Variational quantum circuits are becoming tools of choice in quantum optimization and machine learning. In this paper we investigate a class of variational circuits for the purposes of supervised machine learning. We propose a circuit architecture suitable for predicting class labels of quantumly encoded data via measurements of certain observables. We observe that the required depth of a trainable classification circuit is related to the number of representative principal components of the data distribution. Quantum circuit architectures used in our design are validated by numerical simulation, which shows significant model size reduction compared to classical predictive models. Circuit-based models demonstrate good resilience to noise, which makes then robust and error tolerant.