Lecture 26 — Proteins, molecules, and structural biology

Lecture 26 — Proteins, molecules, and structural biology#

⚠️ AI-synthesized; not fully reviewed by course staff. Treat as a study aid; released slides, notebooks, and lecture recordings are authoritative.

Thu Apr 23, 2026 · Part 5 — Molecules & biological AI

What this lecture is about#

Before modern biological AI (L27), there is a century of methods for working with proteins and small molecules — sequence alignment, multiple sequence alignment (MSA), position-specific scoring matrices (PSSMs), homology modeling, QSAR, docking. This lecture is an honest tour of the classical toolkit, and the case for why those classical methods are still strong baselines that modern AI papers should be benchmarked against (a recurring theme of L27 and the course in general).

We split the lecture into two halves: proteins, then small molecules.

Half 1: Proteins.

  • Protein biology and structure levels.

    • Primary structure. The amino-acid sequence (length 50-2000 typically).

    • Secondary structure. Alpha helices, beta sheets, loops.

    • Tertiary structure. The 3D fold of a single chain.

    • Quaternary structure. Multi-chain complexes.

  • Experimental structure determination. X-ray crystallography (high resolution, requires crystallization), NMR (smaller proteins, in solution), cryo-EM (large complexes, recent revolution). Each has trade-offs.

  • The protein data gap. ~250 million known protein sequences. ~200,000 experimentally-determined structures (PDB). The 1000× gap is the reason structure-prediction matters — and is what AlphaFold (L27) attacked.

  • Evolutionary relationships and sequence conservation. Proteins related by descent share sequence; conserved positions indicate functional importance.

  • Sequence alignment. Pairwise alignment (Smith-Waterman, BLAST) — find regions of similarity between two sequences. Used routinely for “what does this new sequence look like?”

  • Multiple sequence alignment (MSA). Align hundreds-to-thousands of homologous sequences. Reveals which positions are conserved and which co-evolve. The substrate for everything else in this lecture and for AlphaFold.

  • PSSMs (Position-Specific Scoring Matrices). Per-position amino-acid frequencies extracted from an MSA. Captures “what residue is allowed here?” Used in PSI-BLAST and earlier structure-prediction methods.

  • Coevolution. Pairs of positions that mutate together suggest spatial contact. The 2010s breakthrough (DCA, EVfold) used coevolution to predict 3D contacts from sequence alone.

  • Variant-effect prediction. SIFT (sequence conservation) and PolyPhen (sequence + structural features) predict whether a missense variant is likely to disrupt function. Still strong baselines in 2026.

  • Homology modeling. When a structurally-determined homolog exists, build a 3D model of your protein by aligning to it. Pre-AlphaFold, the standard route to a structure.

Half 2: Small molecules.

  • Small molecules. Drugs, metabolites, ligands. Typically 10-100 atoms, characterizable by 2D structure (atoms + bonds) or 3D conformation.

  • SMILES strings. A linear textual encoding of a molecule. Canonical SMILES is unique; non-canonical is not. Easy to parse, easy to use as ML input.

  • Molecular fingerprints (ECFP / Morgan fingerprints). Bit-vector encodings: bit i is on iff some local substructure of radius r is present. Fast similarity (Tanimoto), fast classification (random forests on fingerprints are still competitive).

  • QSAR — Quantitative Structure-Activity Relationship. “Predict bioactivity from structure.” Random forest on fingerprints + log-transformed activity readouts is the canonical setup.

  • Molecular docking. Predict how a small molecule binds to a protein pocket. AutoDock Vina is the long-standing workhorse; commercial tools (Glide) and recent ML-based docking (DiffDock and successors) are increasingly competitive on benchmarks, though their behavior off-distribution is still debated. Useful when you have a structure for the target.

  • The drug-discovery pipeline. Hit identification → lead optimization → preclinical → trials. ML can help at every stage; usually it helps most at the early ones (hit-finding, virtual screening).

  • Scaffold splits. A train/test split that separates by chemical scaffold, not by random sample. Random splits over-estimate generalization for molecules because similar molecules end up in both train and test. Scaffold splits are the honest evaluation.

  • The classical-baseline challenge. A consistent finding in the molecular-ML literature: a random forest on Morgan fingerprints, evaluated with scaffold splits, is competitive with most reported deep-learning approaches. Not always, but often. The lecture pushes back against “deep learning solved chemistry.”

Why it matters#

Three reasons this lecture is the right setup for L27:

Classical methods exploit strong domain structure. Sequence alignment, MSAs, PSSMs, fingerprints, docking — all of these encode generations of biological knowledge into the model itself. Modern deep-learning methods compete with these only when they encode similar (or stronger) priors.

Evaluation matters as much as architecture. Scaffold splits, conservation-aware evaluation, MSA-aware benchmarks — these are the difference between an over-optimistic claim and an honest one. The same architecture can look great or terrible depending on the split.

Domain knowledge isn’t optional. A protein-prediction model that doesn’t use MSA information is throwing away the strongest signal in the field. A small-molecule model that doesn’t use fingerprints or 3D structure is doing it the hard way.

Things you should walk away believing#

  • Biological sequences and molecules have strong structure (evolution, conservation, chemistry) that classical methods exploit. Don’t dismiss them.

  • MSA-derived features (PSSMs, conservation, coevolution) are the strongest single source of signal in protein modeling.

  • SIFT / PolyPhen are still legitimate variant-effect baselines in 2026.

  • SMILES / fingerprints / random forests are still legitimate small-molecule baselines.

  • Random splits over-estimate molecular-ML performance; scaffold splits are the honest evaluation.

  • Docking is a useful tool, not a black box. Know its assumptions before you trust the score.

  • The PDB has ~230K structures (as of 2025); sequence databases have hundreds of millions. The gap is the reason structure prediction (L27) matters.

  • “Deep learning beats classical X” claims should be checked with proper splits and proper baselines.

How this connects to the rest of the course#

  • L25 set up DNA / RNA / population genetics; this lecture moves to the protein and small-molecule layers.

  • L27 (next lecture) extends this with deep learning — protein language models (ESM-2), AlphaFold-style structure prediction, equivariant 3D models, generative protein design.

  • L15 (foundation models) — protein LMs are foundation models over (protein sequence, ?), and the “?” is non-obvious.

  • L24 (causality, fairness) — drug-discovery models inherit biases of their training data (which compounds were measured, which assays were used).

  • L10’s domain-shift story applies: scaffold splits expose lack of generalization; random splits hide it.

Study guide#

Key terms, self-check questions, and additional resources for active recall.

Key Terms#

Term

Definition

Primary structure

Amino-acid sequence of a protein. Length ~50-2000 typically.

Secondary structure

Local conformations: α-helices, β-sheets, loops.

Tertiary structure

The 3D fold of a single protein chain.

Quaternary structure

Multi-chain complexes.

PDB (Protein Data Bank)

The reference database of experimentally-determined protein structures (~200K).

Protein data gap

~250M known protein sequences vs. ~200K experimental structures. The gap that makes structure prediction matter.

MSA (Multiple Sequence Alignment)

Aligning hundreds-to-thousands of homologous sequences. Reveals conserved positions and co-evolving pairs. The substrate for everything else.

PSSM

Position-Specific Scoring Matrix: per-position amino-acid frequencies extracted from an MSA.

Coevolution

Pairs of positions that mutate together — strongly suggestive of spatial contact.

SIFT / PolyPhen

Classical missense-variant-effect predictors using sequence conservation (SIFT) and conservation + structural features (PolyPhen). Still strong baselines.

Homology modeling

Build a 3D model by aligning to a structurally-determined homolog. Pre-AlphaFold standard.

SMILES

A linear textual encoding of a molecule (atoms + bonds). Canonical SMILES is unique.

ECFP / Morgan fingerprint

Bit-vector encoding: bit i is on iff some local substructure of radius r is present. Default small-molecule featurization.

QSAR

Quantitative Structure-Activity Relationship: predict bioactivity from structure. Random forest on fingerprints + log-activity is the canonical setup.

Tanimoto similarity

|A ∩ B| / |A ∪ B| over fingerprint bits. Standard chemical-similarity measure.

Molecular docking

Predict how a small molecule binds to a protein pocket. AutoDock Vina is the long-standing workhorse; Glide (commercial) and ML-based DiffDock-style methods are increasingly competitive on benchmarks.

Scaffold split

Train/test split by chemical scaffold rather than at random. Random splits over-estimate molecular-ML performance.

Classical Methods Are Strong#

Don’t dismiss them

Sequence alignment, MSAs, PSSMs, fingerprints, docking, QSAR — these encode generations of biological knowledge into the model itself. Modern deep-learning methods compete with them only when they encode similar (or stronger) priors. A random forest on Morgan fingerprints, evaluated with scaffold splits, is competitive with most reported deep-learning approaches on molecular-property prediction. Not always, but often.

Evaluation Matters As Much As Architecture#

Random splits over-estimate; scaffold splits are honest

Molecular datasets contain many near-duplicate compounds (same scaffold, different decorations). Random splits put near-duplicates in both train and test, inflating apparent generalization. Scaffold splits (or stricter time-split / cluster-split evaluations) are the honest test.

Self-Check Questions#

  1. What’s the difference between a PSSM and a learned protein embedding? Why is PSSM still used in 2026?

  2. SIFT scores a missense variant by sequence conservation. PolyPhen adds structural features. When would you reach for SIFT vs. PolyPhen vs. an ESM-2 zero-shot score (L27)?

  3. ECFP fingerprints encode local substructure. Why do they work surprisingly well on QSAR despite ignoring 3D geometry?

  4. AutoDock Vina returns a “docking score.” What does it actually estimate, and what are its known failure modes?

  5. A paper reports RMSE 0.55 on lipophilicity prediction (a regression task) with random splits and RMSE 1.20 with scaffold splits. Which number should the reader take away as the model’s true performance, and why does the gap matter?

Additional Resources#