Protein Design

Designing a protein starts with a desired structure or function rather than a known sequence. A ligand-binding task supplies a small molecule but no ready-made pocket. An enzyme task supplies catalytic atoms whose arrangement must be preserved. The surrounding protein still has to be invented.

This environment tests one system across four settings: free generation, small-molecule binding, protein binding and enzyme scaffolding. A plausible drawing is only the beginning. Can a sequence be found that folds back into the proposed design?

Start from a real input: a ligand without a pocket

What must be invented?

A 150-residue protein must be generated around the ligand. The pocket geometry and fold are outputs, not templates. Sequence design then asks which amino acids could realize that geometry.

Enzyme scaffolding instead supplies catalytic atoms; protein binding supplies a target surface and hotspots. These are different conditioning problems for one shared system.

Ligand coordinates · PDB-derived coordinates (CC0); RFdiffusion3 benchmark selection.

Inside the generator, outside the measuring instrument

The generator gradually resolves noisy atoms around fixed constraints. A separate folding model then tests whether a sequence can realize the design.

1. Condition + noiseMotif, ligand or target atoms held at zero noise; generated atoms perturbed.
2. Connect atoms and residuesCombine local atomic detail with information about the wider protein.
3. Conditional denoisingRepeatedly refine the surrounding structure while keeping the supplied constraint fixed.
4. Generated coordinatesPropose a scaffold that accommodates the fixed molecular geometry.

External assay → 4 inverse-folded sequences → RoseTTAFold3 predictions → track-specific thresholds. A denoising training loss is not the benchmark score.

Build around a constraint, then test the round trip

Interactive 2D schematic. Move the scaffold, then deform its refold: fitting a context and reproducing a fold are different questions. No molecular inference, affinity prediction or benchmark scoring runs here.

Design settingEnzyme scaffoldSmall-molecule binderProtein binderUnconditional Scaffold closure

45%

Refold deformation

20%

Preserve catalytic anchors Align round-trip structures Reset example

1 · Conditional structure

2 · Sequence → refold → compare

LigandMPNN · 4 sequences → RoseTTAFold3

Actual track criterion

Enzyme: motif all-atom RMSD ≤ 1.5 Å after motif-backbone alignment, plus no non-motif N, CA or C atom within 1.5 Å of a ligand atom in the raw design.

Preserving catalytic geometry does not establish catalytic activity. A matching refold is computational evidence, not a functional experiment.

Report four separate target-averaged pass rates and diversity; no composite score. Missing designs remain failures. A design passes if any of four sequences meets its track rule (enzyme also considers five predictor models per sequence).

A computable score example

Illustrative small-molecule design, not a measured result. Each row is one refolded sequence. All four conditions must pass in the same row.

Sequence	Backbone RMSD (Å)	Ligand RMSD (Å)	Interface PAE	ipTM	Result
1	1.2	4.0	1.2	0.79	Fail: ipTM
2	1.6	4.0	1.2	0.85	Fail: backbone
3	1.2	5.5	1.2	0.85	Fail: ligand
4	1.2	4.0	1.8	0.85	Fail: PAE

Sequence 1 ipTM

0.79

Design fails: none of its four sequences meets all conditions.

For one target requesting 100 designs, suppose 20 already pass and 10 outputs are missing. Adding this design's verdict gives 20 / 100 = 20%. Missing outputs stay in the denominator. Four small-molecule target rates are then averaged; tracks remain separate.

Diversity is a different fraction: if those 20 passing structures form 5 clusters, target diversity is 5 / 20 = 0.25. It is averaged across targets, with zero assigned when none passes. Duplicating a successful fold can raise pass count without adding a new cluster.

The real ligand displayed above is a useful example: its atomic coordinates are known, but the protein around it is absent. Protein-binding requests instead provide a target surface and important contact sites. Enzyme requests specify a catalytic arrangement that the new scaffold must accommodate.

Evaluation contains 51 targets and requests 100 designs for each. Development structures come from the Protein Data Bank. Some evaluation structures are old enough to have appeared in upstream design-model training, so these results do not establish generalization to entirely unseen structural families.

The baseline is a compact RFdiffusion3 model with about 12.4 million learned parameters. During training, it receives disturbed coordinates and learns to recover a structure. Supplied ligand, target or catalytic atoms stay fixed while the surrounding atoms are denoised.

This teaches a relationship between a constraint and a compatible scaffold. Keeping catalytic atoms in place is easier than arranging a whole protein that can support them, however. The design must also be realizable by an amino-acid sequence.

For every proposed structure, an inverse-folding model suggests four sequences that might realize it. A structure predictor folds those sequences, and the resulting structures are compared with the intended design. Different tracks check overall shape, interface placement or catalytic geometry.

Coordinate discrepancies are measured in ångströms, a tenth of a nanometre. Confidence scores also enter some interface rules, but they are not measured binding affinity. Passing this computational assay does not establish binding or catalytic activity in a laboratory.

Each track reports its own pass rate and diversity. Repeating one successful structure may increase the number of passing designs while adding no new fold; diversity reveals that distinction. Missing designs remain failures.

The author-reported run passes about 42% of unconditional designs, 6% of enzyme designs, 3% of small-molecule binders and 0.2% of protein binders. The final figure is just one passing protein binder out of 500 requests, making relative improvements easy to overstate.

Enzyme success is also concentrated: 28 of 41 targets have no passing designs. All tracks satisfy the supplied conditioning geometry, showing why obeying a geometric constraint is not enough to survive the sequence-and-fold round trip.

The run was not independently reproduced, and repeated training seeds or matched conditioning ablations are unreported. The local assay uses RoseTTAFold3 rather than the original paper’s AlphaFold evaluation, so the rates are not directly comparable.

Compare designs made with the correct target against designs made with a shuffled target. Inspect how success varies across individual enzymes, rather than only averaging them. For binders, check whether the intended contact surface is engaged in addition to the scored thresholds.

These experiments would separate learned conditional design from reuse of a few convenient shapes. Improvements should preserve diversity and use comparable generation and refolding budgets.