ProtT5-MSCRNet Deep Learning Framework Achieves >95% Accuracy in Anticancer Peptide Prediction

Despite their promise as selective, low-toxicity therapeutics, anticancer peptides (ACPs) face significant hurdles in discovery. Traditional experimental screening is labor-intensive, time-consuming, and costly, limiting the pace of drug development. This creates a critical gap for efficient, high-throughput computational methods to identify potential ACP candidates from vast sequence databases, thereby accelerating the pipeline for novel cancer treatments. The challenge lies in accurately predicting peptide activity based solely on sequence data, a complex task given the diverse mechanisms of action.

Researchers developed ProtT5-MSCRNet, an end-to-end deep learning framework designed for accurate anticancer peptide prediction. The model integrates several advanced computational techniques: ProtT5-based evolutionary representations for capturing sequence information, multi-scale convolutional feature extraction to identify patterns at various granularities, and channel-wise attention recalibration to emphasize important features. The framework also incorporates robust optimization strategies to enhance performance. The model's efficacy was evaluated on two independent benchmark datasets, comparing its predictive power against existing state-of-the-art ACP prediction methods.

ProtT5-MSCRNet demonstrated superior performance compared to current state-of-the-art anticancer peptide prediction methods across two independent benchmark datasets. The model achieved remarkable accuracy and specificity, indicating its strong capability to correctly identify ACPs while minimizing false positives. Specific metrics on Test Set 1 included an ACC of 0.954, SN of 0.874, SP of 0.983, and MCC of 0.881. Performance was even stronger on Test Set 2:

ProtT5-MSCRNet achieved an ACC of 0.984, SN of 0.980, SP of 0.987, and MCC of 0.967 on Test Set 2, highlighting its robust and highly accurate predictive power.

Ablation studies confirmed the contribution of each architectural component to the overall effectiveness, while visualization analyses provided insights into the model's interpretability, supporting its biological meaningfulness.