ABConformer: Sequence-based Ab–Ag Interface Predictor

• ABConformer is a sequence-based model that predicts antibody-antigen interface contacts with a physics-inspired sliding attention mechanism, inferring residue-level interactions without 3D structures.
• The model uses parallel branches for heavy-chain, light-chain, and antigen sequences, integrating a Conformer backbone to capture both local and global biosequence features.
• Evaluations on SARS-CoV-2 datasets demonstrate state-of-the-art performance, with ablation studies highlighting the sliding attention module as crucial for improving precise interface prediction.

ABConformer is a sequence-based model for antibody-antigen interface prediction that was introduced to predict paratopes, epitopes, and pan-epitopes without relying on three-dimensional structural data. In "ABConformer: Physics-inspired Sliding Attention for Antibody-Antigen Interface Prediction" (You et al., 27 Sep 2025), the model is described as being based on the Conformer backbone, with a physics-inspired sliding attention mechanism designed to recover residue-level contacts from antibody and antigen sequences alone. The central objective is accurate Ab-Ag interface prediction for vaccine design, immunodiagnostics, and therapeutic antibody development, with particular emphasis on reliable residue-level inference from sequence inputs.

1. Conceptual scope and prediction tasks

ABConformer addresses three closely related prediction settings. It can predict paratopes and epitopes given the antibody and antigen sequence, and it can also predict pan-epitopes on the antigen without antibody information (You et al., 27 Sep 2025). This positions the model within sequence-based immunoinformatics rather than structure-dependent docking or interface analysis.

The model is framed around the claim that accurate sequence-only Ab-Ag interface prediction remains challenging, especially when residue-level contacts must be inferred without explicit structural supervision beyond interface labels. The proposed solution combines a Conformer encoder, which captures both local and global biosequence features, with a sliding attention mechanism intended to model physically plausible interaction locality (You et al., 27 Sep 2025).

A common source of confusion is nomenclature. The label “ABCONFORMER” also appears in the supplied literature as an alias for ABCD-Conformal, a method for approximate Bayesian computation with deep learning and conformal prediction (Baragatti et al., 2024), and as a label applied in the technical summary of BConformeR, a conformer-based model for unified prediction of continuous and discontinuous B-cell epitopes from antigen sequences (You et al., 16 Aug 2025). In the present usage, however, ABConformer most directly denotes the antibody-antigen interface predictor introduced in (You et al., 27 Sep 2025).

2. Backbone architecture and representation flow

The overall architecture comprises three parallel branches that process heavy-chain (Ab-H), light-chain (Ab-L), and antigen (Ag) separately (You et al., 27 Sep 2025). Each branch begins with an encoding layer. In the antigen branch, the raw sequence is embedded with ESM-2 150M with dimension $d=640$, followed by a two-layer feed-forward network with residual connection and a multi-head self-attention block. The heavy and light branches omit the MHSA block, because empirical ablation showed that it contributes little to paratope accuracy when sliding attention is present (You et al., 27 Sep 2025).

After encoding, the antigen representation interleaves Conformer layers with sliding-attention modules. The Conformer layer design is attributed to Gulati et al. 2020 and combines local and global processing: local patterns are captured by a 1D convolution with kernel size $k=5$ and pointwise feed-forward sublayers, while global long-range dependencies are modeled by MHSA over the entire sequence (You et al., 27 Sep 2025). Residual connections and layer normalization merge these paths into a single representation intended to preserve both short-range motif information and sequence-wide dependencies.

A total of 6 Conformer+sliding layers are stacked by default, explicitly as a trade-off between depth and compute (You et al., 27 Sep 2025). After $T$ sliding steps, antigen embeddings obtained by sliding against Ab-H and Ab-L are fused as

$$X_{\rm Ag} = \alpha\,X_H^{(T)} + (1-\alpha)\,X_L^{(T)}, \quad \alpha\in[0,1].$$

The fused antigen embedding then passes through the remaining Conformer layers and a final convolution plus feed-forward module to produce per-residue interface scores (You et al., 27 Sep 2025). This suggests that ABConformer treats heavy-chain and light-chain interaction evidence as separable but combinable sources of contextualization for antigen-interface inference.

3. Physics-inspired sliding attention

The defining mechanism of ABConformer is the physics-inspired sliding attention inserted between Conformer blocks to model Ab-Ag interactions without 3D structure (You et al., 27 Sep 2025). The mechanism is formulated between a “sliding” sequence $X^{(t)}=\{x_i^{(t)}\}_{i=1}^m$ and a fixed “reference” sequence $Y^{(t)}=\{y_j^{(t)}\}_{j=1}^n$, both $d$-dimensional.

The first component is feature-based attention:

$$a_{ij}^{(t)} = \frac{(x_i^{(t)}E_S)\cdot(y_j^{(t)}E_R)^\top}{\sqrt d}, \qquad A_{ij}^{(t)} = \exp\!\bigl(a_{ij}^{(t)}-\max_{k}a_{ik}^{(t)}\bigr),$$

where $E_S,E_R\in\mathbb{R}^{d\times d}$ are learnable projections (You et al., 27 Sep 2025). The second component is spatial proximity attention:

$$S_{ij}^{(t)} = \exp\!\Bigl(-\,\frac{\bigl(p_i^{(t)}-q_j\bigr)^2}{2h^2}\Bigr),$$

with $k=5$0 the current learnable position of sliding residue $k=5$1, $k=5$2 the fixed integer position of reference residue $k=5$3, and bandwidth

$k=5$4

Here $k=5$5 masks padding, and $k=5$6 are hyperparameters (You et al., 27 Sep 2025).

The combined attention weights are defined as

$k=5$7

with row-normalized and column-normalized forms

$k=5$8

The iterative embedding and position updates are

$k=5$9

and

$T$0

Equivalently, the positional update for residue $T$1 takes the mean-shift style form

$T$2

The paper distinguishes this mechanism from conventional cross-attention in two explicit ways: it softly enforces locality via the Gaussian kernel $T$3, and it updates positions $T$4, simulating a physical docking process so that successive iterations focus the attention window on the true contact region (You et al., 27 Sep 2025). A plausible implication is that the method is designed not merely to align embeddings but to impose a structured inductive bias on which residue pairs can interact.

4. Supervision, optimization, and experimental setup

Each chain, namely Ab-H, Ab-L, and Ag, produces a per-residue probability $T$5 over $T$6 via a final softmax (You et al., 27 Sep 2025). With valid-position mask $T$7 and one-hot target $T$8, the per-chain cross-entropy is

$T$9

The total loss is the average across heavy chain, light chain, and antigen:

$$X_{\rm Ag} = \alpha\,X_H^{(T)} + (1-\alpha)\,X_L^{(T)}, \quad \alpha\in[0,1].$$0

No extra auxiliary objectives were used, beyond standard weight decay; optimization uses AdamW with gradient clipping $$X_{\rm Ag} = \alpha\,X_H^{(T)} + (1-\alpha)\,X_L^{(T)}, \quad \alpha\in[0,1].$$1, mixed-precision, and an EMA of weights for stable evaluation (You et al., 27 Sep 2025).

The training dataset is AACDB, used with 5-fold cross-validation: 7,488 PDB complexes filtered to 3,674 unique Ab-Ag pairs, with antigens clustered into six phylogenetic groups that are evenly split into five folds (You et al., 27 Sep 2025). Average lengths are given as Ab-H $$X_{\rm Ag} = \alpha\,X_H^{(T)} + (1-\alpha)\,X_L^{(T)}, \quad \alpha\in[0,1].$$2, Ab-L $$X_{\rm Ag} = \alpha\,X_H^{(T)} + (1-\alpha)\,X_L^{(T)}, \quad \alpha\in[0,1].$$3, and Ag $$X_{\rm Ag} = \alpha\,X_H^{(T)} + (1-\alpha)\,X_L^{(T)}, \quad \alpha\in[0,1].$$4, with interface rates $$X_{\rm Ag} = \alpha\,X_H^{(T)} + (1-\alpha)\,X_L^{(T)}, \quad \alpha\in[0,1].$$5 (You et al., 27 Sep 2025). A held-out SARS-CoV-2 test set contains 35 PDB structures since 2024 from CoV-AbDab, yielding 46 Ab-H/Ab-L+Ag entries with no overlap to AACDB (You et al., 27 Sep 2025).

Preprocessing extracts Ab-H, Ab-L, and Ag chains; each residue is embedded with ESM-2 in 640 dimensions; interfaces are labeled by heavy-atom distance $$X_{\rm Ag} = \alpha\,X_H^{(T)} + (1-\alpha)\,X_L^{(T)}, \quad \alpha\in[0,1].$$6 Å (You et al., 27 Sep 2025). Evaluation uses binary metrics—IoU, Precision, Recall, F1, and MCC—and continuous metrics—Pearson correlation coefficient, ROC-AUC, and PR-AUC—computed per chain and averaged (You et al., 27 Sep 2025).

5. Reported performance and comparative results

On the SARS-CoV-2 held-out test set, ABConformer is reported to achieve state-of-the-art performance on this recent dataset and to surpass widely used sequence-based methods for antibody-agnostic epitope prediction (You et al., 27 Sep 2025). The reported thresholds are $$X_{\rm Ag} = \alpha\,X_H^{(T)} + (1-\alpha)\,X_L^{(T)}, \quad \alpha\in[0,1].$$7 for Ab-H/Ab-L/Ag.

For paratope prediction on Ab-H, the model attains IoU $$X_{\rm Ag} = \alpha\,X_H^{(T)} + (1-\alpha)\,X_L^{(T)}, \quad \alpha\in[0,1].$$8 versus MIPE $$X_{\rm Ag} = \alpha\,X_H^{(T)} + (1-\alpha)\,X_L^{(T)}, \quad \alpha\in[0,1].$$9, F1 $X^{(t)}=\{x_i^{(t)}\}_{i=1}^m$0 versus $X^{(t)}=\{x_i^{(t)}\}_{i=1}^m$1, and PR-AUC $X^{(t)}=\{x_i^{(t)}\}_{i=1}^m$2 versus $X^{(t)}=\{x_i^{(t)}\}_{i=1}^m$3 (You et al., 27 Sep 2025). For epitope prediction in the Ab-Ag setting, it attains IoU $X^{(t)}=\{x_i^{(t)}\}_{i=1}^m$4 versus MIPE $X^{(t)}=\{x_i^{(t)}\}_{i=1}^m$5, F1 $X^{(t)}=\{x_i^{(t)}\}_{i=1}^m$6 versus $X^{(t)}=\{x_i^{(t)}\}_{i=1}^m$7, and PR-AUC $X^{(t)}=\{x_i^{(t)}\}_{i=1}^m$8 versus $X^{(t)}=\{x_i^{(t)}\}_{i=1}^m$9 (You et al., 27 Sep 2025). For pan-epitope prediction on Ag only, the reported IoU is $Y^{(t)}=\{y_j^{(t)}\}_{j=1}^n$0, which is said to rival structure-based DiscoTope-3.0 at $Y^{(t)}=\{y_j^{(t)}\}_{j=1}^n$1 despite using sequence alone; the corresponding F1 is $Y^{(t)}=\{y_j^{(t)}\}_{j=1}^n$2 versus $Y^{(t)}=\{y_j^{(t)}\}_{j=1}^n$3, and PR-AUC is $Y^{(t)}=\{y_j^{(t)}\}_{j=1}^n$4 versus $Y^{(t)}=\{y_j^{(t)}\}_{j=1}^n$5 (You et al., 27 Sep 2025).

These comparisons support two narrower claims made in the paper. First, the model surpasses prior sequence-based methods on the reported SARS-CoV-2 benchmark (You et al., 27 Sep 2025). Second, in the antibody-agnostic setting, its performance is brought close to structure-aware baselines, though not exceeding DiscoTope-3.0 on the specific pan-epitope metrics listed above (You et al., 27 Sep 2025). This is an important distinction, because the paper does not claim universal superiority over structure-based methods.

6. Ablation findings and interpretive significance

Ablation studies on AACDB isolate the contribution of individual components (You et al., 27 Sep 2025). In the encoding ablation, ESM-2 embeddings are compared with one-hot plus context window. The ESM-2 representation reduces input dimension from $Y^{(t)}=\{y_j^{(t)}\}_{j=1}^n$6 to $Y^{(t)}=\{y_j^{(t)}\}_{j=1}^n$7 and improves Ag-epitope precision from $Y^{(t)}=\{y_j^{(t)}\}_{j=1}^n$8 to $Y^{(t)}=\{y_j^{(t)}\}_{j=1}^n$9 and F1 from $d$0 to $d$1 (You et al., 27 Sep 2025).

The most consequential ablation concerns stage III attention. The full sliding-attention model yields Ag-epitope precision $d$2, recall $d$3, and PCC $d$4. Cross-attention yields precision $d$5 and recall $d$6, while self-attention without inter-chain interaction yields precision $d$7 and recall $d$8 (You et al., 27 Sep 2025). The paper explicitly highlights a $d$9 precision gain of sliding attention over cross-attention, interpreting this as confirmation that sliding attention is the key component driving accurate contact recovery (You et al., 27 Sep 2025).

A separate ablation removes either the convolutional path or MHSA from the Conformer backbone. Each removal costs approximately $$a_{ij}^{(t)} = \frac{(x_i^{(t)}E_S)\cdot(y_j^{(t)}E_R)^\top}{\sqrt d}, \qquad A_{ij}^{(t)} = \exp\!\bigl(a_{ij}^{(t)}-\max_{k}a_{ik}^{(t)}\bigr),$$0–$$a_{ij}^{(t)} = \frac{(x_i^{(t)}E_S)\cdot(y_j^{(t)}E_R)^\top}{\sqrt d}, \qquad A_{ij}^{(t)} = \exp\!\bigl(a_{ij}^{(t)}-\max_{k}a_{ik}^{(t)}\bigr),$$1 points in F1 on paratopes and epitopes, and the full Conformer backbone with both convolution and MHSA is selected for the best overall balance (You et al., 27 Sep 2025). Hyperparameter sensitivity analysis identifies sliding steps $$a_{ij}^{(t)} = \frac{(x_i^{(t)}E_S)\cdot(y_j^{(t)}E_R)^\top}{\sqrt d}, \qquad A_{ij}^{(t)} = \exp\!\bigl(a_{ij}^{(t)}-\max_{k}a_{ik}^{(t)}\bigr),$$2, bandwidth $$a_{ij}^{(t)} = \frac{(x_i^{(t)}E_S)\cdot(y_j^{(t)}E_R)^\top}{\sqrt d}, \qquad A_{ij}^{(t)} = \exp\!\bigl(a_{ij}^{(t)}-\max_{k}a_{ik}^{(t)}\bigr),$$3, and scale $$a_{ij}^{(t)} = \frac{(x_i^{(t)}E_S)\cdot(y_j^{(t)}E_R)^\top}{\sqrt d}, \qquad A_{ij}^{(t)} = \exp\!\bigl(a_{ij}^{(t)}-\max_{k}a_{ik}^{(t)}\bigr),$$4 as robust settings; increasing $$a_{ij}^{(t)} = \frac{(x_i^{(t)}E_S)\cdot(y_j^{(t)}E_R)^\top}{\sqrt d}, \qquad A_{ij}^{(t)} = \exp\!\bigl(a_{ij}^{(t)}-\max_{k}a_{ik}^{(t)}\bigr),$$5 improves precision at slight recall cost, while smaller $$a_{ij}^{(t)} = \frac{(x_i^{(t)}E_S)\cdot(y_j^{(t)}E_R)^\top}{\sqrt d}, \qquad A_{ij}^{(t)} = \exp\!\bigl(a_{ij}^{(t)}-\max_{k}a_{ik}^{(t)}\bigr),$$6 tightens the attention window and further boosts precision (You et al., 27 Sep 2025).

Taken together, these observations support the paper’s specific mechanistic interpretation: the Conformer backbone supplies the local/global sequence encoder, whereas sliding attention supplies the residue-contact inductive bias that most strongly improves epitope precision (You et al., 27 Sep 2025). This suggests that ABConformer should be understood less as a generic Conformer variant than as a coupled architecture whose defining contribution is the iterative, locality-constrained interaction module.

7. Related usages of the name and adjacent conformer-based methods

The name “ABCONFORMER” is not unique in the provided corpus. In "Approximate Bayesian Computation with Deep Learning and Conformal prediction" (Baragatti et al., 2024), the technical summary refers to ABCD-Conformal as “ABCONFORMER.” That method is unrelated to antibody-antigen interface prediction: it removes summary statistics, distance, and tolerance threshold from ABC workflows, uses a neural network with Monte Carlo Dropout to estimate $$a_{ij}^{(t)} = \frac{(x_i^{(t)}E_S)\cdot(y_j^{(t)}E_R)^\top}{\sqrt d}, \qquad A_{ij}^{(t)} = \exp\!\bigl(a_{ij}^{(t)}-\max_{k}a_{ik}^{(t)}\bigr),$$7 and $$a_{ij}^{(t)} = \frac{(x_i^{(t)}E_S)\cdot(y_j^{(t)}E_R)^\top}{\sqrt d}, \qquad A_{ij}^{(t)} = \exp\!\bigl(a_{ij}^{(t)}-\max_{k}a_{ik}^{(t)}\bigr),$$8, and applies split-conformal calibration to obtain confidence sets with frequentist marginal coverage (Baragatti et al., 2024). Its domain is simulator-based inference, not biosequence interface modeling.

A second nearby usage appears in "BConformeR: A Conformer Based on Mutual Sampling for Unified Prediction of Continuous and Discontinuous Antibody Binding Sites" (You et al., 16 Aug 2025). The supplied summary labels that antigen-sequence epitope predictor as ABCONFORMER, but the manuscript title itself uses BConformeR. That model combines a CNN branch for local feature extraction with a Transformer branch for long-range dependencies, linked by Feature Coupling Units and trained for unified prediction of linear and discontinuous B-cell epitopes (You et al., 16 Aug 2025).

These homonymous usages are methodologically distinct. ABConformer in (You et al., 27 Sep 2025) is characterized by three parallel Ab-H/Ab-L/Ag branches, a Conformer backbone, and physics-inspired sliding attention for residue-level contact recovery from sequence alone. ABCD-Conformal in (Baragatti et al., 2024) belongs to likelihood-free Bayesian inference, and BConformeR in (You et al., 16 Aug 2025) concerns antigen-only epitope prediction through CNN-Transformer mutual sampling. The overlap is therefore nominal rather than conceptual.