---
title: "SimpleDesign: A Joint Model for Protein Sequence and Structure Codesign"
description: "SimpleDesign addresses protein design as a joint modeling problem involving amino acid sequences and three-dimensional structures, whose interplay determines protein function. Unlike approaches that first train autoencoders and then model their latent representations, it uses a single-stage, end-to-end objective directly in data space. The objective combines discrete cross-entropy for sequences with a regression objective for structures, while Transformer-based multimodal backbones provide modality-specific processing and global self-attention across both modalities. Trained on more than 2M sequence-structure pairs, SimpleDesign achieves competitive performance on co-design and unconditional sequence and structure generation benchmarks. The work argues, based on these results, that multi-stage training is not necessary for performant co-design models."
---

# SimpleDesign: A Joint Model for Protein Sequence and Structure Codesign

[Apple](https://yomu.fyi/company/apple) · Sep 11, 2026

**Type:** Problem & solution

## Summary

SimpleDesign addresses protein design as a joint modeling problem involving amino acid sequences and three-dimensional structures, whose interplay determines protein function. Unlike approaches that first train autoencoders and then model their latent representations, it uses a single-stage, end-to-end objective directly in data space. The objective combines discrete cross-entropy for sequences with a regression objective for structures, while Transformer-based multimodal backbones provide modality-specific processing and global self-attention across both modalities. Trained on more than 2M sequence-structure pairs, SimpleDesign achieves competitive performance on co-design and unconditional sequence and structure generation benchmarks. The work argues, based on these results, that multi-stage training is not necessary for performant co-design models.

## Context

Protein function depends on the complex interplay between amino acid sequence and three-dimensional structure. Existing generative approaches often use multi-stage training: autoencoders first tokenize data into latent representations, followed by generative modeling in the resulting latent space.

## Approach / What changed

SimpleDesign is trained directly in data space with a single-stage, end-to-end objective combining discrete cross-entropy for sequences and regression for structures. Transformer-based multimodal backbones provide modality-specific processing while retaining global self-attention across both modalities.

## Takeaways

- SimpleDesign models protein sequences and structures jointly rather than training separate latent-space stages.
- Its training objective combines sequence cross-entropy with structure regression, reflecting the two modalities’ different representations.
- Training on more than 2M sequence-structure pairs produced competitive results on co-design and unconditional sequence and structure generation benchmarks.

**Tags:** [Architecture](https://yomu.fyi/topic/architecture), [Generative AI](https://yomu.fyi/topic/generative-ai), [Multimodal](https://yomu.fyi/topic/multimodal)

- Source: [Apple](https://machinelearning.apple.com/research/simpledesign-protein-codesign)
- Source URL: https://machinelearning.apple.com/research/simpledesign-protein-codesign
- Ingested by Yomu: 2026-09-11T16:03:23.531Z

[Read original post](https://machinelearning.apple.com/research/simpledesign-protein-codesign)