Paper List

Journal: ArXiv Preprint
Published: Unknown
BioinformaticsComputational Biology

Consistent Synthetic Sequences Unlock Structural Diversity in Fully Atomistic De Novo Protein Design

NVIDIA | Mila - Quebec AI Institute | Université de Montréal | HEC Montréal | CIFAR AI Chair

Danny Reidenbach, Zhonglin Cao, Zuobai Zhang, Kieran Didi, Tomas Geffner, Guoqing Zhou, Jian Tang, Christian Dallago, Arash Vahdat, Emine Kucukbenli, Karsten Kreis
Figure
Figure
Figure
Figure
Figure

The 30-Second View

IN SHORT: This paper addresses the core pain point of low sequence-structure alignment in existing synthetic datasets (e.g., AFDB), which severely limits the performance of fully atomistic protein generative models.

Innovation (TL;DR)

  • Methodology Introduces a novel high-quality synthetic dataset (D_SYN-ours, ~0.46M samples) by leveraging ProteinMPNN for sequence generation and ESMFold for refolding, ensuring aligned and recoverable sequence-structure pairs.
  • Methodology Proposes Proteína-Atomística, a unified multi-modal flow-based framework that jointly models the distribution of Cα backbone atoms, discrete amino acid sequences, and non-Cα side-chain atoms in explicit observable space without latent variables.
  • Biology Demonstrates that consistent synthetic sequences are critical for unlocking structural diversity, with retrained La-Proteína achieving +54% structural diversity and +27% co-designability, and Proteína-Atomística achieving +73% structural diversity and +5% co-designability.

Key conclusions

  • Only 19.1% of the Foldseek-clustered AFDB dataset (D_AFDB-clstr) meets the standard 2Å all-atom RMSD co-designability threshold when refolded with ESMFold, revealing severe sequence-structure misalignment.
  • Training on the new aligned dataset D_SYN-ours boosts La-Proteína's performance by +54% in structural diversity and +27% in co-designability, setting a new state-of-the-art.
  • The proposed Proteína-Atomística framework, when trained on D_SYN-ours, shows a dramatic +73% improvement in structural diversity and a +5% improvement in co-designability, validating the dataset's broad utility.
Background and Gap： The field faces a bottleneck where large-scale synthetic datasets like the AlphaFold Database (AFDB) contain sequence-structure pairs with low co-designability (only 19.1% recoverable by ESMFold), creating misalignment that hinders the training of high-performance, fully atomistic joint generative models.

Abstract: High-quality training datasets are crucial for the development of effective protein design models, but existing synthetic datasets often include unfavorable sequence-structure pairs, impairing generative model performance. We leverage ProteinMPNN, whose sequences are experimentally favorable as well as amenable to folding, together with structure prediction models to align high-quality synthetic structures with recoverable synthetic sequences. In that way, we create a new dataset designed specifically for training expressive, fully atomistic protein generators. By retraining La-Proteína, which models discrete residue type and side chain structure in a continuous latent space, on this dataset, we achieve new state-of-the-art results, with improvements of +54% in structural diversity and +27% in co-designability. To validate the broad utility of our approach, we further introduce Proteína-Atomística, a unified flow-based framework that jointly learns the distribution of protein backbone structure, discrete sequences, and atomistic side chains without latent variables. We again find that training on our new sequence-structure data dramatically boosts benchmark performance, improving Proteína-Atomística’s structural diversity by +73% and co-designability by +5%. Our work highlights the critical importance of aligned sequence-structure data for training high-performance de novo protein design models. All data will be publicly released.