SimBinder-IF: High-Affinity Antibody Design

• SimBinder-IF is a structure-aware computational framework that adapts the ESM-IF model by freezing its structure encoder and fine-tuning only the decoder to generate high-affinity antibodies.
• It employs Simple Preference Optimization (SimPO) to directly optimize the log-likelihood margin between strong and weak binders, resulting in markedly improved affinity prediction.
• Empirical results show up to 156% improvement in zero-shot affinity correlation and enhanced de novo antibody generation efficiency, validated on comprehensive antibody assays.

SimBinder-IF is a structure-aware computational framework for high-affinity antibody design via affinity-optimized inverse folding. It adapts the ESM-IF inverse folding model into an antibody sequence generator by training only the decoder component with preference optimization while freezing the structure encoder, enabling efficient learning of affinity information from experimental mutational scans. SimBinder-IF exhibits substantial improvements in both affinity prediction and de novo antibody generation efficiency, outperforming prior approaches in benchmark and case study settings (Zhao et al., 19 Dec 2025).

1. Motivation and Rationale

The clinical potency of antibody therapeutics is tightly linked to high-affinity binding. Traditional affinity maturation methods, such as phage display coupled with SPR/BLI validation, are slow, low-throughput, and costly. Most publicly available protein LLMs (PLMs), including ESM-IF, are trained in a self-supervised manner on general protein databases, where antibody–antigen binding interfaces represent less than 3% of PDB entries. This results in a lack of explicit bias toward high-affinity antibody–antigen interactions. Empirically, vanilla ESM-IF achieves only modest mean Spearman correlation ($\rho \approx 0.3$) between model-assigned log-likelihoods and experimental affinity, limiting its utility for prioritizing strong binders.

Existing preference optimization strategies such as Direct Preference Optimization (DPO) introduce significant computational overhead, are prone to catastrophic forgetting, and utilize reference-ratio rewards that misalign with the inference metric (absolute log-likelihood).

2. Model Architecture and SimBinder-IF Adaptations

SimBinder-IF builds on ESM-IF, an inverse folding model comprising a structure encoder $f_\text{enc}$, a sequence tokenizer $f_\text{tok}$, and an autoregressive decoder $f_\text{dec}$. The inputs are 3D backbone coordinates of the antibody–antigen complex and partial sequence tokens. The structure encoder is an SE(3)-equivariant GVP–GNN that converts atomic and geometric features into residue embeddings $E \in \mathbb{R}^{n \times d}$. The decoder predicts residue-level log-likelihoods in an autoregressive fashion: $$p(Y|X) = \prod_{i=1}^n p(y_i | y_{<i}, E), \quad p(y_i | y_{<i}, E) = \text{softmax}(W h_i + b)$$

The core SimBinder-IF adaptation consists of freezing the entire structure encoder and Transformer encoder layers to preserve structural priors, and fine-tuning exclusively the decoder parameters (approximately 25 million weights, or $\simeq$18% of the full ESM-IF model). Inputs at each training step are the wild-type 3D structure $X$ and two antibody sequences: $y_w$ (higher experimental affinity) and $y_\ell$ (lower affinity), each paired with the antigen sequence. The output is the length-normalized log-likelihood score from the decoder.

3. Preference Optimization: Simple Preference Optimization (SimPO)

SimBinder-IF employs Simple Preference Optimization (SimPO), distinguishing itself from DPO by using a reward based on the same quantity needed at inference—model's absolute log-likelihood—rather than log-ratio with a reference model. The SimPO reward is given by: $$r_\text{SimPO}(x, y) = \frac{\beta}{|y|}\sum_{i=1}^{|y|} \log \pi_\theta(y_i | x, y_{<i})$$ and the loss function enforces a fixed margin $\gamma$ between preferred and nonpreferred sequences with a Bradley–Terry loss: $$L_\text{SimPO} = -\log \sigma\big(r_\text{SimPO}(x, y_w) - r_\text{SimPO}(x, y_\ell) - \gamma\big)$$ where $\beta=0.1$ and $\gamma=0.1$. This formulation aligns training and inference objectives, directly increasing the absolute log-likelihood of higher-affinity sequences.

4. Datasets, Training Protocol, and Computational Efficiency

The AbBiBench dataset (Zhao et al. 2025) was utilized, comprising eleven antibody–antigen deep-mutational scanning assays (~155,000 mutants) from both oncology and viral targets, with standardized binding scores ($-\log_{10} K_d$ or log enrichment; higher is stronger). Experimental splits were 60% train, 30% test, 10% validation within assay (supervised), and four complete assays held out as unseen test sets (zero-shot).

Training used AdamW (lr: $1\times10^{-4}$), batch size 32, for 3 epochs on a single NVIDIA H200 GPU. Decoder-only SimPO training used approximately 40% less peak GPU memory and 35% less wall-clock time per epoch compared to full-model fine-tuning.

Frozen Parameters	Trained Parameters	GPU Memory Reduction	Wall-clock Time Reduction
Encoder + Transformer layers	Decoder ($\simeq$18%)	$\simeq$40%	$\simeq$35%

Training protocol and efficiency metrics from (Zhao et al., 19 Dec 2025).

5. Quantitative Results and Benchmark Performance

On the AbBiBench benchmark:

• Mean Spearman correlation ($\rho$) between log-likelihood and experimental affinity:
  • Supervised (11 assays): Vanilla ESM-IF $= 0.264$, SimBinder-IF $= 0.410$ ($+55$\% gain)
  • Zero-shot (4 held-out assays): Vanilla ESM-IF $= 0.115$, SimBinder-IF $= 0.294$ ($+156$\% gain)
• Top-10 precision for %%%%24$\beta=0.1$25%%%% affinity improvement: SimBinder-IF exceeded vanilla ESM-IF, DPO, and SaProt by 5–15 percentage points (p < 0.01, paired Wilcoxon test), in both supervised and zero-shot assays.

Setting	Vanilla ESM-IF $\rho$	SimBinder-IF $\rho$	Relative Gain
Supervised	0.264	0.410	+55%
Zero-shot	0.115	0.294	+156%

6. Case Study: F045-092 Redesign for A/California/04/2009 (pdmH1N1)

A case study involved redesigning F045-092, a human anti-influenza antibody, to bind pdmH1N1 (A/California/04/2009), using AlphaFold 3-predicted structures due to the lack of an experimental complex. Each model (SaProt, vanilla ESM-IF, SimBinder-IF) generated 1,500 CDR-H3 variants with up to five mutations, conditioned on the predicted backbone.

Screening was performed in two stages: a plausibility and affinity prescreen (AntiBERTy PLL and FoldX $\Delta \Delta G$), followed by structural ranking (AF3 pLDDT, epitope $\Delta$SASA, ProteinMPNN PLL, pTM, and ipLDDT). Pareto-optimal selection produced 20 finalists: 1 from SaProt, 12 from vanilla ESM-IF, 7 from SimBinder-IF.

• Mean FoldX $\Delta \Delta G$ (kcal·mol$^{-1}$):

ESM-IF variants: $-46.57$ SimBinder-IF variants: $-75.16$ ($\sim$1.6$\times$ stronger predicted binding)

SimBinder-IF designs exhibited high scores for foldability (ProteinMPNN PLL, pLDDT), developability, and epitope targeting, predominantly focusing on the immunodominant HA head.

7. Mechanistic and Practical Insights

Freezing the structure encoder preserves geometric representations learned during pretraining, maintaining spatial interface fidelity. By employing a reward signal identical at training and inference (absolute log-likelihood), SimBinder-IF avoids train–test distribution mismatches present in DPO-based strategies. Decoder-only optimization allows preference for high-affinity binders to be introduced without erasing structural language knowledge.

Formally, SimBinder-IF maximizes log-likelihood margin between strong and weak binders: $$L_\text{SimPO} = -\log \sigma \left(\frac{\beta}{|y|}\sum_i \log \pi_\theta(y_{w,i}) - \frac{\beta}{|y|}\sum_i \log \pi_\theta(y_{\ell,i}) - \gamma \right)$$ At convergence, this yields higher ranking accuracy and affinity prediction, empirically confirmed by benchmark performance and precision@10 for substantial affinity increases. This parameter- and compute-efficient training strategy enables multi-assay learning on a single high-memory GPU while mitigating overfitting risks associated with small affinity datasets.

Together, the design choices of SimBinder-IF—frozen structural encoding, decoder-only preference optimization with a direct log-likelihood reward, and structured multi-task training—underlie its superior affinity modeling and robust, high-affinity antibody generation relative to full-model fine-tuning and alternative baselines (Zhao et al., 19 Dec 2025).