From SMILES strings to molecular graphs — how we encode chemical structure for machine learning.
Takeaway 1. No representation is neutral. From Hammett parameters (1937) to SMILES, ECFPs, graph neural networks, VAEs and transformers, every representation is a choice about which structural information the model gets to see — and that choice caps what it can ever learn.
Takeaway 2. String-based representations (SMILES, SELFIES, and extensions like BigSMILES for polymers or HELM for proteins) are compact and easy to generate, but always check what they silently drop — stereochemistry, hapticity, and tautomers are common blind spots.
Takeaway 3. Extended-Connectivity Fingerprints (ECFPs) remain a strong, fast, interpretable baseline for small-molecule QSAR on datasets of roughly 100–10k molecules — but they carry no 3-D or conformer information, so they fail on shape-sensitive properties.
Takeaway 4. Expert-crafted descriptors (Hammett, Sterimol, DFT-derived) trade cost and availability for interpretability and data efficiency. ML surrogates (MLIPs, learned BDE/pKa predictors) can approximate expensive quantum-chemistry descriptors at a fraction of the compute, but must be benchmarked against a reference method and evaluated for domain of applicability.
Takeaway 5. Reactions and multi-component systems can be encoded by concatenating per-component descriptors or by encoding the whole reaction (reaction SMILES, DRFP, learned transformer encodings) — universality trades off against mechanistic interpretability. Before modeling, standardize features and visualize with PCA or UMAP to catch structure and outliers, but never over-interpret UMAP cluster sizes or distances.
For learned or unsupervised descriptors, see the lectures on Neural Networks, GNNs, and Transformers.