Protein Language Models (PLMs) are deep learning models trained specifically on proteins — biological molecules — as their subject matter.
In short, they work very much like the large language models (LLMs) behind something like ChatGPT, except that “protein” is treated as a kind of language to be understood.
1. Why think of proteins as a “language”?
In biology, a protein is a long chain formed by 20 basic amino acids linked together in a specific order.
The analogy:
- Human language: 26 letters form words, words form sentences.
- Protein language: 20 amino acids (denoted by letter codes like M, A, S, K) form a “protein sequence.”
The logic: a protein’s amino acid sequence determines its three-dimensional structure, and that structure determines its function. Just as a sentence’s grammatical structure determines its meaning, a protein’s amino acid sequence has a hidden “biological grammar” embedded in it.
2. How do PLMs work?
These models typically use self-supervised learning on large protein databases (such as UniProt):
- Pretraining: the model reads hundreds of millions of known protein sequences from nature and learns the “contextual relationships” between amino acids. It learns by guessing masked-out portions of a sequence (e.g., randomly removing a few letters from a sequence and having the model fill them back in).
- Deep insight: during training, the model automatically picks up on which amino acids tend to co-occur, which positions are structurally critical, and which regions function as active sites.
3. What can PLMs do? (Applications)
PLMs have dramatically accelerated R&D in biomedicine and materials science. Common tasks include:
- Protein structure prediction: predicting the 3D shape a sequence folds into (the famous example being ESMFold).
- Function annotation: inferring what a protein actually does from its sequence — is it an enzyme? Does it bind a specific drug molecule?
- Protein design (de novo design): the most exciting frontier — rather than modifying existing natural proteins, AI “creates from scratch” entirely new proteins with specific functions (for example, designing enzymes that break down plastic, or protein-based antiviral drugs).
- Mutation effect prediction: predicting whether swapping out a particular amino acid in a sequence would enhance, weaken, or introduce toxicity into a protein’s function.