Compression encodes complex data in a lower-dimensional representation that preserves the essential information needed for later reconstruction or use. In machine learning, it is closely related to representation learning and dimensionality reduction: irrelevant variation can be discarded while useful signal remains. section 1.2.2
For molecular design, a generative model can compress each molecular blueprint into coordinates in a continuous latent space. The decoder can reconstruct the original structure or translate a new point into a molecular structure. Training may organize nearby points around chemically or biologically similar molecules, but this behavior is not guaranteed: some points may decode to invalid strings, duplicate molecules, or chemically unrelated neighbors. section 10.1.2
Compression supports molecular design by:
- Reducing the difficulty of storing, searching, and manipulating large collections of detailed molecular structures. section 10.1.2
- Turning discrete graph-editing problems into movement through a lower-dimensional continuous space. section 10.1.2
- Supporting generation by decoding sampled points, including points that do not represent existing molecules. section 10.1.2














