LG AI Research Develops HLA-Peptide Binding Prediction AI Model for Personalized Cancer Vaccine Development

Third-generation immuno-oncology drugs, which emerged to address the toxicity and side-effect issues of existing anticancer drugs, induce immune cells to selectively attack only cancer cells. The binding between HLA (Human Leukocyte Antigen) and neoepitope (peptide) plays a key role in this process.

LG AI Research's Material Intelligence (MI) Lab has developed a new AI model that predicts the binding affinity between HLA and peptides. This model was presented at the MLSB Workshop at NeurIPS 2022, and demonstrated superior performance compared to existing simple concatenation or cross attention approaches.

The key points of the Conditional Attention mechanism proposed by the research team are as follows:

• Structural characteristics: After obtaining an information representation through the peptide sequence, it is used as the Query, while the HLA sequence is used as the Key and Value, allowing the model to naturally learn which region of the HLA the peptide binds to.
• Various embedding experiments: Three embedding methods were compared and analyzed—Learned Embedding (LE), Amino Acid Index (AA), and the ESM Pretrained Model.
• Pooling method optimization: Optimal performance was derived through the [CLS] token, Mean pooling, and the Learned weighting method using learnable weights.

This research integrated data and source code to overcome the limitations of fragmented existing datasets and source code, and secured objectivity by setting up various amino acid feature representations and pooling methods for a fairer performance comparison.