Most de novo binder design today is diffusion. RFdiffusion denoises a backbone out of random coordinates, and it is the default for good reason. But there is a second paradigm that gets far less attention: designing a binder by gradient descent through a structure predictor, which is what ESMFold2 inversion does. Neither is the upgrade over the other. They search design space by different routes, and that difference is exactly what you want when one of them is stuck on a target.
This post walks through the two paradigms, then when to reach for the one most people are not using.
Two ways to sample a binder
Diffusion, RFdiffusion. Start from noise. A model trained on protein structures progressively denoises random 3D coordinates into a backbone, conditioned on the target and the hotspots you specify. The output is a stream of backbones; sequence is added afterward by ProteinMPNN. It is fast and it produces high diversity at scale, which is why it is the workhorse for discovery.
Gradient inversion, ESMFold2 design. Start from a soft, continuous representation of a sequence. Run gradient descent on that representation, backpropagating through the ESMFold2 fold network, so the optimizer jointly pushes the sequence and its predicted binding pose toward a target-engaging complex. Sequence and structure fall out of the same optimization rather than in two separate steps. On the Ranomics platform this is the minibinder mode of the ESMFold2 design tool, generating a 60 to 200 residue de novo scaffold.
The mechanical difference is real: one samples a distribution of structures and reads sequence off afterward, the other rides a fold model’s own landscape downhill with sequence and structure coupled the whole way.
When to reach for the gradient method
The point is not that gradient beats diffusion, or the reverse. It is that they get stuck in different places.
Prior diversity is the reason. A diffusion model and a gradient-inverted fold model have learned different notions of what a plausible complex looks like, so they fail and succeed on different targets. When one is producing nothing that validates, the other is the cheapest thing to try before you conclude the target is undesignable.
Why the priors differ
RFdiffusion learned a distribution over protein structures and samples from it. ESMFold2 is a structure predictor built on a protein language model, and inverting it means optimizing against the fold landscape that model has internalized from hundreds of millions of sequences. One prior is shaped like “what protein structures look like,” the other like “what the language model believes folds.” Those are different biases, and different biases reach different backbones. That is the whole reason to keep both in the kit.
The trap
Two ways this goes wrong.
The first is ranking the paradigms, deciding diffusion is the serious method and gradient inversion is a curiosity, or the reverse. That throws away the diversity that makes running both worthwhile. On any given target one may clearly win, but which one is not known in advance, and that is the argument for having both rather than for crowning one.
The second is the trap that applies to every generator: the fold model’s own confidence is not a binding measurement. A gradient run that converges to a low loss with a high ipTM has found a sequence its own fold network likes. That is a hypothesis. It becomes a binder when a display screen says so, not when the optimizer stops.
Quick reference
| RFdiffusion (diffusion) | ESMFold2 design (gradient inversion) | |
|---|---|---|
| How it samples | denoises a backbone from random coordinates | gradient descent through a fold network |
| Sequence step | added afterward by ProteinMPNN | jointly optimized with structure |
| Prior | structure-trained diffusion | language-model fold landscape |
| Best for | default discovery, volume, diversity at scale | a target diffusion is stuck on, prior diversity |
| Watch out | one prior’s blind spots | converged loss is a prediction, not binding |
Practical recommendations
- Default to diffusion for discovery. RFdiffusion’s throughput and diversity make it the right first pass on most targets.
- Reach for gradient inversion on a stuck target. When RFdiffusion’s designs will not validate, a different prior is the cheapest next move.
- Run both in parallel when the target is hard. Prior diversity across paradigms surfaces binders neither finds alone.
- A converged run is not a bound complex. A confident design is a candidate. Take it to display.
Run RFdiffusion or the gradient-based ESMFold2 design tool on tools.ranomics.com against a target structure. For the wider question of which generators to keep in the kit, see where PXDesign fits.
Summary
De novo binder design has two paradigms, not one. RFdiffusion denoises a backbone from noise and is the default for discovery. ESMFold2 inversion designs by gradient descent through a fold model, coupling sequence and structure, and it rides a different prior. Neither is better. They get stuck in different places, so when one produces nothing that validates, the other’s different view is the cheapest thing to try. Run both on a hard target, and remember that a converged run proves the model likes the design, not that the design binds.
Related Ranomics services
- How RFdiffusion works: the designer’s guide to the diffusion paradigm.
- Where PXDesign fits: choosing which generators to keep in the kit, and why we run several.
- Binder Pilot: take designs from either paradigm to a ranked, wet-lab-validated hit list on a fixed-scope campaign.