Optimal subset selection can accurately and efficiently mine important information from high-dimensional data to build a reduced regression model. In recent years, it has been used more and more in machine learning, image processing and biomedicine. However, most of the existing optimal subset selection methods are based on least squares or maximum likelihood, resulting in insufficient robustness when dealing with heterogeneous data. In order to effectively deal with the heterogeneity of high-dimensional data and comprehensively analyze the conditional distribution of response variable, a robust optimal subset selection algorithm based on ${\ell }_0$ penalty and smoothing quantile loss function is designed in this paper. In practice, the real number of important variables is usually unknown. In this paper, a truncated sequential search algorithm is proposed for efficient and accurate selection of the number of important variables. In the simulation, comparing with the existing variable selection methods, it is found that the proposed method has more advantages in variable selection and parameter estimation accuracy. Finally, the new method is used to analyze the gene data related to the production of riboflavin by Bacillus subtilis. The experimental results show that the new method has a smaller quantile prediction error in estimating the riboflavin production rate.