Designing a Protein from Scratch

The Problem, Run Backwards

A protein is a chain of twenty kinds of amino acid, useful only once it folds. Computing the shape of a given sequence is the alphafold plate’s subject. Design goes the other way: fix the shape, fix the surface it must hold, then find a sequence that folds into it.

The difficulty of that direction is one of sheer size. The Baker lab’s 2016 review opens with it: a 200-residue protein has 20²⁰⁰ possible sequences, of which evolution has sampled only an infinitesimal subset.

On the plinth is one such design: the white porcelain is the target protein, the brass ribbon the binder made for it, three helices lying in its groove. Both are computed on this page; neither is a published design.

Design = the inverse of structure prediction · 200 residues → 20²⁰⁰ sequences (Huang et al. 2016)

The First New Fold

Before deep learning, design ran on physics: write an energy function, then let the computer go back and forth between sequence and backbone until it finds the pair with the lowest energy. That method is Rosetta.

In 2003 Kuhlman and colleagues used it to make Top7: 93 amino acids, an alpha/beta protein with, in the paper’s words, a topology not present in the Protein Data Bank. The protein came out extremely stable; its circular dichroism spectrum at 98 °C is nearly indistinguishable from the one at 25 °C.

The crystal structure, at 2.5 Å, sits 1.17 Å from the design model over all backbone atoms — 1.2 Å as the abstract rounds it. From then on “de novo” meant something: a shape could be settled on paper and then made.

Top7 · 93 amino acids · crystal at 2.5 Å · 1.17 Å from the design model (backbone atoms) · PDB 1QYS. The bundle on the bench is still this page’s own binder, not Top7, whose core is a β sheet.

A Backbone Out of Noise

RFdiffusion, in 2023, went about it differently: it fine-tuned the structure prediction network RoseTTAFold into a denoising diffusion model. It works not on atom coordinates but on residue frames — an alpha carbon and an N–Cα–C orientation for each amino acid. Training noises structures from the Protein Data Bank for up to 200 steps: Gaussian noise in three dimensions on the positions, Brownian motion on the matrices for the orientations.

Generation runs that backwards: start from a cloud of random frames and denoise step by step until they line up into a backbone that holds together. The paper adds self-conditioning, which lets each step see the previous step’s prediction, and which improved its benchmarks markedly. On an NVIDIA RTX A4000, a 100-residue protein can be generated in as little as 11 seconds.

The path on the plinth is the forward noising formula read backwards: at each step √ᾱ of the design is kept and √(1−ᾱ) of the noise, with the orientations turned back in the same proportion. It is this page’s interpolation, not a network; real denoising runs the network at every step.

RoseTTAFold fine-tuned into a diffusion model · residue frames: Cα + N–Cα–C orientation · 200 steps · here: the forward formula, backwards

Designed Against a Target

Generating a shape with nothing to hold gives a handsome fold and little else. What is useful is conditioning: put the target’s structure in, and grow the backbone against one patch of its surface. RFdiffusion can condition on symmetry, on a binding target, or on a functional motif that must be kept.

How well it holds starts with how many residues reach the other side. The bench counts alpha-carbon pairs within 10 Å, and the interface residues that follow from them; change the face the target offers and the count changes with it. Here the binder is lifted a little to show its seat: the gilded patch is the part of the target’s surface within 7 Å of the binder’s alpha carbons.

A groove is the easiest hold: a helix drops in and is touched on three sides. A flat face is held on one side only, and a bulge offers least of all. The bench passes a design with at least 22 interface residues, no clashes, and hydrophobic moments whose mean alignment is 0.6 or better; of the three faces, only the groove clears it.

Conditioning: symmetry · a target · a functional motif · here: alpha-carbon pairs within 10 Å

The Interface, Cut Open

Cut the target’s porcelain open and its own steel ribbon shows, lying against the brass of the binder. The layer between them is the interface.

Two kinds of contact cross it: water-avoiding side chains pressed together, squeezing the water out — the copper studs — and the charged or polar side, which can make hydrogen bonds, drawn as thinner pale ones.

The page sorts residues by Kyte and Doolittle’s 1982 hydropathy scale and then sorts the cross-chain alpha-carbon pairs within 8.5 Å into those two kinds. A real interface is read from every atom and from the angles of its hydrogen bonds; this one knows only backbone distances.

Here: cross-chain alpha-carbon pairs within 8.5 Å, in two kinds — water-avoiding, and able to hydrogen-bond

A Sequence for the Backbone

A backbone is not yet a protein; it needs a sequence. ProteinMPNN, in 2022, does that job: a message-passing network over the backbone graph, three encoder and three decoder layers, taking the distances between N, Cα, C, O and a virtual Cβ, and giving back an amino acid for every position.

It does not write from one end to the other; it samples a random decoding order each time. On 402 native monomer backbones its sequences match the native one 52.4 % of the time, against 32.9 % for Rosetta’s fixed-backbone design. For 100 residues it takes 1.2 seconds, where Rosetta takes 4.3 minutes.

The bench’s own rule is a toy: sort each position into core, boundary or surface by how buried it is, allow what that layer and the local shape take readily, and draw one that suits the neighbours already decided. The order is random, as ProteinMPNN’s is — which is why the letters appear scattered rather than in a line.

ProteinMPNN: 52.4 % native sequence recovery against Rosetta’s 32.9 % (402 monomer backbones) · random decoding order

The Exam Before the Bench

Generating is cheap and experiments are dear, so an exam goes in between: take the designed sequence, predict its structure with AlphaFold2, and compare that against the backbone the design asked for.

RFdiffusion states the mark precisely: from a single sequence, the AF2 prediction must have a mean predicted aligned error below 5, lie within 2 Å backbone r.m.s.d. of the designed structure, and within 1 Å on any scaffolded functional site. The paper says this measure has been found to correlate with experimental success. Eight ProteinMPNN sequences per backbone usually sit the exam.

Bennett and colleagues measured what the exam is worth in 2023: using AlphaFold2 or RoseTTAFold to judge whether a sequence adopts the designed structure, and whether that structure binds the target as designed, increased design success rates nearly ten-fold.

There is no structure prediction on the plinth. In its place is a check the page can compute: each helix’s hydrophobic moment, in the manner of Eisenberg and colleagues in 1982 — add up each residue’s hydropathy along the direction its side chain points. The copper arrow is that helix’s water-avoiding face, and it should point at the core or at the target. The helix lying in the groove is buried on both sides, so its water-avoiding residues are spread all round it: its arrow is short and its direction unclear.

The AF2 filter: pAE < 5 · within 2 Å of the backbone · within 1 Å at a site · 8 sequences per backbone · here: the hydrophobic moment

Into the Laboratory

A design that passes gets a synthetic gene, is expressed in bacteria, purified, and then measured to see whether it binds at all. RFdiffusion’s paper ran that for five targets — influenza haemagglutinin, IL-7Rα, PD-L1, the insulin receptor and TrkA — with 95 designs tested for each.

The overall experimental success rate was 19 %, which the paper calls an increase of roughly two orders of magnitude over their previous Rosetta-based method on the same targets, with fewer than 100 designs tested per target against thousands in earlier studies. Affinities reached the nanomolar range with no experimental optimisation.

The strongest influenza binder, HA_20, binds with a dissociation constant of 28 nM. Its complex with haemagglutinin was solved by cryo-EM at 2.9 Å; the structure sits 0.63 Å from the design model, and 0.6 Å over the binder alone.

Five targets · 95 designs each · 19 % overall · HA_20: 28 nM · cryo-EM at 2.9 Å · 0.63 Å from the design

What Has Been Made

Binders are what the field does best so far. Cao and colleagues designed binders in 2022 against 12 targets of differing shapes and surface properties, all smaller than 65 amino acids, which after experimental optimisation bound with nanomolar to picomolar affinities; five of the complexes were solved by crystallography.

Antivenom is one concrete use. Vázquez Torres and colleagues used RFdiffusion and ProteinMPNN in 2025 to design proteins that bind three-finger toxins: the short-chain neurotoxin binder has a dissociation constant of 0.9 nM and a melting temperature of 78 °C, the long-chain one 1.9 nM and above 95 °C. Given 15 minutes after a lethal dose of neurotoxin, they kept all five mice in the group alive.

Enzymes are harder. Lauko and colleagues designed serine hydrolases in 2025 whose abstract reports catalytic efficiencies up to 2.2 × 10⁵ M⁻¹ s⁻¹, crystal structures within 1 Å of the design models over the alpha carbons, and five folds distinct from those of natural serine hydrolases.

12 targets, all under 65 amino acids (Cao et al. 2022) · antivenom: 0.9 nM, five mice of five alive · designed hydrolase: 2.2 × 10⁵ M⁻¹ s⁻¹

What It Cannot Do Yet

Success rates are still low. Bennett and colleagues write in their 2023 discussion that much about interface energetics remains poorly understood and that success rates among the targets remain low, under 1 %. Cao and colleagues reported that for TIE2 and CD3δ, fewer than 10 binders came out of libraries of 100,000 designs. Some surfaces are simply hard: very flexible ones, very polar ones.

What is made tends to be simple. Kortemme’s 2024 review says the experimental successes so far are relatively simple problems: idealised folds, with all-helical proteins over-represented, symmetrical assemblies, and interfaces between proteins.

Shape is easier than function. The same review puts success rates for shape above 10 % and sometimes above 50 % depending on the goal, while a designed function is usually only a starting point, with activity that can be measured but is low, and that then has to be improved experimentally.

Every design ends up on a bench. ProteinMPNN’s paper puts it plainly: the ultimate test of sequence design methods is experimental characterization. One consideration sits outside the laboratory too: since no computationally designed protein can cause real-world harm unless it is physically produced, the ordering of synthetic DNA is a checkpoint for the field — the words of the 2024 Responsible AI × Biodesign statement, whose signatories include David Baker.

Success rates among targets under 1 % (Bennett et al. 2023) · fewer than 10 binders from 100,000 designs (Cao et al. 2022) · function harder than shape

The 2024 Chemistry Nobel

One half of the 2024 Nobel Prize in Chemistry went to David Baker, “for computational protein design”, and the other half jointly to Demis Hassabis and John Jumper, “for protein structure prediction”.

The press release puts it this way: proteins generally consist of 20 different amino acids, which can be described as life’s building blocks; in 2003, David Baker succeeded in using these blocks to design a new protein that was unlike any other protein.

The two halves are one thing seen from both ends. Once the link between sequence and shape is understood, the shape can be computed from the sequence, and the sequence written for the shape.

Nobel Prize in Chemistry 2024 · one half for computational protein design · the other for protein structure prediction

Taken Apart

The binder comes apart into three helices and two loops. Whichever the pointer rests on is lifted out and named.