“iFold”，苹果AI新成果

Core Insights - Apple has launched a protein folding model called SimpleFold, which has been informally referred to as "iFold" by users [1][2]. Group 1: Model Overview - SimpleFold utilizes a general Transformer module combined with flow matching generation paradigm, achieving performance comparable to Google's AlphaFold2 with a 3 billion parameter version [2][4]. - The model simplifies the complex processes involved in protein folding, which traditionally required extensive computational resources and specialized designs [4][6]. Group 2: Technical Innovations - The core innovation of SimpleFold lies in the introduction of flow matching generation technology, which allows for a smooth mapping from random noise distribution to protein conformation distribution, enabling one-step generation of atomic coordinates [7]. - The training dataset for SimpleFold consisted of 9 million entries, resulting in multi-scale models ranging from 100 million to 3 billion parameters, with the SimpleFold-3B model achieving 95% of AlphaFold2's performance on the CAMEO22 benchmark [9]. Group 3: Performance Metrics - In the CAMEO22 benchmark, SimpleFold-3B achieved a TM-score of 0.837 and a GDT-TS score of 0.916, indicating strong performance relative to other models [10]. - The model also demonstrated efficiency, with inference times of only two to three minutes for processing 512 residue sequences on a MacBook Pro equipped with the M2 Max chip, significantly faster than traditional models that require hours [11]. Group 4: Research Team Background - The lead author of the research, Yuyang Wang, has a strong academic background in mechanical engineering and machine learning, with experience at Apple focusing on diffusion models [12]. - The corresponding author, Jiarui Lu, has a background in machine learning and has contributed to Apple's open-source project ToolSandbox, showcasing expertise in large model benchmarking [15].