Introduction to protein design: the pipeline, AlphaFold, and protein language models

From an interaction target to a protein that binds it. The pipeline that made de novo design routine, and the learned models underneath it.

CCBS · EST 5053 · Fall 2026 · lecture 15 · core page · 2026-10-29 (Thu) 10:40–12:15 · students present, lecturer-backed

This lecture steps out of the cell. The object is one molecule, and the question is whether you can design it to do what you want. A decade ago designing a binder to a named target was a research programme; now it is a pipeline with published recipes. The lecture walks the pipeline end to end and then looks underneath, at what the learned models are actually doing: structure prediction, inverse folding, and the protein language models that treat sequence space as a corpus. The critical question is the one this course always asks: what do these models get right, what can they not say, and how would you know.

Still to be written into this core page

This lecture is new in 2026 or substantially re-scoped, so parts of it are not yet carried over from the 2025 notes. The teaching team should treat this list as the commissioning brief, and the lecturer as the backlog.

Protein language models: ESM and relatives, what a masked-token objective learns about structure and function, and where it fails.
AlphaFold in more depth: what the confidence metrics mean, and why a confident structure is not a confident function.
The honest evaluation question: success rates in the literature, what counts as a hit, and what the failures look like.
The connection back to the course: a designed binder is a binding reaction whose affinity you chose, so it is a designed edge in a binding-catalysis network.

Part 1

The pipeline

From binder target to designed sequence, and the tools that made it automatic.

1The de novo binder pipeline

Gene library → yeast surface display → selection → verification: the experimental half of design, and the LCB1 case.

The 2025 hands-on lecture opens with the classical machinery: Rosetta terms (residue, mover, filter, pose, conformation space, score function), RosettaScripts (the XML workflow language: scorefxn / residue selectors / task operations / movers / protocols), and then the full de novo pipeline¹:

gene library synthesis → yeast surface display → limited protease/heat → FACS + NGS → select designs → E. coli expression → SDS-PAGE, size-exclusion, Octet binding

The canonical result: SARS-CoV-2. Two approaches to the ACE2-binding problem — build de novo proteins off the ACE2 helix, or dock designed scaffolds onto the binding region — yielded LCB1: a 55-residue miniprotein, Tm > 95 °C, designed to block spike–ACE2. This is the floor of the field's proof: designs that fold, bind, and survive a real pandemic.

2RFdiffusion → ProteinMPNN → AlphaFold

The AI pipeline: diffusion generates the backbone, a graph autoregressive model writes the sequence, structure prediction validates — and the loop closes.

The modern trio, each with a role:¹

RFdiffusion (Watson et al., Nature 2023) — generates backbones conditioned on the target. Two recipes from the lecture: unconditional binder design (contigmap.contigs=[B1-193/0 90-90] with hotspot list [B120,B122,B123,B160,B172] and the complex-trained checkpoint), and conditional design carrying a fixed motif ([B1-193/0 A19-42/60-60], general checkpoint).
ProteinMPNN (Ingraham et al., NIPS 2019; Science 2023) — a graph-based, autoregressive sequence model: sequence given structure; fast, in silico validated, and it designs function (binder-affinity optimization).
AlphaFold2 / ColabFold (Jumper et al., Nature 2021) — validates the design: does the designed sequence fold to the designed structure? ColabFold swaps the slow MSA search (jackhmmer/HHblits/HHsearch, hours) for MMseqs2 k-mer prefiltering (<4 min on one core) — the difference between a chore and a day-trip exercise.

What "the era of AI" did not change

The pipeline still ends in the lab: protease stability screens, FACS, binding validation. RFdiffusion→MPNN→AF2 (the lecture's closing diagrams) replaced the search, not the measurement — and the knot example (same seeds, different topology; recycles up to 12 for β-barrels) is the reminder that generative tools are sampled, not solved.

†References & note sources

Where this page's claims and numbers live.

F. Xiao, CCBS 2025 Lecture 9 — Introduction to Protein Design (slides: Rosetta, de novo binder pipeline, SARS-CoV-2/LCB1, RFdiffusion, ProteinMPNN, AlphaFold2/ColabFold/MMseqs2, hands-on recipes). PDF This lecture traces here, including the two run_inference recipes.
F. Xiao, CCBS 2025 Lecture 11 — Growth machine (growth law, chemostat/Monod, consumer–resource, proteome partition λ = εφ_R, flux balance, upshift/downshift, Cobb–Douglas/Solow economics analogy, V̇ = λV frontier). PDF This lecture traces here.
L. Cao et al., Science 2020 — SARS-CoV-2 miniprotein inhibitors (LCB1). The case study for this lecture.
J. Watson et al., Nature 2023 — RFdiffusion; J. Ingraham et al., Science 2023 — ProteinMPNN; J. Jumper et al., Nature 2021 — AlphaFold2; M. Steinegger & J. Söding, Nat. Biotechnol. 2017 — MMseqs2. Toolkit references (cited on the lecture slides).
M. Scott, C. Gunderson, E. Mateescu, Z. Zhang, T. Hwa, Science 2011 — the proteome-partition growth law.The backbone result for this lecture.
R. Phillips & J. Kondev, Cell biology by the numbers (book.bionumbers.org) — metabolic yields and energy scales.The numbers for this lecture.