Understanding protein secondary structure is crucial for analyzing their functions and designing drugs. Accurately predicting its eight states is both vital and challenging. Despite numerous proposed methods, research on representing proteins from multiple views and features remains limited. Moreover, effectively selecting and fusing this information remains an unresolved issue, which hampers prediction accuracy. In this paper, the Multi-View Multi-Feature Evolutionary Deep Fusion (MVMF-EDF) method, which consists of two stages, is proposed to address this challenge. First, a multi-view multi-feature (MVMF) strategy is introduced to construct protein representations from various views and features. Notably, MVMF-EDF incorporates semantic information from a pre-trained language model to enhance protein representation. Second, an evolutionary deep fusion (EDF) approach is employed to automatically select and fuse the extracted features. Experimental results on four benchmark datasets show that MVMF-EDF outperforms baseline models in predicting protein secondary structure. This approach leverages deep neural networks for feature extraction and evolutionary algorithms for feature integration, achieving high accuracy without requiring manual feature fusion design.

This paper studies a multi-stage multi-product production and inventory planning problem with random yield derived from the cold rolling process in the steel industry. The cold rolling process has multiple stages, and intermediate inventory buffers are kept between stages to ensure continuous operation. Switching products during the cold rolling process is typically very costly. Backorder costs are incurred for unsatisfied demand while inventory holding costs are incurred for excess inventory. The process also experiences random yield. The objective of the production and inventory planning problem is to minimize the total cost including the switching costs, inventory holding costs, and backorder costs. We propose a stochastic formulation with a nonlinear objective function. Two lower bounds are proposed, which are based on full information relaxation and Jensen’s inequality, respectively. Then, we develop two heuristics from the proposed lower bounds. In addition, we propose a two-stage procedure motivated by newsvendor logic. To verify the performance of the proposed bounds and heuristics, computational tests are conducted on synthetic instances. The results show the efficiency of the proposed bounds and heuristics.