© ROOT-NATION.com - Use of content is permitted with a backlink.
In a recent scientific study, Apple researchers presented SimpleDesign – an optimized AI model capable of simultaneously generating protein sequences and their structures.
Last September, Apple researchers published a paper titled “SimpleFold: Folding Proteins is Simpler than You Think, ” in which they described a simplified approach to predicting a protein’s three-dimensional structure based on its amino acid sequence.
In short, SimpleFold uses a flow-matching model to generate a protein’s three-dimensional architecture directly from its amino acid sequence. The flow-matching technology starts from a random, noisy baseline and constructs a relatively straight path to the final result. This differs from diffusion models, which typically operate by gradually removing noise until the final state is reached. Both methods are most commonly associated with image generation, although researchers have also explored diffusion models for generating text and code. Returning to SimpleFold, Apple’s developers combined flow matching with universal Transformer blocks, allowing the model to avoid some of the computationally intensive techniques characteristic of traditional protein folding models, such as DeepMind’s well-known AlphaFold.
Now, Apple researchers have introduced SimpleDesign, which extends the idea of simple, universal architectures to the broader task of protein design, rather than just predicting a protein’s spatial structure.
As Apple researchers note in their new study, “SimpleDesign: A Joint Model for Protein Sequence and Structure Codesign,” existing models often rely on a multi-stage training process. In the first stage, specialized algorithms first convert complex data into simpler internal codes, and in the second stage, the main model learns to generate new results based on these codes. The authors hypothesized that such multi-stage training is not necessary for creating effective co-design models and proposed SimpleDesign – an effective multimodal protein design model trained directly in the data space.
In other words, unlike many existing solutions that rely on complex procedures, SimpleDesign learns to generate amino acid sequences and continuous 3D structures through a single end-to-end process. Many modern alternatives first train a separate component to convert protein structures into discrete representations or “tokens,” and only then run the generative model.
SimpleDesign bypasses this intermediate step by training directly on the linked amino acid sequences and 3D coordinates, rather than first compressing the structures into a tokenized format.
SimpleDesign’s training method proved to be quite unconventional. The team utilized over 2 million pairs of protein sequences and structures, primarily from the AFESM dataset, which combines predicted structures from the AlphaFold database with additional samples. During training, both components of each pair were artificially distorted: amino acids in the sequence were hidden behind masked tokens, and noise was added to the 3D structure.
The researchers varied the degree of corruption for each element. If the sequence remained nearly intact while the structure underwent significant distortions, the task resembled protein folding, where the model reconstructed the architecture based on a known sequence. When, however, the structure remained clear and the sequence was masked, the scenario resembled reverse folding: in this case, the model had to generate a sequence capable of forming a given configuration. When both sides were partially damaged, the system simultaneously addressed both problems, training for collaborative protein design.
According to the study, SimpleDesign demonstrated competitive results in benchmarks for co-design, structure generation, and sequence generation, despite a much simpler training pipeline. The authors also found that the developed model is capable of generating plausible protein structures, and the resulting amino acid sequences are on par with or even superior to the results of most competing multimodal models.
Finally, the researchers noted that SimpleDesign’s achievements are currently limited to computational evaluations, as the generated proteins have not been experimentally tested for actual folding, functionality, or safety in living biological systems. However, the results obtained look extremely promising, and the full scientific paper contains a much deeper analysis of the architecture, the training process, benchmarks, and results.