When using autoencoders for biological data, don't assume reconstruction loss is a good guide—use this benchmark to find hyperparameters that actually improve downstream task performance.
BBOmix is an open-source benchmark with 105,000 pre-computed results for tuning autoencoders on biological data. It shows that reconstruction loss (what autoencoders optimize for) doesn't always predict how useful the learned representations are for downstream tasks, and provides baselines for hyperparameter optimization methods.