Adding noise while training non-Ab inverse folding models improves self-consistency while worsening sequence recovery

Summary

Introducing noise to structures during training improves self-consistency RMSD of generic inverse folding models but not antibody inverse folding models, while worsening sequence recovery (1–2). It also improves self-consistency RMSD when forward-folding de novo designed proteins (3).

Figures

Exp/Pred	Layer Decay	OAS Gaussian Noise	Test Masking	FR Avg.	CDR1H	CDR2H	CDR3H	CDR1L	CDR2L	CDR3L
Exp	-	-	None	0.898	0.731	0.712	0.569	0.723	0.736	0.718
Exp	-	✓	None	0.898	0.735	0.698	0.566	0.716	0.702	0.713
Exp	✓	-	None	0.895	0.741	0.700	0.584	0.716	0.741	0.725
Exp	✓	✓	None	0.894	0.727	0.702	0.573	0.720	0.728	0.727
Exp	-	-	CDRs	0.894	0.680	0.637	0.432	0.677	0.689	0.661
Exp	-	✓	CDRs	0.894	0.696	0.651	0.434	0.692	0.680	0.659
Exp	✓	-	CDRs	0.890	0.675	0.657	0.431	0.666	0.689	0.658
Exp	✓	✓	CDRs	0.891	0.681	0.653	0.430	0.666	0.698	0.655
Pred	-	-	None	0.909	0.753	0.716	0.561	0.738	0.731	0.722
Pred	-	✓	None	0.905	0.749	0.704	0.558	0.729	0.725	0.722
Pred	✓	-	None	0.907	0.750	0.730	0.572	0.746	0.737	0.730
Pred	✓	✓	None	0.903	0.744	0.713	0.554	0.744	0.733	0.718
Pred	-	-	CDRs	0.904	0.706	0.650	0.445	0.691	0.687	0.665
Pred	-	✓	CDRs	0.901	0.709	0.657	0.435	0.701	0.690	0.658
Pred	✓	-	CDRs	0.903	0.695	0.654	0.435	0.675	0.675	0.654
Pred	✓	✓	CDRs	0.898	0.699	0.647	0.433	0.682	0.682	0.658

Table from (5)

Figure 3b from (3)

Ref (4)

See also

• ProteinMPNN
• ESM-IF
• Adding noise sometimes leads to greater diversity during inverse folding inference

1.

Dauparas J, Anishchenko I, Bennett N, Bai H, Ragotte RJ, Milles LF, et al. Robust deep learning–based protein sequence design using ProteinMPNN. Science. 2022;378(6615):49–56. Available from: https://doi.org/10.1126/science.add2187

2.

Hsu C, Verkuil R, Liu J, Lin Z, Hie B, Sercu T, et al. Learning inverse folding from millions of predicted structures. openRxiv; 2022. Available from: https://doi.org/10.1101/2022.04.10.487779

3.

Ren M, Yu C, Bu D, Zhang H. Accurate and robust protein sequence design with CarbonDesign. Nature Machine Intelligence. 2024;6(5):536–47. Available from: https://doi.org/10.1038/s42256-024-00838-2

4.

Ruffolo JA, Bhatnagar A, Beazer J, Nayfach S, Russ J, Hill E, et al. Adapting protein language models for structure-conditioned design. openRxiv; 2024. Available from: https://doi.org/10.1101/2024.08.03.606485

5.

Høie MH, Hummer AM, Olsen TH, Aguilar-Sanjuan B, Nielsen M, Deane CM. AntiFold: improved structure-based antibody design using inverse folding. Bioinformatics Advances. 2025;5(1):vbae202. Available from: https://doi.org/10.1093/bioadv/vbae202