Affinity-VAE: incorporating prior knowledge in representation learning from scientific images