SLAE

SLAE (Strictly Local All-atom Environment) is a framework for learning protein representations directly from atomic-resolution structure, developed by Yidan Chen, Tianfan Lu, Chuqiao Zhao, Hannah K. Wayment-Steele, and Po-Ssu Huang at Stanford University. Most protein representation models are built either from sequence-pretrained language models or from backbone-only structural graphs. Both discard side-chain geometry and fine-grained chemical detail — precisely the features that determine packing, catalysis, and molecular recognition. SLAE is designed to capture that missing information by modeling each residue through its strictly local all-atom neighborhood.

Rather than encoding a global fold or a residue sequence, SLAE represents a protein as a collection of local atomic environments described only by atom types and interatomic geometries. To learn expressive features from these environments, the authors introduce a multi-task autoencoder objective that jointly optimizes coordinate reconstruction, sequence recovery, and energy regression. The result is a latent space that encodes chemical and geometric context at all-atom resolution.

Posted to bioRxiv in October 2025, SLAE demonstrates that a physically grounded, all-atom representation can be transferred across diverse downstream problems. As of this writing the work is a preprint and no public code or model weights have been released.

#Key Features

• Strictly local all-atom encoding: Each residue is represented by its surrounding atomic neighborhood using only atom types and interatomic geometries, capturing side-chain and chemical detail that backbone-only or sequence-only models omit.
• Multi-task autoencoder objective: Training combines coordinate reconstruction, sequence recovery, and energy regression, encouraging the encoder to learn features that are simultaneously geometric, evolutionary, and physical.
• High-fidelity reconstruction: SLAE reconstructs all-atom structures with high accuracy from its latent residue environments, indicating that the compressed representation retains fine structural detail.
• Transferable representations: The learned features achieve state-of-the-art performance across diverse downstream tasks through transfer learning, without task-specific retraining of the representation.
• Chemically informative latent space: The latent space is environmentally sensitive, enabling quantitative assessment of structural quality and smooth interpolation between conformations at all-atom resolution.

#Technical Details

SLAE is an autoencoder that operates on local atomic environments rather than on whole structures or sequences. For each residue, the encoder ingests the identities and relative geometries of nearby atoms and maps them to a latent residue-environment vector; the decoder reconstructs the all-atom coordinates from that latent. The training signal is a composite objective: a coordinate-reconstruction term that enforces geometric fidelity, a sequence-recovery term that ties structure back to amino-acid identity, and an energy-regression term that grounds the representation in physical plausibility. This combination yields a latent space that is both faithful for reconstruction and discriminative for downstream prediction, and that varies smoothly enough to support interpolation between conformational states. The authors report state-of-the-art results across a range of downstream tasks when the frozen or fine-tuned SLAE representation is transferred to them.

#Applications

SLAE targets researchers building structure-aware models for protein science who need representations that reflect side-chain chemistry and local packing rather than backbone geometry alone. Because its latent space is environmentally sensitive, it can be used to score structural quality, compare conformations, and interpolate between states at all-atom resolution — capabilities relevant to model quality assessment, conformational analysis, and feature extraction for property prediction. As a transferable encoder, it can serve as a drop-in feature source for downstream tasks in place of sequence language-model embeddings.

#Impact

SLAE argues that all-atom local environments are a productive substrate for learned protein representations, recovering chemical information that dominant sequence-based and backbone-based approaches leave out. Reaching state-of-the-art across diverse tasks through transfer learning suggests the representation captures generally useful structural and chemical signal rather than task-specific patterns. The work is currently a preprint awaiting peer review, and no code or model weights have been released publicly, which limits independent reproduction and immediate adoption until an implementation becomes available.