Answers are grounded in this paper and cite their sources.
Protein folding is a complex spatiotemporal process whose accurate modeling remains a challenging task. Physics informed neural networks (PINNs) have shown strong potential for capturing folding dynamics by integrating physical constraints with data driven learning however, their performance is highly sensitive to training configurations particularly hyperparameters, optimizers and loss functions. In this work, we propose a Bayesian Optimization with Hyperband (BOHB) based training pipeline for the PhyFold architecture to systematically identify optimal hyperparameter configurations. Building on this optimized baseline, a comparison of optimizer-loss configurations is conducted across SGD, Adagrad, RMSprop and Adam combined with MSE, Huber and Log-cosh losses evaluated on two protein folding datasets over five independent random seeds. Optimizer RMSprop and Adam, each paired with MSE loss, emerge as the jointly most consistent configurations across both datasets. On the KPro kinetics dataset under the standard split, they achieve RMSE of 0.090 ± 0.005 and 0.089 ± 0.003, MAE of 0.061 ± 0.003 and 0.058 ± 0.002 and Pearson correlation of 0.892 ± 0.012 and 0.893 ± 0.007 respectively, with comparable results on the Rocklin stability dataset. These findings confirm that optimizer-loss comparison substantially impact convergence behavior and predictive performance. Furthermore, adaptive optimizers particularly RMSprop and Adam consistently outperform SGD in terms of convergence stability and predictive accuracy. Nevertheless, the magnitude of improvement varies across datasets, indicating that configuration choices should be dataset aware. Overall, this study provides a systematic, BOHB guided hyperparameter optimization together with an empirically grounded characterization of optimizer-loss comparison for PINN based protein folding models. The code to reproduce the work is available at https://github.com/jamshaidwarraich/Intelligent-Co-Optimization-Framework.git .
No takes yet. Share an insight, caveat, or question.
Rahman et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: