High scRMSD and active site distortion (RFdiffusion -> LigandMPNN -> AF3)

[Valentin-Bio]

I am trying to troubleshoot high RMSD values in a self-consistency analysis (scRMSD). My pipeline involves designing a structure around a theozyme (5 residues + FMN cofactor) using RFdiffusion2, designing the sequence with LigandMPNN, and validating with AlphaFold3.

When running RFdiffusion2, I ensured the theozyme residues and rotamers were fixed. I verified that the output backbones maintained the correct x,y,z coordinates for the catalytic geometry.

I then ran LigandMPNN on 50 designs using the code below, explicitly fixing the theozyme residues:

for i in *.pdb
do

    filename=$(basename "$i" .pdb)

    python ./run.py \
        --model_type "ligand_mpnn" \
        --seed 111 \
        --pdb_path "$i" \
        --out_folder "./lama_50_designs/${filename}_results" \
	 --fixed_residues "A17 A63 A79 A138 A311" \
	 --repack_everything 0
done

However, after predicting the structures with AlphaFold3 (default AlphaFold Server settings), I am getting a high global RMSD (~3.22 Å). More critically, the theozyme geometry is completely disturbed. For instance, the distance between HIS17 in the AF3 prediction and the expected position from RFdiffusion2 is 19.0 Å.

Could I be missing a LigandMPNN setting, or is this likely an issue with the foldability of the generated sequence?

[oteri]

Hello @Valentin-Bio , you might probably quickly access the structure enabling the repacking in LigandMPNN. It is supposed to be good and, if even in this case you do not get a good geometry, then you can discard this backbone and find something else.

If the geometry seems correct, either the generated backbone is not compatible and you have hit a limitations in this model or maybe it is a AF3 limitations. You can try to use physic based approaches as orthogonal validation like energy minimisation and /or MD simulations on both some selected LigandMPNN and/or AF3 structures. It is probably expensive but you might get some insight on the correctness of the sequences.