeling, and discuss how these approaches are being adapted to bispecific antibodies and nanobodies. We argue that AI is most useful when model outputs are interpreted in the context of assay design, uncertainty, and antibody format, and when they are used to guide candidate selection and experimental design rather than serve as stand-alone surrogates for developability.

Statement of Significance This review summarizes how AI is moving antibody developability assessment beyond affinity-first screening toward multi-objective design. We highlight data resources, property predictors, and optimization strategies that help prioritize balanced candidates while accounting for assay context, uncertainty, and antibody format.

Introduction

Therapeutic antibodies are a major class of biotherapeutics and address major unmet needs in oncology, autoimmune diseases, and infectious diseases [1, 2]. More than 150 antibody-based therapeutics have received Food and Drug Administration (FDA) approval, and antibodies continue to account for an important share of recent drug approvals [3, 4]. Despite this progress, their discovery and development pipeline remains lengthy, costly, and inefficient, often spanning more than a decade and requiring hundreds of millions in investment per approved drug [5, 6].

A primary bottleneck lies in developability—the suite of physicochemical and biophysical properties that determine whether a candidate can advance to a viable, manufacturable, and safe therapeutic [7, 8]. High-affinity binders might fail in later stages due to issues such as aggregation, high viscosity that complicates subcutaneous delivery, poor solubility, thermal or chemical instability, immunogenicity, or unfavorable pharmacokinetics [9, 10]. These liabilities lead to substantial attrition, wasting resources when failures occur in expensive clinical trials or during scale-up. Historically, antibody engineering prioritized maximizing binding affinity and specificity to the target antigen, often through targeted mutagenesis in complementarity-determining regions (CDRs) [11, 12]. However, this single-objective focus frequently introduces trade-offs, such as hydrophobic patches that promote aggregation or charge alterations that compromise stability or introduce off-target binding [13, 14].

The multi-objective optimization challenge lies at the heart of modern antibody engineering: properties are highly interdependent in a vast, nonlinear, high-dimensional sequence–structure space. Optimizing one attribute often degrades others, resulting in complex Pareto fronts where the goal is to identify sets of nondominated candidates balancing multiple objectives rather than a single best molecule [15, 16].

Recent advances in artificial intelligence (AI) and machine learning have introduced complementary computational tools for addressing these challenges. Protein language models [17–19], structure prediction tools [20, 21], and generative architectures [22–24] are increasingly used to estimate antibody properties from sequence or structure [25]. These methods have also benefited from larger datasets that combine antibody sequences, experimental structures, interaction records, and developability measurements [26]. Their practical value lies in helping teams prioritize experimental testing, flag likely developability risks earlier, and guide optimization when assay data are limited.

Despite this progress, antibody developability modeling remains fragmented. Many studies still examine isolated endpoints, such as affinity prediction or aggregation propensity alone [27, 28], often using inconsistent datasets and with limited connection to downstream decision-making. Prior reviews have summarized important components of this field, including antibody structure modeling, property prediction, AI-assisted antibody discovery, high-throughput experimentation, and deep sequencing–driven synthetic antibody design [5, 29, 30]. Fewer reviews connect these components into an end-to-end design workflow: which datasets support which property models; how single-property predictors should be combined under uncertainty; and how model outputs can guide Pareto optimization, Bayesian optimization, active learning, or generative search in realistic development settings. This review therefore treats data resources, property prediction, and multi-objective optimization as linked parts of the same antibody design problem.

This review is organized around three questions (Fig. 1). First, what data are available for training and evaluating antibody developability models? Second, which properties can currently be predicted from sequence, structure, or learned representations, including affinity, specificity, aggregation, viscosity, solubility, stability, immunogenicity, and pharmacokinetics? Third, how are optimization methods being used to balance affinity with manufacturability, safety, and formulation constraints? Throughout, we focus less on model novelty alone than on what level of evidence is sufficient to support synthesis, testing, reformulation, or elimination of a candidate. Because many recently proposed approaches remain at an early stage of validation, we emphasize practical evidence supporting their use in antibody developability assessment rather than algorithmic novelty alone. We also discuss how these questions change for nanobodies and bispecific antibodies, where architecture-specific liabilities can make standard IgG-based heuristics insufficient.

Figure 1.

From affinity-first screening to AI-supported multi-objective antibody design. Left: the conventional pipeline progresses sequentially from antibody library construction to affinity screening, hit selection, and only then developability assessment, such that liabilities are often discovered late and require re-engineering. Right: a hierarchical AI workflow in which large-scale sequence, structure, and developability assay data form the data backbone; multimodal representation learning converts these inputs into sequence, embedding, and graph-based features; and multi-objective prediction jointly estimates affinity, specificity, stability, viscosity, solubility, and pharmacokinetics. These predictions are then coupled with Pareto optimization, Bayesian optimization, active learning, and generative models in a closed-loop design-build-test-learn cycle that enables proactive selection of balanced antibody candidates.

The data backbone of antibody developability

The usefulness of a computational developability model depends on the data used to train and evaluate it. Current resources fall into three broad groups: assay-labeled developability datasets, antibody–antigen interaction datasets, and large sequence or structure repositories used for pretraining, annotation, and structure-aware modeling. These groups support different parts of the design workflow. Assay-labeled datasets provide supervised endpoints, interaction datasets support binding and specificity modeling, and sequence or structure repositories provide the background distribution needed for representation learning and format-specific analysis. Table 1 summarizes representative resources across these categories. In practice, these resources are not interchangeable. Some are most useful for benchmarking under controlled caveats, some are better suited for pretraining or representation learning, and some require careful handling before they can support supervised developability prediction.

Table 1.

Summary of key datasets for antibody developability and design.

Category	Dataset	Scale/size	Key features & scope	Ref.
Developability	Jain et al.	137 Abs	Widely used benchmark with 12 biophysical measurements (e.g. titer, purity, aggregation); limited by size and survivor bias.	[31]
	FLAb	13 384 measurements	Fitness landscapes from 17 mutational scans; covers six properties including immunogenicity and binding.	[32]
	FLAb2	>4 million seqs	Expands FLAb to 32 studies; annotated with seven fitness dimensions including pharmacokinetics.	[33]
	PROPHET-Ab	246 Abs	Standardized assay panel with 10 developability measurements generated through automated workflows.	[34]
	DOTAD	Aggregated	Dedicated developability repository integrating experimental and predicted attributes with clinical metadata.	[35]
Interaction	ASD	>1 million records	865 k unique Abs paired with 9.5 k antigens; includes affinity metrics (KD, IC) and structural identifiers.	[36]
Sequence	OAS	> 1 billion seqs	Large collection of unpaired and paired variable-region sequences from over 80 studies; commonly used for PLM pretraining.	[37]
Structure	AbDb	PDB-derived structures	Antibody structure database with standardized sequence, structure, and numbering annotations.	[38]
	SAbDab	12 367 Fv structures	Curated antibody structures from PDB; standardized numbering and annotation.	[39]
	Thera-SAbDab	Therapeutic set	Tracks unique therapeutic antibodies with clinical trial status and target specificity.	[40]
	CoV-AbDab	Specialized	Focused repository for anti-coronavirus antibodies.	[41]
	sdAb-DB	Specialized	Database dedicated to single-domain antibodies (nanobodies).	[42]

Assay-labeled datasets have moved from small, carefully measured panels toward larger collections that better support deep learning. The Jain et al. dataset remains a widely used benchmark because it reports 12 biophysical measurements for 137 clinical-stage antibodies, including expression titer, purity, thermal stability, aggregation, hydrophobicity, self-association, and polyreactivity [31]. Its main limitation is scale: most entries are successful clinical candidates, so the dataset contains relatively few failed molecules from which models can learn attrition liabilities. FLAb and FLAb2 address a different need by collecting mutational fitness landscapes. FLAb aggregates 13 384 measurements from 17 landscapes across expression, thermostability, immunogenicity, aggregation, polyreactivity, and binding affinity [32]. FLAb2 expands this resource to >4 million antibody sequences from 32 studies and adds pharmacokinetics as a seventh fitness dimension [33]. In practical terms, small, curated panels such as Jain et al. are useful for benchmarking and sanity-checking model behavior, whereas landscape-style resources such as FLAb and FLAb2 are more informative for learning local sequence–fitness relationships than for claiming universal developability ranking. They are useful for testing whether sequence and structure representations capture local fitness changes, but they still inherit differences in assay design across source studies.

More standardized assay panels reduce one source of label noise. PROPHET-Ab contains measurements for 246 approved therapeutics, clinical-stage candidates, and preregistration molecules across 10 assays, including expression titer, purity, monomer content, thermostability, hydrophobicity, polyreactivity, self-association, and heparin binding [34]. Its standardized workflow makes the labels easier to compare across antibodies, although the sample size remains modest for training high-capacity models. Database Of Therapeutic Antibody Developability (DOTAD) takes a broader aggregation strategy by linking therapeutic antibody sequences with experimental and predicted developability attributes, clinical status, structural identifiers, and immunogenicity rates [35]. This makes DOTAD useful for retrieval and model evaluation, but its mixed label provenance means experimental and predicted attributes should not be treated as equivalent training targets. As a result, PROPHET-Ab is especially useful when the goal is assay-consistent comparison across candidates, whereas DOTAD is more useful for curation, contextualization, and hypothesis generation than for naive supervised training.

Interaction resources provide a complementary view of antibody function. The Antigen-Specific Antibody Database (ASD) contains >1 million curated interaction records from 15 sources and 25 datasets, covering 865 153 unique antibodies and 9575 antigens [36]. Entries include heavy- and light-chain sequences where available, affinity measurements such as and IC values, qualitative confidence categories, structural identifiers, and experimental metadata. ASD brings heterogeneous binding data into a common format, but antigen imbalance and variation in measurement methods remain important caveats for affinity and specificity modeling.

Sequence and structure repositories are also central, even when they do not provide direct developability labels. The Observed Antibody Space (OAS) collates over 1 billion variable-region sequences from over 80 studies and is widely used for protein language model pretraining [37]. Structural repositories such as AbDb, SAbDab, and Thera-SAbDab provide 3D context for geometric deep learning, epitope mapping, and therapeutic annotation [38–40]. Specialized resources, including CoV-AbDab, Ab-Cov, sdAb-DB, and Antibodypedia, support domain-specific fine-tuning and dataset curation [41–44]. These repositories are most powerful when used for pretraining, annotation, and structure-aware modeling; by themselves, they do not provide the assay-grounded labels needed to claim clinically meaningful developability prediction.

Taken together, these resources show why no single dataset is sufficient for multi-objective antibody design. Practical workflows need to combine assay-labeled endpoints with sequence, structure, interaction, and clinical metadata while tracking label provenance. This provenance is especially important when models are used for optimization because a Pareto or Bayesian search can only make meaningful trade-offs when each objective is measured or predicted from evidence that is traceable, calibrated, and comparable across assays and models.

AI-based property prediction for antibody developability

Moving an antibody from discovery to market requires evaluation across production, storage, formulation, and administration, not only target binding [45]. Traditional assessment is often linear and resource-intensive, with developability liabilities detected only after affinity-driven hit selection [7, 46]. Computational models are now being used earlier to estimate properties such as solubility, thermal stability, immunogenicity, and pharmacokinetics, so that weak candidates can be deprioritized before costly experimental campaigns [45, 46].

We organize property prediction by endpoint because each developability liability is driven by different molecular determinants and is best modeled with different representations. Model choice therefore depends not only on predictive accuracy but also on whether the representation, assay context, and uncertainty estimate match the endpoint being modeled. Recent work combines Convolutional Neural Networks (CNNs), protein language models, structure-aware predictors, molecular dynamics–derived descriptors, and reference-distribution scoring [47–51]. Figure 2 maps these endpoints to the molecular determinants most often used in developability modeling: binding competence depends mainly on paratope geometry and antibody–antigen interface features; physical developability depends on surface hydrophobicity, charge distribution, conformational stability, and formulation context; and clinical performance depends on immunogenicity, Fc-mediated behavior, nonspecific clearance, and aggregation-related safety risks [50, 52–54]. The following subsections therefore review predictors by the design decision they support, rather than by architecture alone.

Figure 2.

Antibody developability is governed by interconnected molecular determinants. The schematic maps three major groups of antibody properties to the regions and molecular features that most commonly influence them; the arrows indicate dominant associations rather than exclusive regional control. ‘Binding competence’ is primarily shaped by the paratope, where affinity depends on shape complementarity, electrostatic complementarity, interface area, intermolecular contact networks, and conformational flexibility, while specificity and polyreactivity are influenced by paratope topology and nonspecific binding patches such as localized hydrophobic or basic positive surface regions. ‘Physical properties’ relevant to chemistry, manufacturing, and control are whole-antibody liabilities that can arise from both Fab and Fc regions: aggregation and solubility are associated with exposed hydrophobic patches, colloidal stability, and conformational or loop flexibility; viscosity is influenced by charge asymmetry, dipole moments, and charge clusters; and stability reflects chemical degradation liabilities as well as domain-level thermal stability. ‘Clinical safety and efficacy’ also depend on whole-antibody features. Immunogenicity and anti-drug antibody risk are shaped by sequence-derived T-cell epitopes, humanness, and aggregate-driven immune activation, whereas pharmacokinetics and half-life reflect FcRn-mediated recycling, nonspecific clearance driven partly by surface charge or hydrophobicity, and Fc glycosylation.

Predicting binding competence: affinity and specificity

The first prediction task in antibody design is usually binding competence: high affinity for the target with limited off-target binding. Affinity models support affinity maturation by ranking variants before experimental testing, while specificity models help detect nonspecific binding and polyreactivity that can lead to poor safety or pharmacokinetic behavior [47, 55, 56]. Data-driven methods have expanded the sequence space that can be explored during affinity maturation. For example, high-throughput binding data generated by assays such as Activity-specific Cell-Enrichment have been used to train deep contextual language models that predict the affinity of unseen variants [47].

Nonspecific binding is a separate failure mode rather than the absence of affinity. Sakhnini et al. showed that nonspecific binding can be predicted from the heavy-chain variable region and its complementarity-determining regions [55]. Their approach combined the Evolutionary Scale Modeling framework (ESM-1v) with logistic regression, while parallel analyses of sequence-based biophysical parameters identified theoretical isoelectric point as a major contributor to nonspecificity. These predictors are most useful when they are treated as early triage tools: they can identify variants that appear potent but are likely to fail because of charge-driven stickiness, poor specificity, or unfavorable humanness. Related computational design strategies have also been used in clinical programs such as AU-007, where binding, stability, and humanness were optimized together [57]. In this setting, the practical question is usually not whether a model can replace affinity maturation assays but whether it can help teams avoid spending experimental effort on variants with obvious downstream liabilities.

Physical properties: aggregation, viscosity, solubility, and stability

Aggregation, viscosity, solubility, and stability determine whether a potent antibody can be manufactured, stored, and delivered [7, 45, 52]. These endpoints are coupled: reduced stability can promote partial unfolding and aggregation, while poor solubility can worsen aggregation and viscosity at the high concentrations required for subcutaneous delivery [52, 58–60]. Computational models therefore need to capture both static sequence or structure features and dynamic behavior under stress conditions.

Aggregation is often triggered when antibody domains partially unfold and expose hydrophobic patches known as aggregation-prone regions. Stressors such as low pH, elevated temperature, or shear stress during pumping and filling can lower the energy barrier for unfolding and accelerate aggregation [58, 59]. Aggregates can cross-link B-cell receptors, break immune tolerance, and induce anti-drug antibodies (ADAs), which may cause rapid drug clearance, loss of efficacy, or severe hypersensitivity reactions [7, 61]. Structure-based tools such as Aggrescan 3D (A3D) identify aggregation-prone regions from static structures [62]. Newer models add conformational information. The Antibody Language Ensemble Fusion (AbLEF) framework, for example, predicts properties such as aggregation temperature from 3D conformational ensembles generated using Molecular Operating Environment (MOE) stochastic titration and LowModeMD conformational sampling [48]. Its transformer-based fusion architecture combines residue distance maps from multiple conformations, allowing the model to use dynamic information that is absent from a single structure. For design, this distinction matters because static predictors can flag exposed liabilities, whereas ensemble-based models may better identify liabilities that emerge only after conformational fluctuation or stress.

The industry shift toward subcutaneous administration has elevated viscosity to a primary developability concern because these products often require high-concentration formulations [45, 59]. At these concentrations, electrostatic attraction between oppositely charged patches on the Variable Fragments (Fv) is a primary driver of viscosity [63, 64]. Computational frameworks such as the Spatial Charge Map (SCM) and DeepSCM map and score these negative electrostatic patches [63, 64]. These scores have been reported to distinguish high-viscosity candidates in clinical antibody datasets and can outperform conventional net-charge calculations in such settings. Viscosity prediction is most useful when interpreted together with solubility and formulation conditions because concentration, buffer, pH, ionic strength, and excipients jointly determine whether a high-concentration antibody solution remains injectable without phase separation or gelation.

Solubility describes the concentration limit at which a monomeric protein remains dissolved in a given buffer without precipitation [60]. Poor solubility affects storage, production, and delivery, and often reflects net charge, surface-exposed hydrophobic patches, or both [60, 65]. Solubility predictors have moved from early sequence-based SVMs and Gradient Boosting Machines, such as SOLpro and PaRSnIP, to specialized predictors such as SOLart [65–67], which detect motifs associated with limited solubility. These models can flag sequences that may cause viscosity or aggregation problems in concentrated formulations, but their predictions should be interpreted in the buffer and concentration context in which the candidate will be used.

Stability measures whether an antibody maintains structure, binding affinity, and function over time under storage or stress conditions [52]. It is central to manufacturability because unstable antibodies are more susceptible to unfolding, which can feed into aggregation and alter solubility or viscosity [52, 58]. Harmalkar et al. developed an Machine Learning (ML) model that uses sequence and structure features to predict the thermostability of scFv antibodies and identify stabilizing mutations [68]. TEMPRO, developed by Alvarez and Dean, uses ESM-2-derived embeddings to predict the melting temperature () of nanobodies [69]. These examples support using stability predictors early in design, when stabilizing mutations can still be introduced before unfolding, aggregation, solubility, and viscosity liabilities become coupled.

Clinical safety and efficacy: immunogenicity and pharmacokinetics

An antibody therapeutic must also have an acceptable safety profile and pharmacokinetic (PK) behavior in the human body [70]. These endpoints are harder to model than many in vitro biophysical properties because they depend on molecular sequence, aggregation state, patient immune background, Fc-mediated recycling, and nonspecific tissue interactions [53, 54, 71]. Immunogenicity is a major safety concern, as ADAs can compromise efficacy and induce adverse reactions [71, 72]. Therapeutic antibodies, particularly those engineered or derived from nonhuman species, contain sequences that may be recognized as foreign. In the immunogenic cascade, antigen-presenting cells take up the therapeutic, degrade it into peptides, and present these peptides through Major Histocompatibility Complex (MHC) Class II molecules to T-helper cells [73]. Predictive models therefore usually focus on T-cell epitopes, namely, peptides predicted to bind human MHC class II alleles.

The computational prediction of antibody humanness is often used as a surrogate for immunogenicity risk. This approach assumes that sequences statistically similar to the natural human repertoire are more likely to be tolerated, although humanness alone cannot capture all T-cell epitope, aggregation-related, or patient-specific contributors to immunogenicity. Current AI-driven predictions often assess B-cell epitope exposure, T-cell epitope content, or repertoire-level humanness. Tools such as the Immune Epitope Database Analysis Resource (IEDB-AR), EpiMatrix, and iTope examine primary sequences to identify epitopes [74, 75]. More recently, the Deane group trained a Random Forest classifier on nearly 650 million nonredundant antibody sequences to produce a humanness score used by the tool HumAb [76]. The open-source platform BioPhi also integrates Sapiens, a Transformer-based model, to aid antibody humanization [77]. In design workflows, these tools are most defensible as relative triage methods for comparing variants and prioritizing humanization targets, not as stand-alone predictors of clinical immunogenicity. In therapeutic programs, apparent improvements in humanness do not automatically translate into lower patient-level risk without supporting assay and clinical context.

A dual-library sequence-based method, ImmunoSeq, has recently been proposed for immunogenicity risk prediction and sequence optimization. By comparing antibody peptide fragments against a large self-library of human peptides and a nonself library of mouse antibody peptides, ImmunoSeq generates a normalized hit rate that correlates with clinical ADA incidence [78].

PK clearance determines the half-life of an antibody [53]. Clinical-stage analyses have linked nonspecific-binding propensity, hydrophobicity, and charged regions to rapid clearance [54]; in one curated clinical dataset, statistical analyses and a Random Forest classifier identified polyspecificity reagent binding and estimated isoelectric point as useful discriminators of fast- and slow-clearing antibodies [54]. When combined with affinity and safety predictors, PK models help rank candidates by both potency and likely clinical behavior, especially when model outputs are calibrated against comparable antibody formats and dosing contexts.

Holistic scoring frameworks

Single-endpoint predictors are useful, but antibody selection usually requires a broader view of the candidate’s developability profile [7, 50]. Holistic scoring frameworks combine several liability predictions, reference distributions, or learned representations to support early triage and candidate ranking.

Reference-distribution approaches

The Therapeutic Antibody Profiler is a reference-distribution method, analogous to Lipinski’s rule of five for small-molecule drugs [50]. It asks whether the physicochemical profile of a query antibody falls within the range observed for clinical-stage therapeutics. AB-Panda extends this idea by combining AlphaFold2 structure generation with CDR-level metrics, including unit-area hydrophobic value and charge, to flag outlier sequences using recommended ranges derived from 919 clinical antibodies [79]. These visual surface–property readouts can guide sequence engineering and help detect developability liabilities. Their main strength is interpretability: they show which local surface features make a candidate unusual relative to therapeutic antibodies.

Sweet-Jones and Martin proposed a triage pipeline that uses protein language models without explicitly modeling individual physicochemical liabilities [80]. Sequences are encoded with the antibody-specific transformer AntiBERTy [81], and an unsupervised kernel-PCA layer projects these embeddings into a 2D space in which clinically approved antibodies form a tight cluster. Candidates outside the variance of this cluster are discarded. A second supervised layer trains a linear support-vector classifier to distinguish 115 approved from 150 discontinued therapeutics using 2500 features selected by F-regression. The pipeline does not directly compute aggregation propensity, thermostability, or pI; instead, it estimates developability from the similarity between a candidate’s learned representation and those of antibodies that have reached the market. This approach can be useful for early filtering, but it also depends on the assumption that approved antibodies define the relevant region of developable sequence space.

Feature-engineering pipelines

Recent hybrid frameworks use deep learning to predict biophysical behavior while preserving links to interpretable molecular descriptors. Wu and colleagues used DeepSP to generate surface descriptors, including Spatial Aggregation Propensity and Charge Map features, directly from variable region sequences [49]. These learned descriptors produced predictive accuracy comparable to Molecular Dynamics (MD)–derived features, especially for properties influenced by electrostatic and hydrophobic interactions, while reducing computational cost. Park and Izadi’s MolDesk uses Gaussian-accelerated MD (GaMD) for conformational sampling and computes descriptors such as electrostatic potentials and hydrophobic patches for developability prediction [82]. PROPERMAB provides an open-source framework for building models for specific developability problems and scaling them to repertoire-level sequence datasets by pretraining simple models to predict structure-derived features from fast sequence inputs [51]. These pipelines are valuable when they turn slow structural or simulation-derived measurements into scalable sequence-level screens.

The main shift in antibody property prediction is not simply from shallow to deep models. It is a move from isolated descriptors toward predictors that combine sequence, structure, dynamics, assay labels, and reference distributions [48–51, 79]. This point is especially important for generative design, where large libraries can contain many apparent binders but only a small subset may survive simultaneous developability filters. The remaining challenge is decision quality: a predictor must be accurate enough, calibrated enough, and interpretable enough to guide which candidates are synthesized, tested, reformulated, or discarded. Its value also depends on whether the training labels match the intended assay, formulation condition, antibody format, and sequence distribution. This requirement leads directly to multi-objective optimization, where affinity must be weighed against viscosity, stability, immunogenicity, pharmacokinetics, and other developability constraints.

Multi-objective optimization

Once prediction moves beyond single endpoints, the problem becomes how to choose among competing objectives. A useful antibody must bind its target, but it also has to remain stable, soluble, manufacturable, and safe. Multi-objective optimization (MPO) turns these trade-offs into a ranking problem for candidate selection and experiment prioritization.

In practice, MPO works only when the underlying models are calibrated and the assay labels match the intended use. Objectives are properties to improve, such as affinity, specificity, stability, solubility, viscosity, PK, and immunogenicity. Constraints rule out sequences with clear liabilities, including unwanted glycosylation sequons, extreme charge, poor manufacturability, or formulation limits. Uncertainty tells us when a predicted gain is too weak to trust, especially for candidates far from the training distribution.

Pareto optimization approaches

Pareto optimization is a common framework for handling competing objectives in protein and molecular design [15, 16, 83]. Rather than collapsing all properties into a single score, it identifies a frontier of nondominated solutions. On this frontier, improving one property requires sacrificing another. Figure 3 illustrates this idea in a simplified two-objective setting.

Figure 3.

Pareto optimization identifies balanced antibody candidates across competing objectives. Each point represents a candidate antibody evaluated in a two-objective space defined here by affinity and developability liability, where the lower-left region corresponds to more desirable overall performance and the upper-right region is suboptimal. The Pareto frontier comprises nondominated solutions for which improvement in one objective can only be achieved by sacrificing the other. Candidates near the upper-left portion of the frontier reflect affinity-biased solutions, whereas those near the lower-right portion are developability-biased. The highlighted point represents a balanced multi-objective optimum selected as a practical compromise between potency and downstream developability, in contrast to single-objective optima that over-emphasize one property at the expense of the other.

In antibody engineering, Pareto-based workflows can be used to examine the affinity–developability trade-off. Multiple machine learning models, each predicting a specific property, can be combined within a Pareto framework so that project teams can compare candidates under specific constraints. The goal is not to maximize a single metric but to select candidates with an acceptable compromise across potency and downstream developability.

The Pareto frontier does not by itself identify the best antibody. It identifies the candidates that remain after dominated alternatives are removed. Final selection still depends on project priorities, acceptable risk, assay confidence, and experimental capacity. This distinction is important because a poorly calibrated predictor can make an apparent frontier look more reliable than it is [84, 85]. For antibody MPO, property scores should therefore be calibrated against assay-specific endpoints whenever possible, rather than treated as directly comparable raw predictions.

Bayesian optimization

Bayesian optimization (BO) provides an algorithmic strategy for choosing which candidates or conditions to test next after objectives and constraints have been defined. It is commonly used when function evaluations are expensive, as in antibody engineering where wet-lab validation is resource-intensive [86, 87]. BO builds a probabilistic surrogate model, often a Gaussian Process, and uses an acquisition function to select the next experiment. The acquisition function balances exploitation of promising regions with exploration of uncertain regions.

Recent literature describes three main strategies for applying BO to antibody design and development: restricting sequence search with biological or developability priors, optimizing experimental or formulation variables under constraints, and using uncertainty estimates to avoid model-exploitation artifacts.

Constraining the sequence space

A major challenge in protein design is the size of sequence space, which can lead to inefficient exploration. CloneBO addresses this by adding a prior based on natural B-cell evolution [88]. The underlying assumption is that clonal evolution already searches across affinity, stability, and expression. By training an large language model (CloneLM) on clonal families, CloneBO focuses the search on mutations that are evolutionarily plausible. This constraint narrows the search space and helped the method identify stable binders more efficiently than less constrained searches in in vitro validation experiments. The broader lesson is that biological priors can make BO more sample-efficient by avoiding regions that are unlikely to yield viable antibodies.

Similarly, AntBO uses a combinatorial strategy to design CDRH3 regions within a trust region [89]. This region restricts the search to sequences that satisfy hard developability constraints, such as net charge within , no amino acid repeats exceeding five, and no N-linked glycosylation sequons. AntBO uses Gaussian processes with specialized kernels and iteratively proposes neighbors around the best sequences, reducing the number of experimental evaluations needed to find developable candidates. This example shows how hard constraints can make search experimentally tractable before more nuanced Pareto ranking is applied.

Optimizing formulation variables

BO is not limited to sequence design; it can also support process development. Waibel et al. demonstrated a BO workflow for antibody formulation, treating melting temperature (), diffusion interaction (), and monomer retention as independent Gaussian Process surrogates [90]. The workflow explored six formulation variables, including sorbitol, arginine, pH, and the fractional contributions of aspartic, glutamic, and acetic acids, under osmolality constraints. A constrained Nondominated Sorting Genetic Algorithm II (NSGA-II) search, an evolutionary multi-objective optimizer, generated a Pareto front from the surrogates, and the next experiment was chosen either to minimize distance to the front or to maximize distance from already sampled space. After 33 experiments, the workflow identified formulations that increased by 1.3 C and improved . This study shows how BO can navigate formulation space while preserving process constraints and reporting interpretable trade-offs.

Uncertainty-aware filtering

Model confidence becomes especially important when exploring novel sequence space. RESP2 introduces an uncertainty-aware pipeline that optimizes antibody binding affinity across multiple antigen variants while enforcing developability constraints [91]. A surrogate model maps paired antibody–antigen sequences to an enrichment score that serves as a proxy for affinity. During training, the model also estimates predictive uncertainty, so sequences far from the observed data receive higher variance. This uncertainty estimate is used in a modified simulated-annealing scheme that accepts a mutation only if it improves or preserves predicted affinity against every antigen variant and if the posterior variance remains below an adaptive threshold. Candidates from this in silico directed-evolution step are then passed through a developability filter and ranked by a Pareto-like compromise between mean predicted affinity and model confidence. This design helps reject sequences that appear favorable mainly because they fall outside the model’s training distribution, a common failure mode when optimization exploits model blind spots.

Active learning for efficient optimization

While BO provides one algorithmic strategy, Active Learning (AL) describes the broader iterative workflow. BO can serve as one acquisition policy within an AL loop: the model proposes variants or conditions, the wet lab tests them, and the results are used to update the model for the next round.

In academic settings, the ALLM-Ab framework illustrates how this loop can be implemented in low-data regimes [92]. It generates candidate sequences from a fine-tuned pLM and evaluates them with a multi-objective function that combines predicted affinity with in silico developability metrics such as isoelectric point, hydropathy, and instability index. Top-ranked Pareto-optimal candidates are selected for laboratory testing, and the new experimental data are used to retrain the pLM. To reduce overfitting when updating a billion-parameter model with limited new data, ALLM-Ab uses Parameter-Efficient Fine-Tuning through LoRA (Low-Rank Adaptation), which updates only a small set of adapter parameters. It also uses a Learning-to-Rank (ListMLE) strategy, training the model to rank candidates rather than predict exact affinity values. Offline validation on Deep Mutational Scanning datasets and online in silico trials showed that ALLM-Ab can find high-affinity variants while preserving key developability metrics. The framework was also applied to bispecific antibody optimization. This ranking-based formulation is useful because relative candidate ordering may be more stable than exact property prediction when data are scarce.

Conditional generative models

While Pareto methods, BO, and active learning are often used to rank, select, or iteratively refine candidates, conditional generative models extend MPO by proposing entirely new sequences conditioned on desired property profiles. In other words, the optimization problem shifts from choosing among existing variants to generating candidates that already reflect target binding and developability constraints. In therapeutic antibody development, these methods are best viewed as tools for candidate suggestion and hypothesis generation rather than autonomous design systems. Pereira et al. introduced an energy-based framework that replaces rigid hard constraints with a scalarized energy function [93]. The model samples from a Boltzmann distribution in which sequence probability depends on an energy score and a prior distribution of natural antibodies derived from a pretrained language model. This prior helps keep generated sequences close to human antibody repertoires while the energy term pushes candidates toward the desired property profile. The study used Metropolis–Hastings Markov Chain Monte Carlo (MCMC) and GFlowNet algorithms to sample diverse candidate libraries along the Pareto front. It also incorporated epistemic uncertainty from Gaussian Process predictors into the energy function, allowing the balance between exploration and exploitation to be adjusted through upper and lower confidence bound strategies. This type of formulation connects generative design with the same uncertainty-aware decision rules used in BO.

Diffusion models, which generate data by reversing a noise process, are also being adapted for MPO. Zhao et al. used a guided discrete diffusion framework with Soft Value–Based Decoding in Diffusion to steer generation toward candidates that meet multiple developability criteria [94]. The framework combines a masked diffusion model trained on paired sequences from the Observed Antibody Space with quantitative regressors based on ESM-2 embeddings. To optimize properties such as hydrophobicity and self-association at the same time, the method scalarizes the objectives by equally weighting their negative normalized values during denoising. This biases sampling toward sequences with favorable predicted developability scores without retraining predictors on partially masked data. The advantage of this strategy is modularity: property predictors can guide generation even when the generative model was not trained directly on every target endpoint.

Taking this further, AbNovo adapts Direct Preference Optimization to antibody design [95]. It pretrains an antigen-conditioned diffusion model for co-designing CDR sequences and structures, then fine-tunes the model with binding-related rewards while using explicit developability constraints to limit unacceptable candidates. This constrained preference formulation is more important than the specific training machinery for the present discussion: it shows how generative models can be aligned with project-level selection criteria rather than optimized only for affinity. AbNovo illustrates a broader trend in which preference and constrained-optimization methods are becoming a bridge between generative modeling and developability-aware candidate selection.

Together, these generative approaches differ in architecture but share the same design principle: generation is no longer treated as unconstrained sequence sampling but as property-conditioned search shaped by priors, developability filters, preference functions, or uncertainty-aware objectives.

Across these approaches, the central challenge is not only generating high-scoring sequences but preventing optimization from exploiting poorly calibrated predictors. Over-optimization against imperfect surrogate models can easily produce attractive in silico candidates that do not survive experimental validation, especially when generation moves beyond the training distribution. Practical MPO workflows therefore need calibrated property models, explicit constraints, uncertainty-aware candidate selection, and experimental feedback loops. These requirements become even more important when moving beyond canonical IgG antibodies to formats with different structural and manufacturing liabilities.

Extensions to other antibody formats

Beyond conventional monoclonal antibodies, bispecific antibodies and single-domain antibodies such as Variable domain heavy chain antibody (VHHs) have become important therapeutic formats because they can enable dual targeting, improve tissue penetration, and access epitopes that are difficult for full-length IgGs to engage [96, 97]. These advantages, however, come with format-specific developability risks. Standard IgG rules provide a useful baseline, but they should not be treated as universal priors for every antibody architecture. Format changes can alter the relevant failure modes, the assays needed to detect them, and the model inputs required for reliable prediction. Extending AI developability workflows to new antibody formats is therefore a distribution-shift and objective-redefinition problem, not only a broader application of IgG-trained predictors. Figure 4 summarizes how the main developability rules change across IgG, nanobody, and bispecific antibody formats.

Figure 4.

Different antibody formats introduce distinct developability rules and liabilities. Left: conventional IgG monoclonal antibodies provide the reference developability framework, with paired heavy and light chains, an Fc region that supports long serum half-life, and generally well-established rules for stability, manufacturability, and formulation. Middle: nanobodies (VHHs) are compact single-domain binders with a convex paratope and often long CDR3 loops, enabling access to challenging epitopes but also introducing distinct developability concerns, including the absence of an Fc region with consequent rapid clearance and the presence of exposed hydrophobic surface patches. Right: bispecific antibodies combine binding modules from different parental antibodies to achieve dual targeting, but this increased architectural complexity creates both assembly-related challenges, such as heavy/light-chain mispairing and incorrect assembly, and broader physicochemical liabilities, including charge asymmetry, polyreactivity, high viscosity, and product heterogeneity. Together, these comparisons illustrate why developability assessment and multi-objective optimization must be adapted to antibody format rather than transferred directly from standard IgG-based rules.

Bispecific and multispecific antibodies can range from IgG-like molecules to engineered fragment-based constructs, and these architectures introduce liabilities that standard monoclonal antibody workflows may miss [96]. Non-native linkers and variable domains such as single-chain variable fragments can reduce thermostability and increase aggregation propensity compared with parental antibodies. In a comparative study of 64 constructs across eight format families, engineered formats including bispecifics and antibody fragments showed greater stability and aggregation risks than full-length antibodies, with more complex formats carrying higher developability risk [98]. Engineered domain interfaces may create neoepitopes, alter immune effector activation, or change Fc-mediated behavior. Complex variable-domain geometries can also create charge imbalances and hydrophobic surface patches that increase risks of polyreactivity and high-concentration viscosity [96].

These liabilities create modeling targets that are absent from ordinary sequence-level IgG screening. For example, correct heavy/light-chain pairing is a format-level assembly constraint in IgG-like bispecific antibodies. Barlow et al. used Rosetta to design orthogonal constant-domain interfaces that reduce heavy/light-chain mispairing while preserving antigen binding and biophysical developability in tested bispecific constructs [99]. This example illustrates the type of architecture-aware computational workflow needed for noncanonical formats: computational screening should model not only the properties of individual binding domains but also assembly fidelity, purification heterogeneity, formulation behavior, and in vivo stability.

Therapeutic single-domain antibodies, including nanobodies, require a different set of criteria. Gordon et al. addressed this with the Therapeutic Nanobody Profiler, a computational platform calibrated against 36 clinical-stage nanobodies and validated with in vitro data from 108 sequences [97]. The study showed that nanobody developability is influenced by the absence of a light chain, elongated CDR3 loops, and an exposed framework-2 tetrad. By combining CDR3 compactness, tetrad motif identity, and surface patch analyses with conventional metrics such as hydrophobicity and charge, the profiler captures conformational features associated with solubility, stability, and aggregation propensity. Experimental mapping showed that both extended and compact CDR3 subtypes can be developable, suggesting that conformation-aware descriptors are more informative than total CDR length alone.

These examples show that format-specific developability is not only a downstream experimental issue. It changes the modeling problem itself. Datasets, descriptors, benchmarks, and optimization objectives need to distinguish IgG, VHH, bispecific, and other engineered formats rather than pooling them as interchangeable antibody sequences. Without this distinction, a model may appear accurate on aggregate benchmarks while failing in the format-specific settings where developability risk is highest.

Outlook and future directions

For antibody development teams, the central question is not whether an AI model looks technically impressive but whether its predictions are sufficiently reliable to alter an experimental decision. In practice, developability predictors are most useful as early triage tools. They can help deprioritize candidates with obvious liabilities, enrich the synthesis set for balanced candidates, and clarify which trade-offs deserve confirmatory assays. However, their value drops quickly when predictions are treated as portable truth across assay formats, concentration regimes, or antibody architectures. Used appropriately, these models should guide experimental prioritization rather than replace empirical validation.

This shift from proof-of-concept modeling to practical deployment is increasingly visible in industrial antibody discovery programs, although the public evidence remains dominated by selected case studies rather than systematic prospective evaluations [5]. Generate:Biomedicines’ GB-0895, a long-acting anti-TSLP antibody, provides one reported example of AI-supported engineering on a validated target, where potency, specificity, half-life, and dosing convenience were described as design objectives [100]. However, because the AI-design details for several industrial programs are still mainly supported by company communications rather than peer-reviewed prospective evaluations, these examples should be interpreted as early translational case studies rather than standardized evidence of platform-level generalizability.

Together, these examples suggest that the near-term value of AI in antibody development lies less in replacing wet-lab validation than in improving candidate prioritization, compressing design–test cycles, and bringing developability, pharmacokinetic, formulation, and therapeutic-window constraints earlier into discovery. At the same time, most disclosed programs provide limited information on training data, negative designs, blind prospective testing, and failure rates. They should therefore be interpreted as translational case studies rather than standardized evidence that a given platform generalizes across targets, formats, and development contexts.

A related consideration is the maturity of the underlying evidence base. Many foundation models, generative frameworks, and optimization strategies discussed in this review are still at an early stage of validation, with performance often reported in preprints, conference papers, or controlled benchmark settings. While these studies provide important methodological advances, their practical impact in antibody development will depend on independent replication, prospective testing, and demonstration across diverse targets, antibody formats, and experimental workflows.

Future progress will depend on better benchmarks and better integration with experiments. Unlike protein structure prediction, where CASP provides community-wide benchmarks [101], antibody developability prediction currently lacks broadly accepted prospective benchmark datasets and evaluation standards. Useful datasets should include negative outcomes, assay metadata, formulation context, and format-stratified test sets so that performance reflects transfer across antibody classes rather than interpolation within IgG-like sequences. For discovery-stage studies, reports should distinguish initial binding enrichment from orthogonally validated binding and from candidates that pass predefined developability filters. The most informative reports would disclose assay formats, antigen quality controls, validation criteria, developability filters, diversity of surviving candidates, and failure rates. More useful models will also combine sequence, structure, experimental conditions, and downstream measurements while exposing confidence and the features driving each prediction. Future deployable models will likely require calibrated uncertainty estimates to support risk-aware decision-making in experimental campaigns.

Taken together, these needs point toward closed-loop platforms in which generative modeling, active learning, and high-throughput experimentation operate as one iterative system. In such workflows, models propose candidates, prioritize informative experiments, update with newly generated data, and refine both affinity and developability objectives. Although autonomous laboratory demonstrations have emerged in adjacent areas of molecular design [102], antibody developability will require closed-loop systems that account for antibody-specific assay noise, format constraints, and manufacturability endpoints. The long-term opportunity is not fully autonomous antibody design but a more reliable design–build–test–learn cycle in which computation and experimentation improve one another through prospective validation, standardized benchmarking, and transparent reporting of both successful and failed designs.

Author contributions

Qianhui Jiang (Conceptualization [lead], Data curation [lead], Validation [lead], Writing—original draft [lead], Writing—review & editing [lead]), Jiahui Guan (Validation [supporting], Writing—review & editing [supporting]), Dan Yu (Validation [supporting], Writing—review & editing [supporting]), Pradeep Singh (Writing—review & editing [supporting]), George Pelekos (Supervision [supporting]), Edward Chin Man Lo (Supervision [supporting]), and Junwen Wang (Conceptualization [equal], Funding acquisition [lead], Project administration [lead], Supervision [lead])

Conflicts of interest

The authors declare no competing interests.

Funding

This work was supported by Collaborative Research Fund of Research Grants Council (C7015-23G), seed funding for collaborative research (2207101590) and basic research (2201101499) from the University of Hong Kong, and startup funds (207051059, 6010309) from Faculty of Dentistry, the University of Hong Kong to J.W.

Data availability

No new data were generated or analyzed in this review.

Ethics and consent statement

Not applicable; this review did not involve human participants or patient-level data.

Animal research statement

Not applicable.