[Paper Review] Improved motif-scaffolding with SE(3) flow matching

Jason Yim, Andrew M. Campbell+10 more|PubMed|Jan 8, 2024

Protein Structure and Dynamics15 references8 citations

TL;DR

The paper extends FrameFlow with motif-conditioned amortization and motif guidance to perform motif-scaffolding, achieving higher scaffold diversity while maintaining strong designability compared to RFdiffusion.

ABSTRACT

Protein design often begins with the knowledge of a desired function from a motif which motif-scaffolding aims to construct a functional protein around. Recently, generative models have achieved breakthrough success in designing scaffolds for a range of motifs. However, generated scaffolds tend to lack structural diversity, which can hinder success in wet-lab validation. In this work, we extend FrameFlow, an SE(3) flow matching model for protein backbone generation, to perform motif-scaffolding with two complementary approaches. The first is motif amortization, in which FrameFlow is trained with the motif as input using a data augmentation strategy. The second is motif guidance, which performs scaffolding using an estimate of the conditional score from FrameFlow without additional training. On a benchmark of 24 biologically meaningful motifs, we show our method achieves 2.5 times more designable and unique motif-scaffolds compared to state-of-the-art. Code: https://github.com/microsoft/protein-frame-flow.

Motivation & Objective

• Improve motif-scaffolding by extending SE(3) flow matching to two strategies: motif amortization and motif guidance.
• Compare conditional (amortized) and unconditional (guidance) approaches on a PDB-based motif-scaffolding benchmark.
• Show that FrameFlow variants achieve equivalent or better designability with greater scaffold diversity than RFdiffusion.
• Demonstrate that the lightweight FrameFlow model requires fewer parameters and training resources than current state-of-the-art methods.

Proposed method

• Extend FrameFlow with motif conditioning to generate scaffolds around a given motif (motif amortization).
• Develop motif guidance by conditioning the sampling trajectory on the motif using an unconditional FrameFlow model (no extra training).
• Model SE(3) backbone representations via Riemannian flow matching with a decomposed vector field over translations and SO(3) rotations.
• Train motif-amortized FrameFlow with motif data augmentation to simulate motif distributions from unlabeled PDBs.
• Use 500 timesteps Euler-Maruyama sampling for generation and evaluate using designability and diversity metrics.
• Compare against RFdiffusion and TDS on a motif-scaffolding benchmark.

Figure 1: We present two strategies for motif-scaffolding. Top : motif amortization trains a flow model to condition on the motif (blue) and generate the scaffold (red). During training, only the scaffold is corrupted with noise. Bottom : motif guidance re-purposes a flow model that is trained to ge

Experimental results

Research questions

• RQ1 Can motif amortization or motif guidance improve motif-scaffolding performance over prior state-of-the-art (RFdiffusion) for SE(3) flow matching?
• RQ2 Do conditional (amortized) and unconditional (guidance) approaches differ in designability and scaffold diversity?
• RQ3 Does FrameFlow deliver more unique, designable scaffolds per motif while remaining lightweight and trainable?
• RQ4 How do unconditional backbone results support the reliability of motif-scaffolding with FrameFlow-guidance?

Key findings

• FrameFlow with motif amortization solves 21 motifs versus 20 for RFdiffusion on the benchmark.
• FrameFlow-guidance solves 20 motifs, matching RFdiffusion performance.
• Across motifs, FrameFlow methods yield 2.5x more unique designable scaffolds than RFdiffusion.
• Unconditional FrameFlow provides comparable designability to RFdiffusion while achieving higher diversity and novelty.
• FrameFlow-conditioned approach uses a 3x smaller network (16.8M vs 59.8M parameters) and no pre-training.
• Motif-scaffolding with FrameFlow achieves larger scaffold diversity than RFdiffusion and TDS in aggregate.

Figure 2: Motif data augmentation. Each protein in the dataset does not come with pre-defined motif-scaffold annotations. Instead, we construct plausible motifs at random to simulate sampling from the distribution of motifs and scaffolds.