Abstract
We present Protpardelle-1c, a collection of protein structure generative models with robust motif scaffolding and support for multi-chain complex generation under hotspot-conditioning. Enabling sidechain-conditioning to a backbone-only model increased Protpardelle-1c’s MotifBench score from 4.97 to 28.16, outperforming RFdiffusion’s 21.27. The crop-conditional all-atom model achieved 208 unique solutions on the La-Proteina all-atom motif scaffolding benchmark, on par with La-Proteina while having ~10 times fewer parameters. At 22M parameters, Protpardelle-1c enables rapid sampling, taking 40 minutes to sample all 3000 MotifBench backbones on an NVIDIA A100-80GB, compared to 31 hours for RFdiffusion.
Introduction
In protein engineering, function is inextricably linked to form. A common design challenge is "motif scaffolding": building a stable protein structure around a specific functional site—such as a binding pocket or catalytic center—to maintain its precise geometry. Current state-of-the-art tools often struggle with high computational costs, the neglect of sidechain atoms (which are critical for chemical function), and the difficulty of designing multi-chain complexes. Protpardelle-1c addresses these challenges by providing a lightweight, efficient diffusion model that incorporates sidechain-aware conditioning and optimized training data to surpass existing benchmarks in both accuracy and speed.
Methods
Protpardelle-1c is built on the Elucidated Diffusion Model (EDM) framework. Notably, the model is non-equivariant by design, learning to "pick a frame" during denoising by training on randomly rotated and translated protein structures. This architectural choice, combined with a Diffusion Transformer (DiT) style noise residual, allowed the authors to reduce the parameter count to 22M.
Key innovations include:
- Classifier-Free Guidance: The model attends directly to motif coordinates as conditional input, avoiding the brittleness of traditional reconstruction guidance.
- Sidechain Conditioning: By conditioning on both backbone and sidechain atoms, the model ensures the generated scaffold is chemically compatible with the functional motif.
- Augmented Training Data: The model was trained on AI-CATH (ESMFold-predicted structures redesigned by ProteinMPNN) and MD-CATH (molecular dynamics simulations) to better understand the space of designable and flexible proteins.
Results
Protpardelle-1c demonstrates significant improvements over the industry-standard RFdiffusion. On the MotifBench benchmark, Protpardelle-1c achieved a score of 28.16, compared to 21.27 for RFdiffusion. The model's efficiency is a standout feature: it sampled 3,000 backbones in just 40 minutes on a single NVIDIA A100, whereas RFdiffusion required 31 hours for the same task—a 46x speedup.
Furthermore, the model's ability to generate high-quality all-atom structures is evidenced by its performance on the La-Proteina benchmark, where it produced 208 unique solutions. The model successfully generated binders for complex targets like SpCas9, demonstrating its utility for real-world therapeutic design.
Discussion and Limitations
While Protpardelle-1c sets a new bar for efficiency and sidechain-aware design, limitations remain. Users must currently pre-specify motif indices within the sequence, and the model exhibits a slight bias toward alpha-helical structures, likely due to the nature of the training data. Future work will focus on automated motif placement and diversifying the topological variety of generated scaffolds.
Summary updated