Fig. 1: AVAMB and AAMB workflow overview, adversarial autoencoder model schematic representation, and AVAMB and AAMB performance across the benchmark datasets.

a AAMB workflow overview. Tetranucleotide frequencies and abundances across samples are extracted per contig and input to the AAMB encoder. After training, latent representations z and y are retrieved. Then, the VAMB clustering algorithm was applied to generate clusters from the z latent representation, and cluster labels were taken directly from y. Finally, bins from z and y are deduplicated to the final AAMB clusters. These can then potentially be integrated with VAMB generated clusters, in that case named AVAMB. Dark arrows represent forward propagations, grey arrows represent clustering and de-replication steps performed after training AAMB and VAMB. b Adversarial autoencoder model overview. The encoder-decoder was optimised to reconstruct the input contig features from the regularized latent representations z and y. Regularisation is achieved by adversarial competition between the discriminators and the encoder, enforcing the latent encodings to stay close to their prior distributions. Dark arrows represent forward propagations. Dashed arrows represent sampling processes from the latent and priors. c Number of distinct NC genomes reconstructed from the six benchmark datasets for VAMB (blue), AAMB(z) (light green), AAMB(y) (dark green), AAMB(z + y) (light purple), AVAMB (dark purple). GI Gastrointestinal, Urog Urogenital.