Embedding
A numerical vector representation of an object (a word, protein, molecule, or document) that captures its meaning or properties in a form AI models can work with — objects that are semantically similar tend to have similar embeddings.
What it means
An embedding is a fixed-length list of numbers (a vector) that represents an object — a word, a sentence, a protein sequence, a molecule, or an entire document — in a high-dimensional space. The key property is that the distances between embeddings encode meaningful relationships: similar objects end up close together in the embedding space, dissimilar objects far apart.
Embeddings are how AI models convert discrete, symbolic inputs (text characters, amino acids, atoms) into the continuous numerical representations that neural networks can process and compare.
Why embeddings matter for research
Embeddings are the intermediate representation underlying most AI tools used in research:
- Literature search: Semantic Scholar and similar tools embed paper abstracts; searching returns papers whose embeddings are closest to your query — this is why semantic search finds conceptually related work even without keyword matches
- Protein property prediction: ESM-2 and other protein language models generate embeddings for protein sequences; these embeddings are used as input features for predicting properties (stability, binding affinity) without needing hand-engineered features
- Molecular property prediction: Models like Chemprop generate molecular graph embeddings that capture structural information for property prediction
- Document similarity: NotebookLM and other RAG tools embed uploaded documents and retrieve the most relevant passages when you ask a question
What high-dimensional means in practice
Modern embeddings are typically 512 to 4,096 dimensional — meaning each object is represented as a vector with that many numbers. You rarely work with embeddings directly; they’re internal representations the model uses. What you see as a user is the downstream output (a search result, a property prediction, a generated answer).
Embeddings vs. features
In traditional machine learning, researchers hand-crafted features to represent their data (molecular fingerprints for molecules, physicochemical properties for proteins). Embeddings replace this: the model learns representations directly from raw data, often capturing structure that hand-crafted features missed.