La-Proteina: Atomistic Protein Generation via Partially Latent Flow Matching

TL;DR

La-Proteina tackles joint generation of protein sequences and fully atomistic structures by combining an explicit α-carbon backbone with per-residue latent representations using a two-stage framework of a conditional VAE and partially latent flow matching. The approach enables scalable generation up to 800 residues and achieves state-of-the-art results in unconditional all-atom design and atomistic motif scaffolding, with strong structural validity and rotamer realism. Key innovations include maintaining explicit backbone modeling while encoding side-chain details into fixed-size latents and employing independent generation schedules for backbone and latent components. The results demonstrate substantial advances in designability, co-designability, and motif scaffolding versatility, unlocking new atomistic design applications and paving the way for more complex protein architectures.

Abstract

Recently, many generative models for de novo protein structure design have emerged. Yet, only few tackle the difficult task of directly generating fully atomistic structures jointly with the underlying amino acid sequence. This is challenging, for instance, because the model must reason over side chains that change in length during generation. We introduce La-Proteina for atomistic protein design based on a novel partially latent protein representation: coarse backbone structure is modeled explicitly, while sequence and atomistic details are captured via per-residue latent variables of fixed dimensionality, thereby effectively side-stepping challenges of explicit side-chain representations. Flow matching in this partially latent space then models the joint distribution over sequences and full-atom structures. La-Proteina achieves state-of-the-art performance on multiple generation benchmarks, including all-atom co-designability, diversity, and structural validity, as confirmed through detailed structural analyses and evaluations. Notably, La-Proteina also surpasses previous models in atomistic motif scaffolding performance, unlocking critical atomistic structure-conditioned protein design tasks. Moreover, La-Proteina is able to generate co-designable proteins of up to 800 residues, a regime where most baselines collapse and fail to produce valid samples, demonstrating La-Proteina's scalability and robustness.

Figures (32)

• Figure 1: La-Proteina consists of encoder $q_\psi$(a), decoder $p_\phi$(b), and joint denoiser $p_\theta$(c). The encoder featurizes the input protein and predicts per-residue latent variables $\mathbf{z}$ of constant dimensionality. Together with the underlying $\alpha$-carbon backbone $\mathbf{x}_{C_\alpha}$, the decoder outputs sequence $\mathbf{s}$ and all other atoms $\mathbf{x}_{\neg C_\alpha}$ and reconstructs the atomistic protein. To facilitate generation of de novo proteins, a partially latent flow model jointly generates novel $\alpha$-carbon backbone structures $\mathbf{x}_{C_\alpha}$ and latents $\mathbf{z}$. The model is trained in two stages and all networks are implemented leveraging the same transformer architecture geffner2025proteina; see details in \ref{['sec:method']}.
• Figure 2: Fully atomistic La-Proteina samples. Numbers denote residue count. All samples co-designable.
• Figure 3: Atomistic Motif Scaffolding.La-Proteina accurately reconstructs the atomistic motif (red), while generating diverse scaffolds. Visualization overlays generated protein and motif.
• Figure 4: La-Proteina's strong performance for unconditional long length generation.La-Proteina produces co-designable and diverse proteins of over 500 residues, where all all-atom baselines collapse, yielding no co-designable samples. Left plots show backbone metrics (designability, diversity) against backbone and all-atom baselines; right plots show all-atom metrics (all-atom codesignability, diversity). Metrics detailed in \ref{['app:metrics']}.
• Figure 5: La-Proteina produces structures with higher structural validity than existing all-atom generation baselines. MolProbity davis2007molprobity metrics assessing structural quality: overall MP score, clash score, Ramachandran angle outliers, and covalent bond outliers (details in \ref{['app:metrics']}). P(all-atom) limited to 500 residues; generating longer proteins is computationally prohibitive, requiring over 140GB of GPU memory to produce a single sample.