Peter DeWeirdt
Peter DeWeirdt
@pdeweirdt.bsky.social
PhD student @mit studying anti-phage defense systems using machine learning
When applied to 500 diverse bacterial assemblies, DefensePredictor revealed thousands of additional candidate systems awaiting validation.

6/n
January 9, 2025 at 7:32 PM
Among the systems were 7 protein domains that we newly validated in defense, including in DS-8 and DS-9, which have homology to the human innate immune protein SMPDL3A.

5/n
January 9, 2025 at 7:32 PM
We tested the model’s predictions in lab and identified 45 novel defense systems, observing a strong correlation between the validation rate and defense probability.

4/n
January 9, 2025 at 7:32 PM
DefensePredictor demonstrates strong performance on a held-out set of defense genes and leverages biological patterns like low GC content to make its predictions.

3/n
January 9, 2025 at 7:32 PM
We compiled a dataset of known defense systems and trained a gradient boosting model to distinguish defense from non-defense genes, using ESM2 embeddings and genomic features of each gene and its neighbors.

2/n
January 9, 2025 at 7:32 PM