Skip to content
Multigrid

Machine learning on molecules

Fingerprints, SMILES, property prediction and screening — representing chemistry so a model can compute over it.

A molecule is not a vector, a string or a sequence. It is a labelled graph with a three-dimensional shape, and almost every decision in molecular machine learning is about which parts of that object you agree to throw away. A SMILES string discards geometry. A fingerprint discards the graph and keeps a hashed inventory of neighbourhoods. A message- passing network keeps the graph and learns the inventory instead. None of those is the molecule, and the gap between the representation and the chemistry is where the surprising results come from.

These pages work from the representation outward. Each one names the parameter that actually changes the answer — a fingerprint’s radius, a split’s scaffold rule, a docking box — and says what breaks when the obvious choice is made. They describe research methods: a model that ranks a compound highly has produced a hypothesis for a laboratory to test, not a result.

Molecular Fingerprints Explained

How a Morgan/ECFP fingerprint turns a molecular graph into a bit vector, and what the radius and bit length each change about the answer.

10 min read

SMILES Notation Explained

The grammar of SMILES from atoms to stereochemistry, worked on one molecule, plus why two toolkits disagree about the canonical form.

10 min read

Graph Neural Networks for Molecules Explained

Why atoms-as-nodes and bonds-as-edges is the natural representation, how message passing builds an atom's environment, and what depth costs.

10 min read

Predicting Molecular Properties With Machine Learning

A four-descriptor solubility regression worked end to end, and the splitting and error-floor decisions that decide whether the number means anything.

10 min read

Retrosynthesis Prediction Explained

How a model proposes precursors for a target molecule, worked on one amide disconnection, and why top-1 accuracy understates what a good model does.

10 min read

Molecule Generation Models Explained

How generative models emit molecules as strings or graphs, what a validity check actually tests, and why validity is the easiest metric to satisfy.

10 min read

Predicting Chemical Reaction Outcomes With AI

How forward reaction models map reactants to products, worked on one esterification, and why yield is much harder than product identity.

10 min read

Virtual Screening: How AI Narrows a Compound Library

The filter-and-dock funnel from ten million compounds to a shortlist, with the enrichment arithmetic that tells you whether it worked.

11 min read

Molecular Similarity Search With Fingerprint Embeddings

Tanimoto similarity computed by hand, the size bias hiding in it, and the popcount bound that makes billion-compound search tractable.

10 min read

Predicting Protein-Ligand Binding Explained

What a binding-affinity model is actually scoring, the free-energy arithmetic that sets the accuracy it needs, and the shortcuts these models learn instead.

11 min read

What It Costs to Screen a Compound Library With AI

The compute cost of a ten-million-compound screen derived line by line from stated per-compound times and a stated hourly rate.

10 min read

Converting a Chemical Name to a Structure With AI

A lookup-first pipeline that resolves chemical names to SMILES with a deterministic parser, a database and a model fallback that is always validated.

11 min read

Other topics