Bio-motif detection is one of essential computational tasks for bioinformatics and genomics. Based on a theoretical framework for quantitatively modeling the relationship of convolution kernel shape and the motif detection effectiveness, we design and propose a novel convolution-based model, VCNN (Variable CNN), for effective bio-motif detection via the adaptive kernel length at runtime. Empirical evaluations based on both simulated and real-world genomics data demonstrate VCNN's superior performance to classical CNN in both detection power and hyper-parameter robustness. All source code and data are available at https://github.com/gao-lab/VCNN/ freely for academic usage.