Integrating sequence and chemical information via a transformer-based framework to predict anticancer peptides

Anticancer peptides (ACPs) have emerged as promising therapeutic agents due to their high selectivity for malignant cells and favorable safety profiles. However, experimental identification of ACPs remains costly and low-throughput, limiting large-scale discovery. To address these challenges, we present ACP-SCI, a novel computational framework that enhances ACP prediction by integrating sequential and Chemical information. The model employs a dual-stream architecture: the sequence subnetwork utilizes a pre-trained protein language model to generate context-aware residue embeddings, which are processed through a Transformer encoder to capture local and long-range residue dependencies. Simultaneously, The chemical branch encodes RDKit-generated peptide SMILES using ChemBERTa to capture molecular-connectivity and chemical-token patterns. These complementary representations are fused for binary ACP classification. Embeddings from both streams are fused into a unified representation and passed through a fully connected layer for binary classification. Benchmark evaluations demonstrate that ACP-SCI provides competitive performance on the evaluated ACP prediction benchmarks. These findings indicate that SMILES-derived chemical representations can complement protein language model-based sequence embeddings by providing additional molecular-connectivity information. The source code and datasets are made publicly available to facilitate reproducibility and future research.