Multiscale and physically consistent multitask learning for protein melting temperature prediction

Protein thermal stability determines the functional viability of proteins in industrial and biomedical applications. We developed an Evolutionary Scale Modeling Cambrian-based multiscale multitask predictor for protein melting temperature (ESMTmP). The model leverages a frozen 300-million-parameter Cambrian encoder to extract residue representations. These features are refined by residual temporal convolutions. Furthermore, global statistical pooling, local multiscale convolutions, and multi-head attention are integrated through residual gated fusion. This framework unifies adaptive representation learning with continuous temperature regression and ordered multi-threshold classification. It also introduces monotonicity regularization to encourage order-consistent probability outputs. The model was trained on 30,076 observations and evaluated on 531 exact-identity-separated held-out proteins. ESMTmP achieved a root mean square error (RMSE) of 7.42 °C and a mean absolute error of 5.64 °C. Additionally, the coefficient of determination reached 0.8112 with a Pearson correlation coefficient of 0.9014. Compared with an encoder baseline, the RMSE decreased from 7.87 to 7.42 °C. Multitask learning further reduced this error from 7.61 to 7.42 °C relative to a regression-only variant, demonstrating that auxiliary supervision effectively limits large prediction errors. Across five temperature thresholds from 40 to 80 °C, the mean area under the receiver operating characteristic curve reached 0.9462. The mean F1 score was 0.8663 with no monotonicity violations. ESMTmP provides monotonically ordered predictions to efficiently prioritize thermostable protein engineering candidates prior to experimental screening.

The online version contains supplementary material available at https://doi.org/10.1007/s13205-026-05122-1.