Optimization of steel frame structures is typically formulated as a large-scale combinatorial problem. Previous research predominantly employs metaheuristic algorithms, which frequently face challenges such as high computational costs, sensitivity to hyperparameters, and reliance on initial solutions. To overcome these limitations, this study proposes a novel optimization workflow termed DeepMonte-Frame, integrating Monte Carlo Tree Search (MCTS) and Feedforward Neural Networks (FNNs). The FNNs rapidly predict structural responses, enhancing both the expansion and rollout phases of the MCTS, thereby significantly improving optimization performance in scenarios with sparse feasible solutions. Ablation experiments demonstrated the essential contribution of FNNs, while comparative evaluations against metaheuristic algorithms proved the superior performance of DeepMonte-Frame. Moreover, the method maintains robust performance when applied to irregular structural configurations, and exhibits remarkable flexibility in various optimization objectives, including cost and carbon footprint reduction. A comprehensive case study further validated the practical applicability of DeepMonte-Frame, achieving a 15.45 % cost reduction while ensuring compliance with engineering standards. The optimized designs can also be automatically transformed into BIM models, facilitating design decision-making and supporting subsequent interdisciplinary collaboration. Overall, DeepMonte-Frame is a highly effective and adaptable approach, significantly outperforming conventional methods and providing innovative insights for future research in structural optimization.