Claude was trained on protein sequences at scale (xinmingtu.cn)

🤖 AI Summary
The latest Claude models, Fable 5.2 and Opus 5.5, have been trained on protein sequences at an unprecedented scale, demonstrating a superior ability to predict amino acids for little-studied proteins compared to other frontier models. In a next-amino-acid test, these Claude models showed an impressive accuracy of 54% on real protein sequences, while struggling to produce valid predictions from shuffled sequences, which indicates a deep understanding of the sequences rather than mere statistical correlation. This performance is significant as it showcases the potential of large language models in the field of bioinformatics, particularly for understanding less-researched protein families that have little to no existing literature. The technical implications are profound, revealing that the model's training included substantial protein sequence data, enabling it to outperform traditional protein language models like ProtGPT3-10B. This adaptability is evident in its handling of protein families with no prior known function and suggests that these models could have broader applications in predicting protein structures or functions, supporting future research in biological systems. Moreover, the findings underscore the limitations of other models, such as GPT and Gemini, which fail to show the same depth of understanding in protein profiling, indicating a significant leap in how AI can facilitate advancements in AI/ML applications within biological sciences.
Loading comments...
loading comments...