Kapitan Vitalii, Choi Michael
Department of Statistics and Data Science, National University of Singapore, 6 Science Drive 2, Singapore, Singapore.
BMC Bioinformatics. 2025 Jul 1;26(1):161. doi: 10.1186/s12859-025-06185-2.
Protein structure prediction is one of the most important scientific problems, on the one hand, it is one of the NP-hard problems, and on the other hand, it has a wide range of applications including drug discovery and biotechnology development. Since experimental methods for structure determination remain expensive and time-consuming, computational structure prediction offers a scalable and cost-effective alternative and application of machine learning in structural biology has revolutionized protein structure prediction. Despite their success, machine learning methods face fundamental limitations in optimizing complex high-dimensional energy landscapes, which motivates research into new methods to improve the robustness and performance of optimization algorithms.
This study presents a novel approach to protein structure prediction by integrating the Landscape Modification (LM) method with the Adam optimizer for OpenFold. The main idea is to change the optimization dynamics by introducing a gradient scaling mechanism based on energy landscape transformations. LM dynamically adjusts gradients using a threshold parameter and a transformation function, thereby improving the optimizer's ability to avoid local minima, more efficiently traverse flat or rough landscape regions, and potentially converge faster to global or high-quality local optima. By integrating simulated annealing into the LM approach, we propose LM SA, a variant designed to improve convergence stability while facilitating more efficient exploration of complex landscapes.
We compare the performance of standard Adam, LM, and LM SA on different datasets and computational conditions. Performance was evaluated using Loss function values, predicted Local Distance Difference Test (pLDDT), distance-based Root Mean Square Deviation (dRMSD), and Template Modeling (TM) scores. Our results show that LM and LM SA outperform the standard Adam across all metrics, showing faster convergence and better generalization, particularly on proteins not included in the training set. These results demonstrate that integrating landscape-aware gradient scaling into first-order optimizers advances research in computational optimization and improves prediction performance for complex problems such as protein folding.
蛋白质结构预测是最重要的科学问题之一，一方面，它是NP难问题之一，另一方面，它有广泛的应用，包括药物发现和生物技术发展。由于用于结构测定的实验方法仍然昂贵且耗时，计算结构预测提供了一种可扩展且具有成本效益的替代方法，机器学习在结构生物学中的应用彻底改变了蛋白质结构预测。尽管取得了成功，但机器学习方法在优化复杂的高维能量景观方面面临着根本限制，这促使人们研究新方法以提高优化算法的鲁棒性和性能。
本研究提出了一种新的蛋白质结构预测方法，即将景观修饰（LM）方法与用于OpenFold的Adam优化器相结合。主要思想是通过引入基于能量景观变换的梯度缩放机制来改变优化动态过程。LM使用阈值参数和变换函数动态调整梯度，从而提高优化器避免局部最小值的能力，更有效地遍历平坦或崎岖的景观区域，并可能更快地收敛到全局或高质量的局部最优解。通过将模拟退火集成到LM方法中，我们提出了LM SA，这是一种旨在提高收敛稳定性同时促进对复杂景观更有效探索的变体。
我们在不同数据集和计算条件下比较了标准Adam、LM和LM SA的性能。使用损失函数值、预测局部距离差异测试（pLDDT）、基于距离的均方根偏差（dRMSD）和模板建模（TM）分数来评估性能。我们的结果表明，在所有指标上，LM和LM SA均优于标准Adam，显示出更快的收敛速度和更好的泛化能力，特别是在训练集中未包含的蛋白质上。这些结果表明，将基于景观感知的梯度缩放集成到一阶优化器中推动了计算优化研究，并提高了对蛋白质折叠等复杂问题的预测性能。
