Search papers, labs, and topics across Lattice.
This paper introduces SinAE, a novel flow-matching autoencoder designed to unify the generative modeling of small molecules, crystals, and proteins within a single architecture. By employing a vanilla Transformer encoder and an iterative flow-matching decoder, SinAE achieves near-lossless reconstruction while significantly reducing errors compared to existing latent models. The results demonstrate that joint training across domains enhances performance, showcasing the potential for effective cross-domain transfer in atomic system generation.
SinAE achieves near-lossless reconstruction across diverse atomic systems, drastically improving generative performance while simplifying the architecture.
Small molecules, crystals, and proteins all reduce to atoms in 3D space, yet their generative pipelines remain fragmented across domains, each with its Small molecules, crystals, and proteins all reduce to atoms in 3D space, yet their generative pipelines remain fragmented across domains, each with its own graph, equivariant, or frame-based architecture. Cross-domain training would mitigate per-domain data scarcity, but direct generation in 3D coordinate space cannot easily handle the heterogeneous structural priors of all three domains, and no prior latent autoencoder is simultaneously lossless and architecturally general across all three. We introduce SinAE, a single-architecture flow-matching autoencoder for molecules, crystals, and proteins, with vanilla Transformer encoder and decoder and no equivariant, graph, or domain-specific operators. Rather than requiring the encoder to capture fine-grained geometry, SinAE shifts the reconstruction burden into an iterative flow-matching decoder, achieving near-lossless reconstruction across domains and reducing reconstruction errors by orders of magnitude relative to prior latent baselines. The same per-token latent supports a standard Diffusion Transformer prior that reaches strong performance on molecular, crystal, and protein generation benchmarks. Joint molecule--crystal training strictly improves both domains, providing direct evidence of cross-domain transfer through a shared atomic latent. Code is available at https://github.com/BlueWhaleLab/SinAE .
