Finally, we also tested its ability to predict outcomes for convergently evolved receptors. SCORE, a csp22 receptor found in Citrus and relatives, was recently discovered by @brunongou.bsky.social. We preliminarily found mamp-ml can zero-shot predict immunogenic outcomes!

Compared to other state-of-the-art models such as AlphaFold3, mamp-ml can predict these outcomes with higher confidence, even when a solved structure is not available. We anticipate users can now screen for immunogenic outcomes of receptor-ligand variants before spending time and $$ in the lab.

Mamp-ml harnesses the power of protein language models (ESM-2) to build a classifier model for immunogenic outcomes.

Previously immunogenicity was primarily characterized in the lab but was a expensive bottleneck considering the variation captured computationally!

We then generated a pipeline that combines AlphaFold2 and LRR-Annotation to precisely extract the LRR ectodomain as well as generated features to improve model training. This even includes tracking which residues interface with the ligand!

Mamp-ml was built upon two decades of foundational research. To generate the training data required, I manually pulled receptor and epitope sequences from every paper I could find, small or large. In total, we were able to capture over 1,300+ combinations across 11 receptors and 91 plant species.

Sometimes the smallest things warm the soul ❤️ so thankful to be part of such a great lab group!