The RFdiffusion2 paper claims that users can "specify the RASA of each atom" for fine control over ligand burial depth, but this feature is not implemented in the open source codebase? Only molecule-level RASA conditioning is available.

Start by reading get_relative_sasa_inference in rf_diffusion/conditions/v2.py and the relative_sasa_v2 settings in rf_diffusion/config/inference/base.yaml. Compare their molecule-level handling with the partially_fixed_ligand parsing and atom-index mapping described in the issue. Done means the supported configuration and behavior for per-atom RASA are defined and implemented, or the limitation is clearly documented.

Background

I've been exploring RFdiffusion2 for small molecule binder design and noticed what appears to be a discrepancy between the paper's claims and the open source implementation. I'd appreciate help from the community to understand if I'm missing something or if this feature needs implementation.

Paper Claim vs. Implementation

The RFdiffusion2 paper states:

"First, to provide finer control over the depth at which each reactant and/or cofactor is buried within the protein, we enable users to specify the RASA of each atom"

However, in my analysis of the codebase, I only found molecule-level RASA conditioning. Am I looking in the wrong place, or is this feature not yet implemented?

What I Found

Current RASA Implementation

File: rf_diffusion/conditions/v2.py (lines 135-141)

def get_relative_sasa_inference(indep, feature_conf, feature_inference_conf, cache, **kwargs):
    if not feature_inference_conf.active:
        return {'t1d':torch.zeros((indep.length(), feature_conf.n_bins + 1))}
    rasa = torch.full((indep.length(),), feature_inference_conf.rasa)  # Single scalar
    is_feature_applicable = indep.is_sm  # Applied to ALL small molecule atoms
    # ... rest applies same RASA to all ligand atoms

Current Configuration

# rf_diffusion/config/inference/base.yaml
relative_sasa_v2:
  active: False
  rasa: -1  # Single value - no per-atom specification

Atomic-Level Control Exists Elsewhere

Interestingly, I found the partially_fixed_ligand system does support atomic-level control:

# This works for structural fixing:
inference.partially_fixed_ligand="{BCA:[C6B,C5B,C7B,C4B,O2B]}"

Questions for the Community

1. Am I missing something? Is there another configuration parameter or method for atomic-level RASA that I haven't found?

2. Is this a known limitation? Should the paper's claim be interpreted differently?

3. Implementation pathway? If this needs to be implemented, would extending the existing patterns (similar to partially_fixed_ligand) be the right approach?

4. Workaround exists? Is there a current method to achieve per-atom RASA control that I'm not aware of?

Potential Implementation Ideas

Please let me know if this approach makes sense or if there's a better way:

# Hypothetical per-atom RASA syntax:
inference.per_atom_rasa:
  LIG:
    OH: 0.8    # Highly buried hydroxyl
    C1: 0.3    # Partially exposed carbon
    C2: 0.1    # Surface-exposed carbon

The architecture seems to exist since partially_fixed_ligand successfully:

• Parses per-atom specifications
• Maps atom names to internal indices
• Validates atom existence
• Applies individual control per atom

My Analysis Could Be Wrong

I want to emphasize that I might be:

• Looking in the wrong files
• Misunderstanding the configuration system
• Missing a parameter that enables this functionality
• Misinterpreting the paper's claims

Any guidance, corrections, or insights would be greatly appreciated!

Environment

• RFdiffusion2 open source repository
• Analyzed primarily rf_diffusion/conditions/v2.py and config files
• Used standard configuration files without modification

---

If this is indeed a missing feature, I'd be happy to contribute to its implementation with guidance from maintainers and the community.