* Please select a metal ion model and upload a FASTA file containing a single protein sequence. The system will predict metal ion binding sites for the selected model. Do not submit files larger than 50Mb. Click to download test dataset.
To enhance model robustness, we adopted the larger parameter configuration from the Hugging Face T5-large model. Finally, for the downstream classifier, we incorporated the Swish activation function and applied both L1 and L2 regularization to mitigate overfitting. To facilitate a comparison between ProtT5 and other classical models, we constructed a benchmark dataset using ion-specific data from multiple sources. Data for Ca²⁺, Mg²⁺, Zn²⁺, Fe²⁺, Fe³⁺, K⁺, Na⁺, and Mn²⁺ were sourced from IonCom, while data for Co²⁺, Cu²⁺, Ni²⁺, and Fe³⁺ were obtained from MetalNet.