I want to close out by sending a huge thank you to @pcronald.bsky.social, @ellenrim.bsky.social, @jimnotwell.bsky.social, @the-real-og.bsky.social, and @yeahgene.bsky.social. Your guidance and contributions have been essential, and I'm incredibly grateful to work with you all!

Overall, this approach shows utility by learning from the phenotypic effects associated with tested sequence variation and extending predictions towards previously unseen sequence variants. I'm very excited to see what the broader applications of this approach may yield in the future! (11/11)

Genotype-to-phenotype methods are needed to utilize large genomic datasets. Approaches like these can help massively speed up the process of screening sequence variants and identifying top candidates to devote more attention towards in the lab. (10/11)

ESM-2 fine-tuned on this data was able to successfully correlate NUDT15 variation with changes in functionality, and extended predictions on enzyme functionality to sequence variants not generated in the original assay, such as uncharacterized in-frame indels found in some individuals. (9/11)

This approach was also applied to a dataset of NUDT15 sequence variants published by Suiter et al. to see how generalizable the pipeline was. NUDT15 is an enzyme in humans important for processing thiopurine drugs, and loss-of-function variations can cause thiopurine cytotoxicity. (8/11)

We were curious to see if this was accurate given none of our training Pik-1 sequence variants contained indels. Testing two highly-scored variants with insertions in vitro confirmed that they do bind to Avr-PikC more strongly than Pikh-1! (7/11)

These models were used to predict the ligand binding of Pik-1 sequence variants found in the 3,000 Rice Genomes Project dataset. Several variants with in-frame insertions in the Pik-1 ligand binding domain were predicted to bind to Avr-PikC stronger than our wild type control Pikh-1. (6/11)

We used this data to fine-tune the protein language model ESM-2 to correlate Pik-1 variation with changes in Avr-PikC/F binding. This modeling approach outperformed alternate models trained on ESM-2 embeddings, highlighting the advantages of fine-tuning for PLM applications. (5/11)

Directed evolution of the Pik-1 ligand binding domain was previously conducted by ellenrim.bsky.social to identify gain-of-function variations which expand Pik-1 Avr-PikC/F recognition. This produced a set of thousands of Pik-1 variants and their associated binding behaviors to explore. (4/11)

Two variants of Avr-Pik, Avr-PikC and Avr-PikF, escape Pik-1 binding and thus skirt Pik-1 mediated immunity. We wanted to identify any variants of Pik-1 present in rice which may be capable of recognizing these escaping ligands. (3/11)

The rice immune receptor Pik-1 recognizes the peptide Avr-Pik, produced by the fungus M. oryzae. This pathogen is responsible for destroying enough rice to feed 60 million people a year. Upon binding Avr-Pik, Pik-1 activates to confer immunity against the pathogen. (2/11)