OpenAI’s new LLM exposes the secrets of how AI really works. 谷歌DeepMind的AlphaFold五周年，從根本上重塑了結構生物學，加速了全球研究，並擴展了其模擬整個分子系統的能力。這項工作獲得了認可，德米斯·哈薩比斯和約翰·詹珀共同榮獲2024年諾貝爾化學獎。

 

The experimental model won't compete with the biggest and best, but it could tell us why they behave in weird ways—and how trustworthy they really are.

technologyreview.com

OpenAI’s new LLM exposes the secrets of how AI really works

人工智慧概覽

自2020年11月谷歌DeepMind的AlphaFold成功解決蛋白質折疊難題以來，五年間，它從根本上重塑了結構生物學，加速了全球研究，並擴展了其模擬整個分子系統的能力。這項工作獲得了認可，德米斯·哈薩比斯和約翰·詹珀共同榮獲2024年諾貝爾化學獎。

五年來的主要進展

AlphaFold 2的突破（2020年）：AlphaFold 2能夠根據胺基酸序列準確預測蛋白質結構。這曾是科學家50年來面臨的挑戰。此前，確定一個蛋白質結構需要數年時間。

公共資料庫發布（2021-2022）：GoogleDeepMind與歐洲生物資訊研究所（EMBL-EBI）合作推出了AlphaFold蛋白質結構資料庫。這個免費資料庫包含超過2億個蛋白質結構的預測結果，幾乎涵蓋了所有已知蛋白質。全球超過300萬研究人員正在使用該資料庫。

AlphaFold 3 簡介（2024）：AlphaFold 3 正式發表。此次更新將人工智慧的功能從蛋白質擴展到 DNA、RNA 和小分子。這使得研究人員能夠模擬分子複合物的相互作用，這對藥物研發至關重要。

人工智慧合作科學家及未來願景：Google DeepMind 正在開發「人工智慧合作科學家」。該系統基於 Gemini 2.0 構建，旨在成為虛擬的研究合作者。其目標是產生並探討假設，從而加速藥物發現。長期目標是模擬完整的細胞系統，以變革醫學和生物學。

對科學研究的影響

加速研究：AlphaFold 已被超過 35,000 篇科學論文引用，並縮短了研究週期。使用 AlphaFold 的實驗室向公共資料庫提交新的實驗性蛋白質結構的可能性提高了 40%。

疾病理解與藥物發現：此工具有助於理解生物系統和疾病。這包括揭示與心臟病密切相關的載脂蛋白B-100的結構，以及為蜜蜂免疫研究提供資訊。

結構生物學的普及化：AlphaFold 提供免費且易於使用的工具，使包括本科生在內的研究人員能夠學習該領域並發表研究論文。

AlphaFold 已成為生物學的基礎架構，加速了科學發現。

AI Overview

Five years after its landmark achievement in solving the protein folding problem in November 2020, 

Google DeepMind's AlphaFold has fundamentally reshaped structural biology, accelerated global research, and expanded its capabilities to model entire molecular systems. The work was recognized with the 2024 Nobel Prize in Chemistry shared by Demis Hassabis and John Jumper. 

Key Developments Over Five Years

• AlphaFold 2 Breakthrough (2020): AlphaFold 2 accurately predicted protein structures from amino acid sequences. This was a challenge for scientists for 50 years. Determining a single protein structure previously took years of work.
• Public Database Launch (2021-2022): Google DeepMind, with the European Bioinformatics Institute (EMBL-EBI), launched the AlphaFold Protein Structure Database. This free database contains predictions for over 200 million protein structures, including almost all known proteins. Over 3 million researchers worldwide use the database.
• AlphaFold 3 Introduction (2024): AlphaFold 3 was released. This expanded the AI's capabilities beyond proteins to include DNA, RNA, and small molecules. This allows researchers to model how molecular complexes interact, which is important for drug discovery.
• AI Co-Scientist and Future Ambition: Google DeepMind is developing an "AI co-scientist." This is a system built on Gemini 2.0 designed to be a virtual research collaborator. The goal is to generate and debate hypotheses to accelerate discovery. The long-term goal is to simulate complete cellular systems to transform medicine and biology. 

Impact on Scientific Research

• Accelerated Research: AlphaFold has been cited in over 35,000 scientific papers and has reduced research timelines. Labs using AlphaFold are 40% more likely to submit new experimental protein structures to public databases.
• Disease Understanding and Drug Discovery: The tool has helped in understanding biological systems and diseases. This includes revealing the structure of the apoB-100 protein central to heart disease and informing research on honeybee immunity.
• Democratization of Structural Biology: AlphaFold has provided free, accessible tools, allowing researchers, including undergraduates, to learn the field and publish research papers. 

AlphaFold has become a foundational infrastructure for biology, speeding up scientific discovery.