ProtThermoRank: Sequence-based multi-class prediction model for protein thermal stability changes and its application...

Protein thermal stability is a key property of industrial enzymes and biopharmaceutical proteins. The rapid and accurate evaluation of thermal stability across large numbers of variants has become a critical step in protein engineering. However, experimental screening of large mutational landscapes remains time-consuming, labor-intensive, and costly. Here, we present ProtThermoRank, a sequence-based multiclass prediction model for mutation-induced changes in protein thermal stability. Unlike existing approaches that encode reference and mutant sequences independently, ProtThermoRank explicitly modeled their relationship through a cross-attention mechanism. The change in melting temperature (ΔT_m) was classified into four classes, enabling fine-grained discrimination of protein thermal stability changes. ProtThermoRank achieved the highest class-specific recall for the ΔT_m > 10 °C class, representing a three-fold increase in class-specific recall than the previous best-performing model and facilitating the identification of highly thermostabilizing variants. To demonstrate the applicability of ProtThermoRank, we performed in-silico saturation mutagenesis of clostridial cellulose synthase subunit Z (CcsZ) and identified nine candidate thermostabilizing variants using ProtThermoRank. Among them, three candidate variants (D121G, D121V, and Y170T) were experimentally confirmed to exhibit enhanced thermal stability, and molecular dynamics simulations provided structural insights into the enhanced thermal stability of these variants. These results demonstrated that ProtThermoRank provides a sequence-based approach for artificial intelligence-guided protein engineering, enabling in-silico high-throughput screening of thermostabilizing variants and accelerating the discovery and optimization of highly stabilizing proteins.