Optimal transport outperforms mean-pooling on property prediction tasks

Summary

Optimal transport outperforms mean-pooling on protein property prediction (1). This uses a learnable aggregation function to convert a variable-length vector generated by a protein language model into a fixed-length vector used for inference.

Figures

Figures from (1)

See also

• Mean-pooled embeddings outperform other zero-shot approaches for transfer learning of PLMs using the full sequence

1.

NaderiAlizadeh N, Singh R. Aggregating residue-level protein language model embeddings with optimal transport. Bioinformatics Advances. 2024;5(1). Available from: https://doi.org/10.1093/bioadv/vbaf060