Emotion is an important component of music investigated in music psychology. In recent years, the use of computational methods to assess the link between music and emotions has been promoted by advances in music emotion recognition. However, one of the main limitations of applying data-driven approaches to understand such a link is the scarce knowledge of how perceived music emotions might be inferred from automatically retrieved features. Through statistical analysis we investigate the relationship between perceived music emotions (rated by 41 listeners in terms of categories and dimensions) and multi-modal acoustic and symbolic features (automatically extracted from the audio and MIDI files of 24 pieces) in piano repertoire. We also assess the suitability of the identified features for music emotion recognition. Our results highlight the potential of assessing perception and data-driven methods in a unified framework.