 What's new in string version 12.0? String is a database approach in association networks. We've just released a major update, meaning new genomes and updated gene annotations for existing genomes. The co-expression channel features a new algorithm called Favre, which uses variational autoencoders to produce a latent space in which we do the correlation analysis. We further added single-cell transcriptomic data leading to many more associations than earlier versions. The experiment channel has a new scoring scheme for low throughput studies. This takes into account the assay type used to measure them, giving us much better separation between what is reliable and what is less reliable. Finally, the text binding channel has an improved transformer which is based on biomedical Roberta. This time we've trained it to also do cross-sentence extraction, giving us twice as many physical interactions as previous versions at all confidence levels.