AbGen: Antibody Generation & LLM Benchmark

Updated 3 July 2026

AbGen is a framework that integrates multi-objective generative antibody design with rigorous LLM benchmarks for experimental design.
It employs structure-aware, energy-based, diffusion, and retrieval-augmented approaches to co-design antibody sequences and structures.
By leveraging Pareto optimization and precise evaluation metrics, AbGen systematically balances trade-offs in affinity, solubility, and developability.

AbGen encompasses multiple distinct methodologies and benchmark platforms for generative antibody design, optimization, and evaluation, as well as for assessing scientific reasoning tasks performed by LLMs. The term "AbGen" is used both as shorthand for antibody generative modeling frameworks and as the title of specialized benchmarks in LLM experimental workflows. In antibody science, AbGen frameworks integrate structure-aware, energy-driven, and diffusion-based approaches for protein sequence and structure co-design, leveraging deep generative models to optimize biophysical, functional, and developability properties of monoclonal antibodies. In computational evaluation research, "AbGen" refers to a rigorous human-annotated benchmark for probing LLMs' capability in designing ablation studies for scientific papers. Both senses share a common focus on systematic, multi-objective optimization within complex domains.

1. Mathematical Foundations of Generative Antibody Design

AbGen frameworks formulate antibody sequence generation as probabilistic modeling over molecular properties with explicit or implicit structure conditionality. Central mathematical approaches include:

Energy-Based Generative Models: Given an antibody sequence $x$ (typically concatenating CDRH1, CDRH2, CDRH3), the generative distribution $p(x)$ is defined by

$p(x) = Z^{-1} p_{\text{HUM}}(x) \exp\bigl[-E(x)/T\bigr]$

where $p_{\text{HUM}}$ is a pretrained autoregressive "humanness" prior (e.g., IGLM), $E(x)$ is a weighted sum of property proxies, and $T$ a tunable temperature; $E(x) = -w \hat{A}_{\text{aff}}(x) - (1-w)\hat{A}_{\text{sol}}(x)$ captures affinity and solubility tradeoffs, sampled via MCMC or amortized GFlowNet flows (Pereira et al., 2024).

Diffusion-Based Models: Sequence (and structure) evolution is performed by discrete or continuous noising and denoising processes. For categorical-type diffusion,

$q(s_j^t|s_j^{t-1}) = \text{Mult}\big((1-\beta_t)\delta_{s_j^{t-1}} + \beta_t \tfrac{1}{20}\mathbf{1}\big)$

with reverse steps parameterized by neural networks conditioned on context, retrievals, and/or antigen features. Recent models (e.g., RADAb (Wang et al., 2024), AbMEGD (Chen et al., 26 Jun 2025), AlignAb (Wen et al., 2024)) extend this to E(3)- or SE(3)-equivariant sequence–structure diffusion with coupled losses, often enabling joint sequence and structural optimization.

Conjoined ODE/PDE Approaches: AbODE (Verma et al., 2023) formulates co-design as a coupled neural ODE system on joint antibody–antigen graphs, evolving both sequence logits and local 3D structure via differential attention on intra- and inter-molecular edges. The final state $z(T)$ encodes both categorical sequence distributions and structural predictions, optimizing

$\mathcal{L} = \mathcal{L}_{\text{seq}} + \mathcal{L}_{\text{structure}}$

with cross-entropy and geometric (angle/radius) losses.

2. Architectural Strategies and Multi-Objective Conditioning

Antibody generative models deploy several architectural innovations for capturing complex sequence–structure–function dependencies:

Template-Augmented Architectures: RADAb introduces retrieval augmentation, integrating homologous CDR motif fragments (retrieved and aligned via MASTER/Kabsch) into the generative process. This constrains sequence space and injects evolutionary plausibility (Wang et al., 2024).
Dual-Branch Denoising and Attention Fusion: Structural (global geometry) and evolutionary (local MSA-based) pathways process noisy sequences in parallel, employing invariant point attention (IPA), MSA Transformer-style axial attention, and fusion of logit outputs for conditional sampling.
Multi-Scale, E(3)-Equivariant Fusion: AbMEGD constructs representations at both residue and atomic levels, embedding local geometric frames, backbone torsions, and vector-scalar message-passing (ViS-MP), fused by IPA. All operations preserve geometric symmetries critical for robust generalization (Chen et al., 26 Jun 2025).
Integration with Protein LLMs and Pretrained Priors: Models such as RADAb and AlignAb embed evolutionary and humanness priors via ESM2 or transformer encoders, providing context-aware feature augmentation in the generative backbone.
Energy Alignment and Pareto Front Optimization: AlignAb extends direct preference optimization to the multi-objective domain, dynamically steering generation toward trade-offs between attraction (high binding) and repulsion (low steric clash) energies computed from Rosetta REF15 terms (Wen et al., 2024). Empirical Pareto fronts are generated by sweeping weightings in the reward.

3. Training Objectives, Losses, and Sampling Algorithms

Training regimes are shaped by the choice of generative foundation:

Diffusion Models minimize per-step KL divergence between noised forward and model posteriors, typically combining categorical, Gaussian, and rotational (SO(3) or SE(3)) losses over T steps. For instance, AbMEGD's objective is

$p(x)$ 0

where each term corresponds to sequence, position, and orientation denoising (Chen et al., 26 Jun 2025).

Energy-Based Sampling utilizes MCMC (Metropolis–Hastings, with burn-in and Hamming radius constraints), as well as amortized GFlowNet trajectories with trajectory-balance loss to match agent distributions to Boltzmann-Gibbs targets (Pereira et al., 2024).
Direct Preference Optimization in AlignAb applies KL-regularized policy gradient learning, sampling design pairs and optimizing for energy-differentiated rewards across attraction–repulsion space (Wen et al., 2024).
Pretraining and Embedding Fusion: Supervised pretraining (masked language modeling on OAS-derived repertoires) precedes or is fused with structure–function optimization, as in AlignAb's BERT-style transformer pretraining (Wen et al., 2024).

4. Benchmarks, Evaluation Metrics, and Empirical Outcomes

Quantitative evaluation in AbGen research leverages standard and derived metrics:

Metric Name	Description	Typical Range / Effect
Amino Acid Recovery (AAR%)	Fraction of correctly recovered input residues	29.8–70.5% (CDR-H3), >10% improvement with new models
Cα-RMSD (Å)	CDR backbone geometry deviation after alignment	0.65–3.48 Å, lower is better
Plausibility (AntiBERTy PLL)	Pseudo log-likelihood under an antibody PLM	e.g., –0.53 (higher less negative is better)
Improvement (%)	Fraction of designs with ΔG lower than native	20–37% (IMP), higher is better
Binding Energy (ΔΔG)	Rosetta interface energies (score units)	7–162 (lower is better)

Model comparisons for sequence recovery (Table 1, (Wang et al., 2024)) show that RADAb (AAR 57%, scRMSD 2.23) outperforms ablation baselines such as DiffAb and AbMPNN. AbMEGD demonstrates >10% AAR improvement vs. DiffAb, with competitive IMP and RMSD across SAbDab and viral antigen test sets (Chen et al., 26 Jun 2025). AlignAb achieves median repulsion energies ∼20% lower and attraction ∼15% higher than DiffAb, with corresponding gains in calculated ΔG and dominated-volume Pareto metrics (Wen et al., 2024).

Energy-based sampling frameworks produce empirically diverse and novel sequences (mean pairwise Hamming distance ∼4.5–5 mutations), and systematic exploration of solubility-affinity tradeoffs (high solubility sequences sacrifice up to 10× in affinity compared to low-solubility, high-affinity optimals) (Pereira et al., 2024).

5. Retrieval-Augmented and Evolutionary Conditioning

The integration of retrieval-based and evolutionary priors defines a major innovation in AbGen. RADAb restricts generative search to a functionally plausible subspace by:

Retrieving top-k backbone-matched CDR fragments via RMSD thresholds, forming a pseudo-MSA as input for local attention/modeling.
Conditioning denoising on both global geometric features and local evolutionary signal (retrieved homologs), fused in logit space and combined with PLM embeddings.
This approach mitigates overfitting and improves generalization, especially on long CDR-H3 loops, which are underrepresented in training data (Wang et al., 2024).

A plausible implication is that semi-parametric or retrieval-augmented generative models will become standard in AbGen pipelines, serving as "plug-in" components for otherwise de novo frameworks.

6. Multi-Objective Optimization, Alignment, and Trade-Off Management

Advanced AbGen workflows address the intrinsic Pareto-frontier optimization of antibody design problems:

Energy-based models sample from the Boltzmann-Gibbs distribution over human-likeness, affinity, and solubility, with temperature and β acquisition controlling the exploration–exploitation balance (Pereira et al., 2024).
AlignAb's Pareto-Optimal Energy Alignment (POEA) extends DPO-style losses to manage attraction–repulsion conflicts, tracing the empirical Pareto set by varying reward weightings and conducting iterative online policy refinement.
The use of temperature scaling during inference preserves sequence diversity and matches the entropy profile of natural antibody repertoires (Wen et al., 2024).

This suggests that optimal antibody candidates are identified not by maximizing any single property, but by explicit exploration and surface tracing of the Pareto frontier, with post-hoc empirical selection of maximally diverse, high-quality binders.

7. AbGen as a Benchmark for LLM Reasoning in Experimental Science

AbGen is also defined as a 1,500-example, expert-annotated LLM benchmark assessing the generation and evaluation of ablation study designs in scientific research (Zhao et al., 17 Jul 2025). Key characteristics:

Each instance includes a research context (1,800 words, excerpted from an NLP paper), a specified module, and an expert-provided reference ablation (objective, experiment protocol, and results).
LLMs are evaluated on their proposed ablation's importance, faithfulness, and soundness, each scored 1–5 by human judges.
Automated LLM-as-Judge methods (e.g., GPT-4.1-mini, Llama, Gemini) show low correlation with human scoring, both at the instance and system level.
Failure modes include context misalignment, protocol ambiguity, and logical inconsistency, highlighting the difficulty of nuanced scientific reasoning for current LLMs.
Recommendations include specialized constraint-based evaluation, agent-based or multi-turn LLM protocols, and public release of model outputs for future benchmarking.

In sum, AbGen constitutes both a suite of computational antibody generative frameworks (with innovations in retrieval augmentation, diffusion/ODE generative processes, evolutionary priors, and Pareto optimization) and a platform for benchmarking the capabilities of modern LLMs in experimental science design (Wang et al., 2024, Pereira et al., 2024, Chen et al., 26 Jun 2025, Verma et al., 2023, Wen et al., 2024, Zhao et al., 17 Jul 2025).