SimpleDesign: A Joint Model for Protein Sequence and Structure Codesign

September 11, 2026 | Source: Apple ML Research

Tags: Apple ML Research, protein design, drug discovery, multimodal, TMLR, protein codesign, generative models

Apple researchers challenge the standard two-stage protein design pipeline with SimpleDesign: trained end-to-end directly on 2M+ sequence-structure pairs, it matches multi-stage latent-space models on codesign benchmarks — suggesting the autoencoder pretraining stage most prior work depends on is unnecessary.

Details

Protein codesign — jointly generating a protein's amino acid sequence and 3D structure — is foundational to drug discovery and protein engineering. The dominant computational approach has been multi-stage: first train an autoencoder or tokenizer to compress protein data, then train a generative model in that latent space. Apple researchers published SimpleDesign in TMLR (September 2026) to challenge this assumption. SimpleDesign trains in a single stage directly in data space, combining discrete cross-entropy loss for amino acid sequences with a regression objective for structure coordinates. The model uses Transformer backbones that process sequences and structures through modality-specific layers while sharing global self-attention across both modalities — allowing sequence and structure representations to inform each other without collapsing into a single representation. Trained on over 2 million sequence-structure pairs, SimpleDesign achieves competitive results on standard codesign benchmarks and unconditional generation tasks, demonstrating that the two-stage pipeline complexity was not necessary for strong performance. Practically, simplifying the training stack lowers the barrier for researchers building on protein design models. The result follows a broader pattern from the same Apple group (see SimpleFold, 2025) of questioning whether domain-specific architectural complexity borrowed from early landmark papers is actually required.