一项最新研究揭示,Anthropic旗下的Claude系列模型(包括Opus 5.5和Fable 5.2)在训练过程中包含了大规模的蛋白质序列数据 1。该发现表明,最新的Claude模型具备类似专业蛋白质语言模型的序列学习能力 1。相比之下,GPT、Gemini及其他主流开源大语言模型在此任务上表现接近随机猜测,未显示出类似迹象 1。此外,这种知识能泛化到训练截止日期后发布的新蛋白质家族中 1。
为了验证这一假设,研究人员构建了包含来自167个Pfam功能未知家族的668个细菌和古菌蛋白质的测试集 1。实验结果显示,Claude Fable 5.2在真实蛋白质序列上预测下一个氨基酸的正确率为54%,而在打乱顺序的序列上仅为9% 1。相比之下,GPT系列、Gemini以及包括DeepSeek V4.1 Flash在内的七个开源权重模型在该测试中的正确率接近零 1。值得注意的是,Claude Opus 5.5的表现与最大的开源蛋白质语言模型ProtGPT3-10B相当 1。即便面对假设训练截止日期(2026年7月1日)之后发布的新型蛋白质家族,Claude模型仍能保持较高的预测能力 1。
New research published on November 1, 2026, indicates that Anthropic’s latest large language models, specifically Opus 5.5 and Fable 5.2, have been trained on extensive protein sequence data 1. The study demonstrates that these models can accurately predict the next amino acid in bacterial and archaeal proteins that are poorly studied, a capability not observed in other leading LLMs 1. Researchers Xinming Tu and Leo Tianlai Chen conducted the analysis to verify whether Claude models possess sequence learning abilities comparable to specialized protein language models 1.
The test set comprised 668 proteins from 167 Pfam families with unknown functions, drawn from bacteria and archaea 1. Results showed that Claude Fable 5.2 achieved a 54% accuracy rate when predicting the next amino acid in real protein sequences, compared to only 9% for shuffled sequences 1. This significant gap confirms that the model has learned genuine biological patterns rather than random associations 1. Furthermore, Claude Opus 5.5 performed on par with ProtGPT3-10B, one of the largest open-source protein language models available 1.
In contrast, GPT series models, Gemini, and seven open-weight models including DeepSeek V4.1 Flash scored near zero on the same task, behaving similarly to random guessing 1. These findings suggest that while other frontier models lack this specific training component, Anthropic has incorporated large-scale protein datasets into its training pipeline 1. Notably, the Claude models maintained high predictive performance even for new protein families released after the assumed training cutoff date of July 1, 2026, indicating robust generalization capabilities 1.
评论
还没有评论,欢迎留下第一条。