A pivotal challenge in computational neuroscience is identifying sub-populations of neurons that encode behavior, perception, and memory, across varying contexts such as task difficulty or temporal progression. Therefore, analysis tools that provide model interpretability, beyond prediction performance, are essential for studying neuronal population activity. We propose a novel approach conditional-STochastic Gates (c-STG) that combines predictive modeling with contextual feature selection. c-STG comprises two networks: a hypernetwork that maps contextual variables to probabilistic feature selection parameters, and a prediction network, which can be linear or nonlinear, that maps the selected features to the response variable, e.g., trial outcome. c-STG accommodates categorical (e.g., discrete trial condition), continuous (e.g., time within trial), and/or multidimensional contextual variables. Our results show that c-STG outperforms population-wise, instance-wise, and context-specific feature selection techniques in terms of feature selection, prediction performance, and interpretability. Applied to calcium imaging data recorded from pyramidal neurons in the primary motor cortex in mice performing a hand-reach task for a food pellet, a single c-STG model consistently decoded trial outcome from neural activity starting 2 seconds after the cue till the end of the trial and identified that 12\% of neurons encoded trial outcome. In contrast, prior work required training thousands of conventional models to arrive at these conclusions. In a similar experiment where mice received sucrose, quinine or unflavored pellets, a single c-STG model identified sub-populations of neurons encoding outcomes based on flavor. We show that the encodings of sucrose and quinine flavors were most different, while the encoding of unflavored pellets was more similar to sucrose than to quinine. Overall, our findings demonstrate that c-STG is an interpretable tool for analyzing neuronal population activity and can provide novel insights into how complex signals are encoded by brain activity.