Prot2Prop: Structure-informed multitask protein property prediction

Protein engineering often relies on separate models for related developability properties, limiting efficiency and transfer across tasks. We present Prot2Prop, a multitask framework based on a frozen ProstT5 encoder with shared and task-specific adapters for joint prediction of six protein properties: material production, solubility, temperature stability, aggregation propensity, expression yield, and folding stability. Across held-out test data, Prot2Prop achieved strong performance on both classification and regression tasks, including AUROC values ranging from 0.86 to 0.98 for classification endpoints and Spearman correlations ranging from 0.73 to 0.86 for regression endpoints. The model achieved particularly strong performance for temperature stability (AUROC = 0.98) and aggregation propensity (Spearman = 0.86). Post-hoc calibration further improved regression accuracy, reducing folding stability MAE from 0.67 to 0.48. These results demonstrate that parameter-efficient multitask adaptation of protein language models can provide accurate and unified prediction of diverse protein developability properties.