The previous generation of protein models, ESM2, hit a performance plateau with increasing parameters, whereas the new ESM Cambrian model continues to improve without a plateau.

Using sparse autoencoders, researchers found that the latent space of protein language models decomposes into interpretable features corresponding to biological concepts like amino acids, structural m...

The BioHub research team created a protein atlas with up to 7 billion folded protein structures, which is larger than the AlphaFold database.

The improved scaling performance of the ESM Cambrian model was achieved by increasing its training dataset from 50 million to 2.8 billion protein sequences, primarily by incorporating metagenomic data...

The ESM-Crambran model family demonstrates a log-linear scaling law, where performance on predicting long-distance protein contacts improves predictably with increased training compute.

On antibody design tasks, the single-sequence ESMFold2 model outperforms AlphaFold3, achieving a DocQ pass rate of 50 compared to AlphaFold3's 47.

ESMFold2, using only a single input sequence, achieves near-parity with AlphaFold3 on general protein-protein complex prediction, performing within 3 points on the DocQ pass rate metric.