Apple Steps into Computational Biology with a New AI Model, SimpleFold, Aiming to Compete with Google DeepMind’s AlphaFold. The goal is to predict the 3D structures of proteins more efficiently while reducing computational power requirements.
Amazon co-founder MacKenzie Scott has donated over $19 billion to charity in just five years
Diamond batteries powered by nuclear waste promise 28,000 years of clean energy
A Major Scientific Challenge
For decades, unraveling the three-dimensional structure of proteins from their amino acid sequences has been a significant scientific challenge. Until recently, these predictions could take months or even years of computation. The introduction of AlphaFold by Google DeepMind has transformed the field: the task can now be completed in a matter of hours, sometimes minutes, paving the way for new drug discoveries and material innovations.
However, this achievement comes at a price: AlphaFold2, along with other models like RoseTTAFold and ESMFold, relies on highly complex and resource-intensive architectures. These models incorporate elaborate mechanisms such as multiple sequence alignments (MSA), pair representations, and triangular updates, which encode current scientific knowledge into the model.
SimpleFold: A Radically Different Approach
Apple has taken a different path. With SimpleFold, its researchers have moved away from these specialized mechanisms to embrace flow matching models, an evolution of diffusion models used in image and video generation.
Instead of removing noise step-by-step as in traditional models, the flow matching models chart a direct path from random noise to the final structure, significantly speeding up calculation and reducing power consumption.
Apple has tested SimpleFold across different scales (from 100M to 3B parameters) on two benchmark references, CAMEO22 and CASP14. Results indicate that SimpleFold competes with the most advanced models, achieving 95% of the performance of AlphaFold2 and RoseTTAFold2. Even the smallest version (100M) surpasses 90% of the performance of ESMFold, while being much more resource-efficient.
A Step Toward More Accessible AI in Biology
For Apple, SimpleFold is just the beginning. Researchers emphasize the model’s potential to scale with size and training data, and hope that this approach will set a new benchmark for developing faster, less costly, and equally powerful models.
While AlphaFold has already revolutionized structural biology, the entry of players like Cupertino indicates that competition now extends well beyond academic labs and specialized AI giants. The future of biomedical discovery may rely on this new generation of hybrid models that can merge scientific accuracy with computational efficiency.