BeeTLe: An Imbalance-Aware Deep Sequence Model for Linear B-Cell Epitope Prediction and Classification with Logit-Adjusted Losses

The process of identifying and characterizing B-cell epitopes, which are the portions of antigens recognized by antibodies, is important for our understanding of the immune system, and for many applications including vaccine development, therapeutics, and diagnostics. Computational epitope prediction is challenging yet rewarding as it significantly reduces the time and cost of laboratory work. Most of the existing tools do not have satisfactory performance and only discriminate epitopes from non-epitopes. This paper presents a new deep learning-based multi-task framework for linear B-cell epitope prediction as well as antibody type-specific epitope classification. Specifically, a sequenced-based neural network model using recurrent layers and Transformer blocks is developed. We propose an amino acid encoding method based on eigen decomposition to help the model learn the representations of epitopes. We introduce modifications to standard cross-entropy loss functions by extending a logit adjustment technique to cope with the class imbalance. Experimental results on data curated from the largest public epitope database demonstrate the validity of the proposed methods and the superior performance compared to competing ones.

💡 Research Summary

BeeTLe introduces a unified deep‑learning framework for linear B‑cell epitope prediction and antibody (Ig) type classification. The authors first encode each amino acid using a biologically informed embedding derived from the BLOSUM62 substitution matrix. By exponentiating BLOSUM62, performing eigen‑value decomposition (B = U Σ Uᵀ), and constructing the embedding matrix E = U √Σ, each residue is represented as a 20‑dimensional vector whose dot product reflects substitution similarity. Unknown residues receive an orthogonal one‑hot vector and padding tokens are zero‑vectors.

The model processes variable‑length peptide sequences through a bidirectional LSTM (BiLSTM) to capture forward and backward context. The BiLSTM output is then fed into two stacked Transformer encoder blocks that omit positional encodings (the BiLSTM already supplies order information). Each block consists of multi‑head self‑attention, a position‑wise feed‑forward network, residual connections, and layer normalization. After the Transformer layers, a scaled dot‑product attention pooling layer aggregates residue‑level representations into a single peptide‑level vector p. This vector is passed to two parallel fully‑connected heads: one for binary epitope vs. non‑epitope prediction (sigmoid output) and another for multi‑class Ig‑type classification (softmax over IgA, IgE, IgM, etc.). ReLU activations are used in the feed‑forward layers.

To address the severe class imbalance inherent in epitope datasets (few epitopes, many non‑epitopes, and uneven Ig‑type distribution), BeeTLe adopts a logit‑adjusted loss. For each class c with prior probability π_c, the logits are shifted by τ·log π_c (τ is a temperature hyper‑parameter) before applying the standard cross‑entropy. This adjustment effectively re‑weights the loss toward minority classes, aiming to minimize balanced error rather than raw misclassification rate. The same formulation is applied to both the binary and multiclass heads, providing a unified solution.

The authors curated over 120,000 peptide sequences from the Immune Epitope Database (IEDB), removed redundancy at a 90 % similarity threshold, and split the data into training/validation/test sets (8:1:1). Evaluation metrics include AUC, accuracy, and F1‑score. BeeTLe achieves an AUC of 86 % for epitope detection, surpassing previous state‑of‑the‑art methods (which typically reach ~80 % AUC) and improves overall accuracy by 6 percentage points. Ablation studies demonstrate that (i) replacing the BLOSUM‑based embedding with one‑hot reduces performance by ~3 %, (ii) omitting the logit‑adjusted loss degrades minority‑class F1 by ~5 %, and (iii) removing the Transformer blocks lowers overall scores by ~4 %. Multi‑task training also yields better generalization than training each task separately.

Key contributions are: (1) a simple yet biologically grounded amino‑acid embedding that avoids the need for large pre‑trained protein language models; (2) a lightweight Transformer architecture tailored to short peptide sequences, offering superior representation without excessive computational cost; (3) the first application of logit‑adjusted loss to epitope prediction, effectively handling both binary and multiclass imbalance; and (4) a unified multi‑task framework that simultaneously predicts epitope presence and Ig‑type, facilitating downstream vaccine and diagnostic design.

Future directions suggested include integrating large‑scale protein language model embeddings, incorporating structural information from predicted 3D models (e.g., AlphaFold), and extending the approach to other immunological tasks such as T‑cell epitope prediction. BeeTLe’s design makes it accessible to laboratories with modest computational resources while delivering state‑of‑the‑art performance, positioning it as a practical tool for immunoinformatics research.