Ab-Affinity: SARS-CoV-2 Binding Predictor

• Ab-Affinity is a sequence-based antibody affinity predictor that estimates binding strength to a conserved SARS-CoV-2 HR2 peptide using scFv antibody sequences.
• It utilizes a BERT-style transformer encoder (ESM-2) fine-tuned on dense mutational datasets to predict log-transformed dissociation constants and discern affinity classes.
• The model supports rapid antibody screening by providing target-specific insights for optimization while reducing experimental burdens associated with traditional affinity assays.

Searching arXiv for the primary paper and closely related antibody-affinity works to ground the article. Ab-Affinity is a sequence-based antibody affinity predictor specialized for SARS-CoV-2-related binding, introduced as a LLM for estimating how strongly a single-chain fragment variable (scFv) antibody variant binds a fixed target peptide from the viral spike protein, specifically a conserved peptide in the HR2 region (Ashraf et al., 4 Mar 2026). In this formulation, the model does not encode an arbitrary paired antigen sequence; rather, the antigen is implicit in the task definition because all supervised measurements concern one SARS-CoV-2 HR2 peptide (Ashraf et al., 4 Mar 2026). The system is therefore best understood as a target-specialized affinity regressor built by fine-tuning a pretrained protein LLM so that antibody sequence representations become predictive not only of binding affinity itself, but also of affinity classes, improvement over a seed antibody, residue-level interaction patterns, and apparent thermostability trends (Ashraf et al., 4 Mar 2026).

1. Biological problem and task scope

Binding affinity is one of the central determinants of whether an antibody will be useful as a neutralizer or therapeutic lead (Ashraf et al., 4 Mar 2026). Antibodies recognize antigens through paratopes, typically formed by complementarity-determining regions in the variable domains of the heavy and light chains, and strong and specific binding is often required for effective neutralization (Ashraf et al., 4 Mar 2026). In practice, affinity measurement depends on assays such as Surface Plasmon Resonance, ELISA, or Bio-Layer Interferometry, together with generation and purification of many candidate antibodies, making the process expensive and slow (Ashraf et al., 4 Mar 2026). This creates a clear role for computational triage, particularly for SARS-CoV-2, where rapid screening and optimization remain valuable for variant-resilient therapeutic design (Ashraf et al., 4 Mar 2026).

Ab-Affinity addresses a narrower problem than generic antibody–antigen affinity prediction. The training target is a peptide in the spike HR2 region, chosen because it is conserved across SARS-CoV-2 variants and also across broader coronaviruses including SARS and MERS (Ashraf et al., 4 Mar 2026). That choice makes the system relevant to broad-reactive antibody design, but it also means the model is not a universal antibody–antigen predictor (Ashraf et al., 4 Mar 2026). A common misconception is to treat it as a dual-encoder system over arbitrary antibody and antigen sequences. In fact, the model takes only the antibody sequence as input, while the target peptide remains fixed by the dataset and task definition (Ashraf et al., 4 Mar 2026).

The paper explicitly positions antibody affinity prediction as more difficult than generic protein–protein interaction prediction because both paratope and epitope regions can be flexible and may include intrinsically disordered regions, for which structural data are sparse (Ashraf et al., 4 Mar 2026). This motivates a sequence model that learns how amino-acid changes alter affinity, especially in mutational neighborhoods relevant to antibody engineering (Ashraf et al., 4 Mar 2026). This emphasis differs from benchmarking frameworks that treat affinity as a broader cross-target ranking problem, such as AbRank (Liu et al., 21 Jun 2025), and from antigen-specific ranking formulations that adapt to contextual demonstrations at inference time, such as AbICL (Chen et al., 7 Jul 2026).

2. Model architecture and representation strategy

Ab-Affinity is a BERT-style transformer encoder based on ESM-2; the paper states that the architecture is “based on BERT, as implemented in ESM-2” (Ashraf et al., 4 Mar 2026). It consists of $N$ sequential encoder blocks, each containing multi-head self-attention followed by feed-forward layers, and the authors experimented with $N=6,12,33$, corresponding to models of about 8M, 35M, and 650M parameters, respectively (Ashraf et al., 4 Mar 2026). The final and best model is the 33-layer version fine-tuned from pretrained ESM-2 weights (Ashraf et al., 4 Mar 2026).

The antibody is represented as an scFv amino-acid sequence, which in this setting effectively combines variable heavy and variable light information into one sequence (Ashraf et al., 4 Mar 2026). The model processes this amino-acid token sequence through the ESM-2 encoder, and the output of the last encoder layer is taken as the sequence representation, described by the authors as “the embedding” (Ashraf et al., 4 Mar 2026). Depending on model size, the embedding dimension is 320, 480, or 1280 (Ashraf et al., 4 Mar 2026). On top of this embedding, the model places a single fully connected layer to regress binding affinity (Ashraf et al., 4 Mar 2026).

The prediction target is the log-transformed dissociation constant, specifically $\log_{10}(K_d)$ (Ashraf et al., 4 Mar 2026). Since lower $K_D$ indicates tighter binding, lower predicted values correspond to stronger affinity (Ashraf et al., 4 Mar 2026). The training loss is Mean Squared Error, and model selection uses the Pearson correlation coefficient on the validation set (Ashraf et al., 4 Mar 2026). The paper does not report more elaborate heads, alternative pooling strategies, explicit antigen inputs, or non-MSE objectives (Ashraf et al., 4 Mar 2026).

Architecturally, the novelty is not a new transformer block but the specialization of a large pretrained protein LLM to a high-throughput SARS-CoV-2 antibody engineering dataset (Ashraf et al., 4 Mar 2026). This places Ab-Affinity in a different category from sequence-only dual-stream antigen–antibody regressors such as DuaDeep-SeqAffinity (Boutorh et al., 26 Dec 2025), from structure-conditioned inverse-folding affinity rankers such as SimBinder-IF (Zhao et al., 19 Dec 2025), and from listwise or in-context ranking frameworks such as AbLWR (Xu et al., 13 Apr 2026) and AbICL (Chen et al., 7 Jul 2026).

3. Dataset, preprocessing, and training protocol

The training dataset comes from Engelhart et al. and contains 104,972 antibody variants generated by introducing one, two, or three amino-acid changes into antibodies derived from three seed candidates identified from a phage display library (Ashraf et al., 4 Mar 2026). The seed components are described as Ab-14-VH and Ab-14-VL, Ab-91-VH, and Ab-95-VH and Ab-95-VL (Ashraf et al., 4 Mar 2026). Each variant was assayed against the HR2 peptide, and each interaction had three biological replicate measurements of $K_D$, estimated by an indirect competitive binding assay (Ashraf et al., 4 Mar 2026).

Preprocessing is unusually specific. For each antibody–antigen pair, the authors took the arithmetic mean of the two closest $K_D$ values and discarded the third in order to reduce outlier effects (Ashraf et al., 4 Mar 2026). Antibodies with all three replicate values missing were removed (Ashraf et al., 4 Mar 2026). After preprocessing, 71,834 unique antibodies remained for training and evaluation, and the labels were then log-transformed to produce $\log_{10}(K_d)$ as the regression target (Ashraf et al., 4 Mar 2026).

The split strategy was 85% training and 15% validation, with the distribution of affinity values maintained (Ashraf et al., 4 Mar 2026). The paper also refers repeatedly to a held-out test set used for comparison plots, but the exact proportions or construction of that test set are not clearly described in the methods section (Ashraf et al., 4 Mar 2026). This underreporting is one of the main methodological limitations of the study. What is clear is that the best checkpoint in each setup was selected by Pearson correlation on the validation set (Ashraf et al., 4 Mar 2026).

Optimization used Adam with batch size 128 for 100 epochs on four NVIDIA A100 80GB GPUs (Ashraf et al., 4 Mar 2026). The encoder was fine-tuned from pretrained ESM-2 rather than frozen, and the authors also trained a counterpart model from random initialization to assess the contribution of pretrained protein knowledge (Ashraf et al., 4 Mar 2026). The paper states that the 33-layer fine-tuned model was best, implying that both model scale and protein-language pretraining matter (Ashraf et al., 4 Mar 2026). Specific settings such as learning rate, weight decay, warmup, dropout, and scheduler are not reported (Ashraf et al., 4 Mar 2026).

The data regime itself is highly consequential for interpretation. The antibodies occupy mutational neighborhoods around only three seed antibodies, with one to three substitutions each (Ashraf et al., 4 Mar 2026). This makes the benchmark highly relevant for local affinity optimization, but less clearly informative about de novo generalization far from the seed sequences (Ashraf et al., 4 Mar 2026). That limitation parallels broader concerns in affinity benchmarking about scaffold locality versus true out-of-distribution generalization, a distinction made explicit in AbRank (Liu et al., 21 Jun 2025).

4. Empirical performance and comparative evaluation

The principal evaluation metrics are Pearson correlation and Spearman rank correlation between predicted and measured affinity; for downstream classification tasks the paper uses ROC curves and AUC (Ashraf et al., 4 Mar 2026). In the full held-out test-set comparison, the main baselines are DG-Affinity, ESM-2 embeddings plus linear regression, and AbLang embeddings plus linear regression (Ashraf et al., 4 Mar 2026). The body text explicitly gives DG-Affinity’s Pearson correlation as 0.194 and states that Ab-Affinity has the highest Pearson and Spearman among the four methods (Ashraf et al., 4 Mar 2026). The exact numeric test-set correlations for ESM-2, AbLang, and Ab-Affinity are not printed in the text, so they are not reproducible from the paper excerpt (Ashraf et al., 4 Mar 2026).

The most concrete quantitative comparison is on the Ab-14-derived subsets 14H and 14L (Ashraf et al., 4 Mar 2026). The results reported in the paper are as follows.

Benchmark	Method	Reported result
14H	Ab-Affinity	Pearson 0.652, Spearman 0.526
14H	A2Binder	Pearson 0.642, Spearman 0.553
14H	ESM-F	Pearson 0.634, Spearman 0.516
14L	Ab-Affinity	Pearson 0.712, Spearman 0.713
14L	A2Binder	Pearson 0.683, Spearman 0.688
14L	ESM-F	Pearson 0.674, Spearman 0.681

These results show that Ab-Affinity achieved the best Pearson correlation on both 14H and 14L, and the best Spearman correlation on 14L (Ashraf et al., 4 Mar 2026). On 14H, its Spearman correlation is slightly below A2Binder, but the paper states that it “closely followed” (Ashraf et al., 4 Mar 2026). The authors also argue that these Ab-14-specific evaluations understate the broader value of the model because Ab-Affinity performed strongly when evaluated across all three seed antibody families (Ashraf et al., 4 Mar 2026).

The comparative interpretation is narrow but important. DG-Affinity, although previously strong on another antibody dataset, performed poorly on this SARS-CoV-2-specific task (Ashraf et al., 4 Mar 2026). Generic protein pretraining via ESM-2 and antibody-specific pretraining via AbLang remained reasonably competitive, but supervised specialization to the HR2 affinity dataset yielded the best reported correlation (Ashraf et al., 4 Mar 2026). This suggests that target-specific supervised fine-tuning can dominate generic pretrained embeddings in local antibody optimization settings (Ashraf et al., 4 Mar 2026).

5. Embedding geometry, downstream tasks, and interpretability

A major claim of Ab-Affinity is that fine-tuning reorganizes antibody sequence space into a functionally meaningful latent representation (Ashraf et al., 4 Mar 2026). Using t-SNE with perplexity 200 on the learned embeddings, the paper reports that Ab-Affinity arranges antibodies along a smooth affinity gradient, whereas raw ESM-2 embeddings do not clearly order antibodies by $\log K_d$ (Ashraf et al., 4 Mar 2026). This is presented as evidence that affinity supervision reshapes the latent space rather than merely improving scalar regression (Ashraf et al., 4 Mar 2026).

The same embeddings are used for two downstream classification tasks. One assigns antibodies to High, Medium, and Low affinity classes; the other predicts whether a mutant binds more strongly than its seed antibody (Ashraf et al., 4 Mar 2026). In both cases, classifiers built on Ab-Affinity embeddings yield ROC and AUC behavior reported as much better than classifiers built on raw ESM-2 embeddings, although exact AUC values are not given in the text excerpt (Ashraf et al., 4 Mar 2026). This suggests that the representation carries rank- and class-relevant information beyond the final regressed $\log_{10}(K_d)$ value.

The paper also analyzes transformer attention maps using the contact-analysis method of Rao et al. (Ashraf et al., 4 Mar 2026). It compares strong binders, defined as $\log K_d < 0.5$, to weak binders, defined as $N=6,12,33$0, and reports that the largest differences in attention-derived contact maps tend to occur in CDR-H1, CDR-H2, CDR-L1, or adjacent regions (Ashraf et al., 4 Mar 2026). This is biologically plausible because CDRs mediate antigen binding (Ashraf et al., 4 Mar 2026). The result is presented as interpretability evidence that the model’s internal focus aligns with known antibody binding sites (Ashraf et al., 4 Mar 2026).

A further analysis examines thermostability. The authors assembled a small external dataset of experimentally determined thermostability for 26 SARS-CoV-2 antibodies from two prior studies and report that Ab-Affinity embeddings separate antibodies into clusters with similar thermostability more clearly than raw ESM-2 embeddings in t-SNE projections (Ashraf et al., 4 Mar 2026). The paper does not provide a rigorous predictive benchmark for this claim, and the sample size is very small (Ashraf et al., 4 Mar 2026). A plausible implication is that affinity fine-tuning induces representations that capture some broader biophysical variation, but the evidence remains exploratory rather than conclusive.

6. Position within antibody-affinity modeling

Ab-Affinity occupies a specific niche within a rapidly diversifying literature on antibody affinity modeling. It differs from broad ranking benchmarks such as AbRank, which reframes affinity prediction as pairwise ranking across heterogeneous antibodies, antigens, and assay types (Liu et al., 21 Jun 2025). It also differs from antigen-specific contextual rankers such as AbICL, which use support demonstrations to adapt ranking behavior without gradient updates for a given antigen campaign (Chen et al., 7 Jul 2026). Ab-Affinity instead addresses a single conserved SARS-CoV-2 HR2 peptide and predicts a continuous $N=6,12,33$1 target from antibody sequence alone (Ashraf et al., 4 Mar 2026).

This specialization has both strengths and constraints. Relative to regression-averse frameworks such as AbRank (Liu et al., 21 Jun 2025) or listwise ranking systems such as AbLWR (Xu et al., 13 Apr 2026), Ab-Affinity retains a direct thermodynamic-like regression target, which is useful when labels are dense, internally consistent, and assay-matched. Its success therefore supports the view that absolute-value regression can still be effective when the antigen is fixed and the experimental design is highly controlled (Ashraf et al., 4 Mar 2026). By contrast, methods designed for heterogeneous multi-antigen settings often emphasize ranking because cross-assay calibration is difficult (Liu et al., 21 Jun 2025).

Ab-Affinity also differs from structure-first approaches. The AlphaFold 3 study on anti-CD47 antibodies evaluates structure generation followed by external MM/GBSA scoring, and concludes that AF3 is useful primarily as an upstream structural modeler rather than a direct affinity predictor (Xu et al., 18 Nov 2025). Ab-Affinity bypasses this route by learning from sequence only, motivated partly by the scarcity of reliable structural data for flexible antibody–antigen regions (Ashraf et al., 4 Mar 2026). Conversely, a plausible implication is that Ab-Affinity’s scope is strongest where target specialization and dense mutational affinity data are available, rather than in broad complex-level structure generalization.

From a methodological lineage perspective, Ab-Affinity can also be situated against earlier experimental work on sequence–affinity landscapes. Tite-Seq established that massively parallel titration curves can recover physical $N=6,12,33$2 values for antibody variants and thereby disentangle affinity from confounding expression and stability effects (Adams et al., 2016). Later analysis of Tite-Seq data showed that binding free-energy landscapes contain substantial epistasis and beneficial interaction effects (Adams et al., 2017). Ab-Affinity inherits the same general engineering logic—learning affinity-relevant structure from mutational sequence data—but implements it through large-scale protein LLM fine-tuning rather than direct biophysical landscape inference (Ashraf et al., 4 Mar 2026).

7. Limitations, interpretation, and significance

The paper’s limitations are explicit in several places and inferable in others. First, the supervised task concerns antibodies against one selected HR2 peptide rather than arbitrary antigens, so Ab-Affinity is not a universal antibody–antigen affinity predictor (Ashraf et al., 4 Mar 2026). Claims about broader SARS-CoV-2 or pan-antigen generalization should therefore be interpreted narrowly (Ashraf et al., 4 Mar 2026). Second, the dataset consists of mutational neighborhoods around only three seed antibodies, which makes the model particularly relevant for local optimization but leaves de novo extrapolation uncertain (Ashraf et al., 4 Mar 2026). Third, the train/validation/test protocol is incompletely specified, which complicates precise assessment of held-out novelty (Ashraf et al., 4 Mar 2026).

Reporting incompleteness extends to implementation details and some comparison metrics. The main full-test-set numeric correlations are not fully printed, some hyperparameters are missing, and pretrained checkpoint details are only partially described (Ashraf et al., 4 Mar 2026). These omissions do not negate the reported results, but they limit reproducibility. More broadly, the paper offers strong evidence for target-specialized mutational affinity modeling, not for universal antibody design.

Within that scope, the significance is clear. Ab-Affinity shows that supervised fine-tuning of a large protein LLM on a high-throughput SARS-CoV-2 mutational affinity dataset can produce an antibody sequence representation that is predictive of $N=6,12,33$3, useful for ranking and classification, and partly interpretable through latent geometry and attention behavior (Ashraf et al., 4 Mar 2026). Its practical value lies in reducing experimental screening burden for SARS-CoV-2 antibody optimization, while its scientific value lies in demonstrating that large protein LLMs can be specialized into binding-aware sequence encoders for a fixed antigenic target (Ashraf et al., 4 Mar 2026). A plausible broader implication is that future antibody-affinity systems may increasingly combine large pretrained sequence models with tightly scoped, target-specific supervision rather than relying only on universal architectures or antigen-agnostic training.