SimpleDesign: A New Paradigm for Joint Protein Sequence and Structure Design

The Core Challenge of Protein Design

Proteins are the fundamental building blocks of life, and their function is determined by the complex interplay between amino acid sequences and three-dimensional spatial structures. To truly understand proteins, both dimensions must be grasped simultaneously — the sequence dictates how a protein folds, while the resulting structure in turn determines its function. This inherent multimodal relationship is one of the most difficult problems to crack in protein engineering and drug discovery.

Developing generative models capable of understanding the dual relationship between sequence and structure is critical for applications in drug development, enzyme design, antibody engineering, and more. An ideal model must not only generate plausible amino acid sequences, but also ensure that their corresponding spatial conformations are physically feasible and functionally viable. SimpleDesign is proposed as a solution to exactly this joint design (codesign) challenge.

The Multi-Stage Training Bottleneck in Existing Approaches

Most mainstream protein generative models rely on a multi-stage training pipeline. The typical approach works as follows:

In the first stage, an autoencoder is trained to tokenize sequence and structure data separately into latent representations. The goal of this step is to compress high-dimensional, heterogeneous protein information into a more tractable latent space.

In the second stage, a generative model is trained on top of the latent representations produced by the autoencoder to learn the data distribution. In other words, the generative model does not work directly with raw protein data — it operates in the compressed latent space.

While this staged design has been widely adopted in practice, it comes with significant costs: the training pipeline is complex, multiple modules must be optimized separately, errors can accumulate across stages, and the quality of the latent representations directly caps the ceiling of the final generative model. When sequences and structures are encoded separately, the tightly coupled relationship between them may also be artificially severed.

SimpleDesign's Joint Modeling Approach

As its name suggests, SimpleDesign aims to address the co-design of protein sequences and structures through a more streamlined joint model. In contrast to the traditional paradigm that relies on a multi-stage autoencoder plus generative model pipeline, the core value of joint modeling lies in enabling sequences and structures to be understood and generated simultaneously within a single framework.

The direct benefit of this approach is that the multimodal associations between sequence and structure can be captured end-to-end within the model itself, rather than being artificially separated into distinct training stages. For proteins — where sequence and structure are inherently tightly coupled — joint modeling is theoretically more aligned with their fundamental nature.

For researchers, a simplified training pipeline also means lower engineering complexity and better reproducibility. Fewer intermediate steps means fewer opportunities for hyperparameter sensitivity and error propagation.

Implications for Protein Engineering and Drug Discovery

The ability to jointly design protein sequences and structures directly determines the ceiling for numerous downstream applications. In drug discovery, the capacity to generate proteins with both sound structures and functional sequences on demand can accelerate the design of novel biologics, antibodies, and enzymes. In protein engineering, joint design helps explore functional spaces beyond what naturally occurring proteins offer.

The continued advancement of generative AI in this domain is gradually transforming protein design from an experience-driven, trial-and-error experimental process into a computable, generative engineering problem. If the joint modeling direction that SimpleDesign represents can maintain or even improve generation quality while simplifying training, it will have a meaningful methodological impact on the field.

Summary

SimpleDesign addresses the challenge of joint protein sequence and structure design by replacing the traditional multi-stage (autoencoder + generative model) training paradigm with a more streamlined joint model. Its central argument is that protein sequences and structures are inherently multimodal and tightly coupled — they should be co-generated within a unified framework, rather than handled in separate, disconnected stages.

It is worth noting that this article is based on publicly available research abstracts. For a comprehensive technical assessment of the model's specific architecture, experimental metrics, and comparative results, readers should refer to the full paper.