Learned protein embeddings for machine learning
Problem
Converting protein sequences into vector representations that let machine-learning models predict biological properties of unseen sequences.
Why it matters
Better sequence embeddings enable prediction and discovery of proteins with optimal properties.
What the AI did
Learned low-dimensional embeddings of protein sequences were trained on vast unmeasured sequence data and used in Gaussian process models for downstream prediction.
Results
Predictive power comparable to existing representations with orders of magnitude fewer dimensions.
Impact
Embeddings match existing representation quality while drastically reducing dimensionality, per the abstract.
Record
| Venue | Bioinformatics |
|---|---|
| Published | 2018 |
| Work type | article |
| Authors | Kevin Yang, Zachary Wu, Claire N. Bedbrook, Frances H. Arnold |
| Citations | 313 (OpenAlex, at import time) |
| References | 25 |
| Model or method | Learned protein embeddings with Gaussian process models |
| Organization | not stated in source |
| Primary source | https://doi.org/10.1093/bioinformatics/bty178 |
| Open access | bronze |
| Keywords | protein embeddings, machine learning, Gaussian process, sequence representation, bioinformatics |
Importance rating
real but narrower advance
Better protein sequence embeddings enable prediction and discovery of proteins with optimal properties; a useful methodological advance but no application scale given.