蛋白質(zhì)工程：從定向進(jìn)化到計(jì)算設(shè)計(jì)

曲戈，朱彤，蔣迎迎，吳邊，孫周通

蛋白質(zhì)工程：從定向進(jìn)化到計(jì)算設(shè)計(jì)

曲戈1*，朱彤2*，蔣迎迎1，吳邊2，孫周通1

1 中國(guó)科學(xué)院天津工業(yè)生物技術(shù)研究所，天津 300308 2 中國(guó)科學(xué)院微生物研究所，北京 100101

定向進(jìn)化通過(guò)建立突變體文庫(kù)與高通量篩選方法，快速提升蛋白的特定性質(zhì)，是目前蛋白質(zhì)工程最為常用的蛋白質(zhì)設(shè)計(jì)改造策略。近十年隨著計(jì)算機(jī)運(yùn)算能力大幅提升以及先進(jìn)算法不斷涌現(xiàn)，計(jì)算機(jī)輔助蛋白質(zhì)設(shè)計(jì)改造得到了極大的重視和發(fā)展，成為蛋白質(zhì)工程新開(kāi)辟的重要方向。以結(jié)構(gòu)模擬與能量計(jì)算為基礎(chǔ)的蛋白質(zhì)計(jì)算設(shè)計(jì)不但能改造酶的底物特異性與熱穩(wěn)定性，還可從頭設(shè)計(jì)具有特定功能的人工酶。近年來(lái)機(jī)器學(xué)習(xí)等人工智能技術(shù)也被應(yīng)用于計(jì)算機(jī)輔助蛋白質(zhì)設(shè)計(jì)改造，并取得矚目的成績(jī)。文中介紹了蛋白質(zhì)工程的發(fā)展歷程，重點(diǎn)評(píng)述當(dāng)前計(jì)算機(jī)輔助蛋白質(zhì)設(shè)計(jì)改造方面的進(jìn)展與應(yīng)用，并展望其未來(lái)發(fā)展方向。

蛋白質(zhì)工程，計(jì)算設(shè)計(jì)，結(jié)構(gòu)模擬，能量函數(shù)，人工智能

地球生命歷經(jīng)40億年的自然進(jìn)化，孕育了無(wú)數(shù)功能豐富、結(jié)構(gòu)復(fù)雜的蛋白質(zhì)，但天然的蛋白質(zhì)在穩(wěn)定性、耐受性、選擇性等方面往往無(wú)法滿足工業(yè)生產(chǎn)的需求，促使人類探索高效的蛋白質(zhì)改造方法。蛋白質(zhì)的三維結(jié)構(gòu)決定其生物學(xué)功能，而三維結(jié)構(gòu)又由其內(nèi)在的氨基酸序列決定。蛋白質(zhì)的序列空間極為龐大，以一條100個(gè)氨基酸長(zhǎng)度的蛋白質(zhì)為例，每個(gè)位點(diǎn)可以突變成20種天然氨基酸，它的序列空間達(dá)到20100(約10130)，這一數(shù)字甚至超過(guò)了宇宙中所包含原子的總和[1]。因此，自然界需要花費(fèi)數(shù)百萬(wàn)年來(lái)進(jìn)化得到具有新功能的蛋白質(zhì)，人類要想獲得性能優(yōu)異的酶還得依靠自身的智慧[2]。

20世紀(jì)80年代，聚合酶鏈?zhǔn)椒磻?yīng) (Polymerase chain reaction, PCR) 技術(shù)的出現(xiàn)為人類改造蛋白質(zhì)結(jié)構(gòu)提供了高效的分子操作手段，蛋白質(zhì)工程應(yīng)運(yùn)而生。研究者運(yùn)用PCR技術(shù)在基因特定位點(diǎn)引入突變，從而改變蛋白質(zhì)對(duì)應(yīng)位置的氨基酸殘基種類。最初的蛋白質(zhì)設(shè)計(jì)案例中，由于彼時(shí)缺少蛋白結(jié)構(gòu)與機(jī)理研究，突變位點(diǎn)的選擇完全依靠研究人員的經(jīng)驗(yàn)，因此是一種初級(jí)理性設(shè)計(jì)策略，適用性較窄[3]。在此背景下，定向進(jìn)化 (Directed evolution) 策略誕生，該技術(shù)通過(guò)對(duì)蛋白質(zhì)進(jìn)行多輪突變、表達(dá)和篩選，引導(dǎo)蛋白質(zhì)的性能朝著人們需要的方向進(jìn)化，從而大幅縮短蛋白質(zhì)進(jìn)化的過(guò)程[4]。在這之后，定向進(jìn)化與理性設(shè)計(jì)結(jié)合，形成了半理性設(shè)計(jì) (Semi-rational design) 策略，旨在構(gòu)建“小而精”的突變體文庫(kù)，進(jìn)一步提高效率。近年來(lái)，隨著結(jié)構(gòu)生物學(xué)、計(jì)算生物學(xué)及人工智能技術(shù)的迅猛發(fā)展，計(jì)算機(jī)輔助蛋白質(zhì)設(shè)計(jì) (Computer-assisted protein design, CPD) 策略為蛋白質(zhì)工程領(lǐng)域注入了新的學(xué)術(shù)思想和技術(shù)手段，出現(xiàn)了基于結(jié)構(gòu)模擬與能量計(jì)算來(lái)進(jìn)行蛋白質(zhì)設(shè)計(jì)的新方法，以及使用人工智能 (Artificial intelligence, AI) 技術(shù)指導(dǎo)蛋白質(zhì)改造的新思路?？傮w來(lái)看，蛋白質(zhì)工程經(jīng)歷了從初級(jí)理性設(shè)計(jì)、定向進(jìn)化、半理性設(shè)計(jì)，再到計(jì)算設(shè)計(jì)的發(fā)展歷程 (圖1)。

圖1 蛋白質(zhì)工程發(fā)展歷程

1 定向進(jìn)化與半理性設(shè)計(jì)

1.1 定向進(jìn)化

為了加速蛋白質(zhì)的進(jìn)化，定向進(jìn)化這一策略在20世紀(jì)80–90年代被開(kāi)發(fā)出來(lái)，通過(guò)位點(diǎn)飽和突變(Saturation mutagenesis, SM)、易錯(cuò)PCR (Error-prone polymerase chain reaction, epPCR) 及DNA重組 (DNA shuffling) 等技術(shù)，可有效產(chǎn)生序列多樣性的隨機(jī)突變體文庫(kù)，表達(dá)并篩選特定性狀提高的目標(biāo)突變體。將這一過(guò)程重復(fù)數(shù)輪，通過(guò)連續(xù)積累有益突變，最終可成功獲得性能改進(jìn)或具有新功能的蛋白質(zhì)[3]。定向進(jìn)化是一種工程化的改造思路，不需要事先了解結(jié)構(gòu)信息及催化機(jī)制，通過(guò)迭代有益突變，實(shí)現(xiàn)蛋白質(zhì)性能的飛躍。2018年諾貝爾化學(xué)獎(jiǎng)得主Frances H. Arnold團(tuán)隊(duì)在這一領(lǐng)域作出了杰出貢獻(xiàn)，其團(tuán)隊(duì)通過(guò)使用隨機(jī)突變及單點(diǎn)飽和突變策略改造P450氧化酶，實(shí)現(xiàn)了 碳-硅成鍵[5]、碳-硼成鍵[6]、烯烴反馬氏氧化[7]、卡賓及氮賓的碳-氫鍵插入[8-9]等一系列令人矚目的成果。

定向進(jìn)化技術(shù)在工業(yè)化應(yīng)用領(lǐng)域也扮演了重要角色。例如美國(guó)Codexis公司通過(guò)對(duì)羰基還原酶和鹵醇脫鹵酶進(jìn)行定向改造，實(shí)現(xiàn)降膽固醇藥物立普妥關(guān)鍵手性砌塊的產(chǎn)業(yè)化生產(chǎn)[10]；Codexis聯(lián)合默克公司 (現(xiàn)已合并) 對(duì)轉(zhuǎn)氨酶進(jìn)行多輪定向進(jìn)化，實(shí)現(xiàn)不對(duì)稱氨化合成Ⅱ型糖尿病治療藥物西他列汀的工業(yè)化應(yīng)用[11]；近年來(lái)，華東理工大學(xué)許建和團(tuán)隊(duì)通過(guò)基因挖掘及定向進(jìn)化羰基還原酶，實(shí)現(xiàn)了 ()-硫辛酸的綠色制造工藝[12]。

1.2 半理性設(shè)計(jì)

全隨機(jī)突變策略是以隨機(jī)的方式引入突變，它的瓶頸在于突變體文庫(kù)的規(guī)模非常大，不利于篩選。借助蛋白質(zhì)保守位點(diǎn)及晶體結(jié)構(gòu)分析，通過(guò)非隨機(jī)的方式選取若干個(gè)氨基酸位點(diǎn)作為改造靶點(diǎn)，并結(jié)合有效密碼子的理性選用，構(gòu)建“小而精”的突變體文庫(kù)是克服這一瓶頸的有效方式，這種方式被稱為半理性設(shè)計(jì)。20世紀(jì)90年代，Manfred T. Reetz教授在酶的不對(duì)稱催化改造工作中發(fā)現(xiàn)影響手性選擇的氨基酸位點(diǎn)主要集中在底物結(jié)合口袋區(qū)域，在此基礎(chǔ)上開(kāi)發(fā)了組合活性中心飽和突變策略(Combinatorial active-site saturation test, CAST) 及迭代飽和突變技術(shù) (Iterative saturation mutagenesis, ISM)，廣泛應(yīng)用于酶的立體/區(qū)域選擇性、催化活力、熱穩(wěn)定性等酶參數(shù)的改造[10-15]。例如通過(guò)CAST/ISM策略對(duì)P450-BM3單加氧酶進(jìn)行改造，并與醇脫氫酶或過(guò)氧化物酶偶聯(lián)，使其成功應(yīng)用于高附加值手性二醇及衍生物的不對(duì)稱催化合成[16-17]。最近，Reetz教授與吳起團(tuán)隊(duì)合作，在有效密碼子的選取方面作了改進(jìn)，提出Focused Rational Iterative Site-specific Mutagenesis (FRISM) 策略，并應(yīng)用于南極假絲酵母脂肪酶B (CALB) 的不對(duì)稱催化，成功獲得了雙手性中心底物所對(duì)應(yīng)的全部4種異構(gòu)體，且選擇性均在90%以上[18]。

在CAST基礎(chǔ)上，孫周通等通過(guò)理性選擇 3種氨基酸密碼子作為飽和突變的構(gòu)建單元，開(kāi)發(fā)了三密碼子飽和突變技術(shù)TCSM (Triple code saturationmutagenesis)，進(jìn)一步降低了篩選工作 量[19-20]。除此之外，Gjalt W. Huisman團(tuán)隊(duì)基于統(tǒng)計(jì)學(xué)方法開(kāi)發(fā)的ProSAR (Protein sequence activity relationships)[21]及Miguel Alcalde團(tuán)隊(duì)基于序列同源性開(kāi)發(fā)的MORPHING (Mutagenic organized recombination process by homologousgrouping)[22]工具也廣泛應(yīng)用于蛋白酶的設(shè)計(jì)改造。其他半理性設(shè)計(jì)策略及應(yīng)用已在前文進(jìn)行過(guò)評(píng)述[23]。基于筆者經(jīng)驗(yàn)，半理性設(shè)計(jì)是建立在已有知識(shí)(如保守序列、晶體結(jié)構(gòu)、催化機(jī)制、通量篩選方法、前期實(shí)驗(yàn)數(shù)據(jù)等)的基礎(chǔ)上，對(duì)目標(biāo)蛋白進(jìn)行再設(shè)計(jì)。因此，前期基礎(chǔ)的豐富程度直接影響到半理性設(shè)計(jì)的成功與否。

2 計(jì)算機(jī)輔助蛋白質(zhì)設(shè)計(jì)

2.1 概述

隨著計(jì)算機(jī)運(yùn)算能力持續(xù)提升、先進(jìn)算法相繼涌現(xiàn)，以及蛋白質(zhì)序列特征、三維結(jié)構(gòu)、催化機(jī)制之間關(guān)系不斷被挖掘和解析，計(jì)算機(jī)輔助蛋白質(zhì)設(shè)計(jì)策略得到前所未有的重視和發(fā)展，人類迎來(lái)了蛋白質(zhì)從頭設(shè)計(jì)的時(shí)代[24-25]。蛋白質(zhì)計(jì)算設(shè)計(jì)一般以原子物理、量子物理、量子化學(xué)揭示的微觀粒子運(yùn)動(dòng)、能量與相互作用規(guī)律為理論基礎(chǔ)，也有部分研究以統(tǒng)計(jì)能量函數(shù)為算法依據(jù)。研究者在計(jì)算機(jī)的輔助下，通過(guò)運(yùn)用分子對(duì)接 (Molecular docking)、分子動(dòng)力學(xué)模擬 (Molecular dynamic simulations)、量子力學(xué) (Quantum mechanics) 方法、蒙特卡羅 (Monte Carlo) 模擬退火 (Simulated annealing) 等一系列計(jì)算方法 (相關(guān)方法及使用已有文章綜述[26])，預(yù)測(cè)并評(píng)估數(shù)以千計(jì)的突變體在結(jié)構(gòu)、自由能、底物結(jié)合能等方面的變化?；谟?jì)算結(jié)果，從中篩選可能符合改造要求的突變體并進(jìn)行實(shí)驗(yàn)驗(yàn)證(如突變體能否正常表達(dá)、折疊及行使預(yù)期功能等)；再根據(jù)實(shí)驗(yàn)結(jié)果制定下一輪計(jì)算方案，循環(huán)往復(fù)直到獲得符合需求的酶 (圖2)。與定向進(jìn)化相比，蛋白質(zhì)計(jì)算設(shè)計(jì)可提供明確的改造方案，大幅降低建立、篩選突變體文庫(kù)所需的工作量，目前已在蛋白質(zhì)從頭設(shè)計(jì)、酶的底物選擇性與熱穩(wěn)定性設(shè)計(jì)等方面取得了眾多成果，更有部分成果達(dá)到了工業(yè)應(yīng)用水平[27-28]。

圖2 計(jì)算機(jī)輔助蛋白質(zhì)設(shè)計(jì)流程

2.2 蛋白質(zhì)從頭設(shè)計(jì)

蛋白質(zhì)的從頭設(shè)計(jì)于2016年被雜志列入年度十大科學(xué)突破[29]，其目標(biāo)是創(chuàng)造自然界不存在的、可成功折疊的蛋白質(zhì)并賦予其特定功能，在開(kāi)發(fā)新型病毒疫苗[30-32]、進(jìn)行腫瘤治療[33]等領(lǐng)域發(fā)揮作用。1997年，Stephen L. Mayo團(tuán)隊(duì)提出了包括設(shè)計(jì)、模擬、實(shí)驗(yàn)和分析四步的蛋白質(zhì)設(shè)計(jì)策略PDA (Protein design automation) 并開(kāi)發(fā)了相應(yīng)的軟件，以鋅指結(jié)構(gòu)域?yàn)槟０宄晒υO(shè)計(jì)了一個(gè)由28個(gè)氨基酸組成的ββα蛋白，其核磁檢測(cè)結(jié)構(gòu)與預(yù)測(cè)結(jié)構(gòu)高度一致[34]。2003年，華盛頓大學(xué)的David Baker團(tuán)隊(duì)設(shè)計(jì)并構(gòu)建了一個(gè)非天然結(jié)構(gòu)模板[35]，為蛋白質(zhì)設(shè)計(jì)領(lǐng)域開(kāi)辟了新的方向，其開(kāi)發(fā)的Rosetta軟件如今已發(fā)展為集蛋白質(zhì)從頭設(shè)計(jì)、酶活性中心設(shè)計(jì)、配體對(duì)接、生物大分子結(jié)構(gòu)預(yù)測(cè)等功能為一體的生物大分子計(jì)算建模與分析軟件組合[36]。目前該程序中最常用的工具包括用于設(shè)計(jì)蛋白質(zhì)骨架氨基酸序列的Rosetta Design，以及用于評(píng)估序列變化對(duì)蛋白質(zhì)穩(wěn)定性影響的Rosetta DDG等[37-38]。

表1總結(jié)了最近十年蛋白質(zhì)從頭設(shè)計(jì)的部分案例。大部分案例采用了基于Rosetta的設(shè)計(jì)策略，其中最具代表性的便是Baker團(tuán)隊(duì)提出的“Inside- out”設(shè)計(jì)策略，該策略的主體流程如下：在催化機(jī)理完全明確的前提下，研究者首先運(yùn)用量子化學(xué)方法設(shè)計(jì)酶的活性中心，確定酶的關(guān)鍵催化基團(tuán)與底物形成的過(guò)渡態(tài)構(gòu)象 (Theozyme)；然后使用RosettaMatch搜索蛋白質(zhì)結(jié)構(gòu)數(shù)據(jù)庫(kù)，嘗試將theozyme與已有蛋白質(zhì)結(jié)構(gòu)匹配，篩選能維持theozyme構(gòu)象的蛋白質(zhì)骨架結(jié)構(gòu)；接下來(lái)使用Rosetta Design設(shè)計(jì)位于活性中心但不直接參與催化的氨基酸，運(yùn)用基于蒙特卡羅的模擬退火算法進(jìn)行多輪采樣，獲得經(jīng)過(guò)優(yōu)化的完整酶結(jié)構(gòu)；最后制定評(píng)分標(biāo)準(zhǔn)，依據(jù)過(guò)渡態(tài)能量、配體位置取向等多項(xiàng)參數(shù)評(píng)估設(shè)計(jì)結(jié)果，挑選排名靠前的結(jié)構(gòu)開(kāi)展活性驗(yàn)證實(shí)驗(yàn)[39-40]。運(yùn)用這套策略，Baker團(tuán)隊(duì)成功從頭設(shè)計(jì)了多種酶，其中Retro-Aldol反應(yīng)酶催化的C-C鍵斷裂反應(yīng)速率比無(wú)酶反應(yīng)體系高出4個(gè)數(shù)量級(jí)[41]；Kemp消除反應(yīng)酶催化的消除反應(yīng)速率比無(wú)酶反應(yīng)體系高出5個(gè)數(shù)量級(jí)[42]；Diels-Alder反應(yīng)酶可催化兩個(gè)底物發(fā)生[4+2]雙烯環(huán)加成反應(yīng)，形成具有兩個(gè)手性碳原子的產(chǎn)物，對(duì)映選擇性達(dá)到97%[43]。

表1 最近十年計(jì)算機(jī)輔助的蛋白質(zhì)局部與全局設(shè)計(jì)的部分案例

蛋白質(zhì)序列從頭設(shè)計(jì)的主要困難是計(jì)算模型的精度不夠，導(dǎo)致設(shè)計(jì)成功率低。大量天然蛋白質(zhì)的已知序列結(jié)構(gòu)數(shù)據(jù)可用來(lái)改善模型精度，乃至建立精度更高的新計(jì)算模型。中國(guó)科學(xué)技術(shù)大學(xué)劉海燕團(tuán)隊(duì)發(fā)展了蛋白質(zhì)序列設(shè)計(jì)的統(tǒng)計(jì)能量模型ABACUS (A backbone based amino acid usage survey)。用其對(duì)不同折疊類型的天然蛋白質(zhì)骨架進(jìn)行從頭序列設(shè)計(jì)，實(shí)驗(yàn)證明這些序列全自動(dòng)設(shè)計(jì)的蛋白質(zhì)能夠可溶性表達(dá)并正確折疊，對(duì)Dv_1ubq、D_1cy5_M2兩個(gè)蛋白分子的結(jié)構(gòu)解析結(jié)果表明其與設(shè)計(jì)目標(biāo)高度吻合[67]。之后程序算法進(jìn)一步優(yōu)化，加入了范德華能量函數(shù)，設(shè)計(jì)準(zhǔn)確率再度提高，原本需要其他策略輔助提高折疊成功率的1r26蛋白分子，也能應(yīng)用新一代程序一步設(shè)計(jì)到位[68]。

“Inside-out”設(shè)計(jì)策略針對(duì)的是酶的活性中心以及催化的反應(yīng)類型，而酶的骨架結(jié)構(gòu)仍采用自然界已有的蛋白質(zhì)結(jié)構(gòu)，若骨架結(jié)構(gòu)也能從頭設(shè)計(jì)，便能獲得完全由人類設(shè)計(jì)而非自然進(jìn)化形成的酶[69]。目前，主鏈骨架設(shè)計(jì)還更多依賴于基于“規(guī)則”的啟發(fā)式方法，通用計(jì)算方法仍在發(fā)展之中。在此方面，劉海燕團(tuán)隊(duì)提出了建立氨基酸序列待定、只依賴于主鏈結(jié)構(gòu)的統(tǒng)計(jì)能量模型的設(shè)計(jì)方法，他們已發(fā)表的TetraBASE (Tetrahedron- based backbone statistical energy) 驗(yàn)證了這條途徑原理上是可能的[70]。進(jìn)一步，他們已發(fā)展了基于神經(jīng)網(wǎng)絡(luò)統(tǒng)計(jì)能量項(xiàng)的SCUBA (Sidechain unspecialized backbone arrangement) 模型，能在序列待定情況下進(jìn)行主鏈骨架的全柔性設(shè)計(jì)和連續(xù)采樣，這類方法的發(fā)展將極大拓寬可選蛋白質(zhì)骨架結(jié)構(gòu)范圍。

2.3 酶的底物選擇性與熱穩(wěn)定性設(shè)計(jì)

酶對(duì)底物的特異選擇性一方面避免了各種副反應(yīng)的發(fā)生，但另一方面限制了酶的底物譜。從自然界中挖掘新酶的速度并不能滿足生物技術(shù)產(chǎn)業(yè)快速增長(zhǎng)的需求，因此改變酶的底物特異性是蛋白質(zhì)計(jì)算設(shè)計(jì)的重要應(yīng)用方向[39]。Baker團(tuán)隊(duì)對(duì)鳥(niǎo)嘌呤脫氨酶進(jìn)行改造，運(yùn)用同源建模和量子化學(xué)計(jì)算方法設(shè)計(jì)以三聚氰酸酰胺為底物的活性中心結(jié)構(gòu)，確定關(guān)鍵側(cè)鏈基團(tuán)的位置后優(yōu)化殘基所在的loop區(qū)序列以保證新活性中心結(jié)構(gòu)的穩(wěn)定，在刪除2個(gè)、突變4個(gè)氨基酸殘基后使得酶對(duì)新底物的活性提高約100倍[71]。在另一項(xiàng)工作中，Baker團(tuán)隊(duì)使用Rosetta Design與Foldit改造苯甲醛裂解酶的結(jié)合口袋，使其催化3分子甲醛聚合形成二羥丙酮反應(yīng)的活性提高近100倍，為實(shí)現(xiàn)核心生物代謝分子磷酸二羥丙酮合成途徑的重構(gòu)打下基礎(chǔ)[72]。中國(guó)科學(xué)院微生物研究所吳邊團(tuán)隊(duì)則使用Rosetta Design與高通量分子動(dòng)力學(xué)模擬方法對(duì)天冬氨酸酶的活性中心進(jìn)行設(shè)計(jì)，得到一系列突變體，分別能高效催化巴豆酸、(E)-2-戊烯酸、富馬酸單酰胺、(E)-肉桂酸的β-加氨反應(yīng)，其中巴豆酸β-加氨合成β-氨基丁酸的反應(yīng)底物濃度可達(dá)300 g/L，反應(yīng)轉(zhuǎn)化率大于99%，立體選擇性大于99%，已具有工業(yè)生產(chǎn)的潛力[28]。

工業(yè)酶時(shí)常需要在高溫環(huán)境下發(fā)揮功能，因此提升酶的熱穩(wěn)定性在工業(yè)生產(chǎn)方面具有重要意義。荷蘭格羅寧根大學(xué)的Dick Janssen團(tuán)隊(duì)提出了一種運(yùn)用計(jì)算方法提升酶熱穩(wěn)定性的FRESCO策略 (Framework for rapid enzyme stabilization by computational libraries)，主要流程如下：研究者首先使用Rosetta DDG、FoldX、動(dòng)態(tài)二硫鍵挖掘、保守序列分析等計(jì)算軟件或方法，預(yù)測(cè)能提高蛋白質(zhì)熱穩(wěn)定性的單點(diǎn)突變；然后通過(guò)分子動(dòng)力學(xué)模擬分析單點(diǎn)突變對(duì)酶的影響，刪除化學(xué)角度上不合理或可能提高蛋白質(zhì)結(jié)構(gòu)靈活度的突變方案；接下來(lái)對(duì)剩余單點(diǎn)突變方案進(jìn)行實(shí)驗(yàn)驗(yàn)證，表達(dá)、純化突變后的酶并測(cè)定熔融溫度與催化活性，篩選出熱穩(wěn)定性提高的突變體；最后將一部分單點(diǎn)突變疊加，獲得熱穩(wěn)定性大幅提高的多點(diǎn)突變體[73]。該策略指導(dǎo)下的酶改造過(guò)程中，需要表達(dá)、純化的單點(diǎn)突變體不超過(guò)200個(gè)，單點(diǎn)預(yù)測(cè)成功率大于10%，酶的熔融溫度一般可提升20–35 ℃[74]。Janssen團(tuán)隊(duì)運(yùn)用該策略成功地提升了檸檬烯環(huán)氧化物水解酶[73]、鹵代醇脫鹵酶[75]、鹵代烷脫鹵酶[76]等多種酶的熱穩(wěn)定性，其中以檸檬烯環(huán)氧化物水解酶為代表，該酶的熔融溫度從50 ℃提升至85 ℃，55 ℃下的半衰期延長(zhǎng)250倍，催化活性提高的同時(shí)仍保留反應(yīng)的立體選擇性。吳邊團(tuán)隊(duì)同樣運(yùn)用FRESCO策略，將一種木聚糖酶的熔融溫度提升了14 ℃，改造后的酶在70 ℃下反應(yīng)5 h的產(chǎn)物量相比野生型提高了10倍，為其應(yīng)用于工業(yè)生產(chǎn)打下基礎(chǔ)[77]。

2.4 人工智能技術(shù)在計(jì)算設(shè)計(jì)中的應(yīng)用

得益于計(jì)算速度的大幅提升以及海量數(shù)據(jù)集的出現(xiàn)，當(dāng)前人工智能技術(shù)的發(fā)展如火如荼。在人工智能領(lǐng)域，機(jī)器學(xué)習(xí) (Machine learning) 已經(jīng)成為開(kāi)發(fā)計(jì)算機(jī)視覺(jué)、語(yǔ)音識(shí)別、自然語(yǔ)言處理、機(jī)器人操控和其他應(yīng)用范疇的首選方法[78]。近年來(lái)，機(jī)器學(xué)習(xí)等人工智能方法也被應(yīng)用于蛋白質(zhì)工程，包括Frances H. Arnold、Manfred T. Reetz等定向進(jìn)化先驅(qū)所領(lǐng)導(dǎo)的實(shí)驗(yàn)室均涉足機(jī)器學(xué)習(xí)領(lǐng)域，利用其指導(dǎo)蛋白質(zhì)設(shè)計(jì)改造[79-80]。

蛋白質(zhì)突變體及其對(duì)應(yīng)的實(shí)驗(yàn)數(shù)據(jù)本身是無(wú)法被機(jī)器學(xué)習(xí)算法直接識(shí)別的，其序列、結(jié)構(gòu)、功能等特征 (Feature) 信息必須以向量或數(shù)組的形式展現(xiàn)出來(lái)，才能構(gòu)建被機(jī)器學(xué)習(xí)算法識(shí)別的模型。模型的好壞取決于特征提取。以氨基酸在蛋白序列上的位置信息為特征，是比較常見(jiàn)的處理方式；另外，氨基酸殘基位點(diǎn)的理化性質(zhì) (如帶電性、親疏水性、側(cè)鏈空間體積等) 或所處的二級(jí)結(jié)構(gòu)信息均可作為特征。問(wèn)題在于應(yīng)優(yōu)先選取哪些特征，以及這些特征能在多大程度上決定蛋白質(zhì)擬改造性能是需要進(jìn)行考量的[81]。目前已經(jīng)有一些蛋白質(zhì)/氨基酸特征工具箱可供參考，包括AAIndex[82]、ProFET[83]等。一旦特征提取之后，將交付機(jī)器學(xué)習(xí)算法進(jìn)行學(xué)習(xí)并生成可以描述數(shù)據(jù)模型的目標(biāo)函數(shù)，并對(duì)蛋白質(zhì)序列進(jìn)行虛擬進(jìn)化，通過(guò)訓(xùn)練和測(cè)試評(píng)估效能，最終給出預(yù)測(cè)結(jié)果 (圖3)。

圖3 機(jī)器學(xué)習(xí)指導(dǎo)的蛋白質(zhì)設(shè)計(jì)改造流程

作為人工智能領(lǐng)域常用技術(shù)，機(jī)器學(xué)習(xí)基于大量的數(shù)據(jù)進(jìn)行訓(xùn)練，通過(guò)各種算法解析數(shù)據(jù)并從中學(xué)習(xí)，然后對(duì)處理任務(wù)作出決策。機(jī)器學(xué)習(xí)包括3種：1) 有監(jiān)督學(xué)習(xí) (Supervised learning)，向計(jì)算機(jī)提供原始數(shù)據(jù)及其所對(duì)應(yīng)的結(jié)果 (或稱標(biāo)簽，labels)，最終計(jì)算機(jī)給出定性(分類，classification) 或定量(回歸，regression) 的預(yù)測(cè)；2) 無(wú)監(jiān)督學(xué)習(xí) (Unsupervised learning)，只給計(jì)算機(jī)訓(xùn)練數(shù)據(jù)而不提供結(jié)果，最終得到聚類 (Clustering) 的學(xué)習(xí)結(jié)果；3) 半監(jiān)督學(xué)習(xí) (Semi-supervised learning)，其訓(xùn)練數(shù)據(jù)一部分是有對(duì)應(yīng)的結(jié)果，另一部分則無(wú)結(jié)果。由于蛋白質(zhì)設(shè)計(jì)改造過(guò)程中可產(chǎn)出大量的突變體實(shí)驗(yàn)數(shù)據(jù)，因此有監(jiān)督學(xué)習(xí)應(yīng)用在該領(lǐng)域最為普遍[81]。

目前尚未有任何一種普適的學(xué)習(xí)算法，可以應(yīng)對(duì)所有的學(xué)習(xí)任務(wù)，在蛋白質(zhì)設(shè)計(jì)改造領(lǐng)域亦是如此。因此，科研工作者需要在相應(yīng)的情況下，通過(guò)測(cè)試比對(duì)等方式選用合適的算法進(jìn)行設(shè) 計(jì)[84]。常見(jiàn)算法包括線性模型(Linear models)、隨機(jī)森林(Random forests)、支持向量機(jī)(Support vector machines)、高斯過(guò)程(Gaussian processes) 等。以Frances H. Arnold團(tuán)隊(duì)近期改造一氧化氮雙加氧酶 (NOD) 立體選擇性的工作為例，先后通過(guò)K最近鄰、線性模型、決策樹(shù)、隨機(jī)森林等多個(gè)算法構(gòu)建NOD的立體選擇性催化模型，將76% ()-ee初始突變體提升至93% ()-ee及反轉(zhuǎn)至79% ()-ee[85]。此外，表2列舉了近十幾年來(lái)機(jī)器學(xué)習(xí)指導(dǎo)蛋白質(zhì)設(shè)計(jì)的應(yīng)用實(shí)例，涉及蛋白質(zhì)的熱穩(wěn)定性、催化活性、對(duì)映體選擇性、光敏性及可溶性等多個(gè)方面。從中不難看出，機(jī)器學(xué)習(xí)的可應(yīng)用性非常強(qiáng)，各常用算法均可覆蓋某些蛋白質(zhì)性能的改造。

表2 機(jī)器學(xué)習(xí)指導(dǎo)蛋白質(zhì)設(shè)計(jì)應(yīng)用

除了上述傳統(tǒng)機(jī)器學(xué)習(xí)方法，自2006年以來(lái)，深度學(xué)習(xí)(Deep learning)成為機(jī)器學(xué)習(xí)中的一個(gè)新興領(lǐng)域，如2018年DeepMind團(tuán)隊(duì)開(kāi)發(fā)的AlphaFold在Critical Assessment of protein Structure Prediction (CASP) 全球競(jìng)賽中獲勝，總計(jì)43個(gè)蛋白質(zhì)結(jié)構(gòu)的預(yù)測(cè)結(jié)果中，有25個(gè)獲得了最高分?jǐn)?shù)。

深度學(xué)習(xí)通過(guò)訓(xùn)練深度神經(jīng)網(wǎng)絡(luò)，學(xué)習(xí)由低到高的特征層次，進(jìn)而對(duì)輸入數(shù)據(jù)進(jìn)行分層抽象處理，原始特征數(shù)據(jù)能夠被映射成更高層次和更抽象的數(shù)據(jù)表示，能有效增強(qiáng)辨別能力和減輕無(wú)關(guān)因素的影響，因此深度學(xué)習(xí)深刻變革了機(jī)器學(xué)習(xí)領(lǐng)域[113-114]。相比之下，傳統(tǒng)的學(xué)習(xí)技術(shù) (如支持向量機(jī)、高斯回歸和人工神經(jīng)網(wǎng)絡(luò)等) 則強(qiáng)烈依賴于人工提取的特征，由于它們明確的特征編碼原理，這些方法可能會(huì)丟失隱藏在輸入數(shù)據(jù)中的敏感特征?，F(xiàn)在已有利用卷積神經(jīng)網(wǎng)絡(luò) (Convolutional neural networks, CNN) 和循環(huán)神經(jīng)網(wǎng)絡(luò) (Recurrent neural network, RNN) 進(jìn)行蛋白質(zhì)結(jié)合親和力的預(yù)測(cè)報(bào)道[104,106]。相信深度學(xué)習(xí)未來(lái)可以在蛋白質(zhì)設(shè)計(jì)改造領(lǐng)域扮演更加重要的角色。

3 結(jié)論及展望

自20世紀(jì)80年代以來(lái)，蛋白質(zhì)工程經(jīng)歷了輝煌發(fā)展的30多年，兩次獲得諾貝爾化學(xué)獎(jiǎng) (1993年授予Michael Smith教授及2018年授予Frances H. Arnold教授)。從定向進(jìn)化、半理性設(shè)計(jì)到理性設(shè)計(jì)，每個(gè)階段均涌現(xiàn)了一系列廣泛應(yīng)用的改造策略和技術(shù)，同時(shí)對(duì)計(jì)算技術(shù)的依賴也逐漸加深。定向進(jìn)化不依賴蛋白質(zhì)晶體結(jié)構(gòu)及催化機(jī)制等信息，但存在篩選瓶頸；半理性設(shè)計(jì)兼顧了序列空間多樣性和篩選工作量；而理性設(shè)計(jì)則可以構(gòu)建自然界不存在的新酶新反應(yīng)。在開(kāi)展具體的蛋白質(zhì)工程案例時(shí)，應(yīng)充分考慮到上述因素，基于改造目的靈活選用合適的改造策略[115–116]。

如今，數(shù)據(jù)驅(qū)動(dòng)的人工智能技術(shù)正在全球范圍內(nèi)蓬勃興起，為蛋白質(zhì)設(shè)計(jì)改造注入了新動(dòng)能。蛋白質(zhì)的智能化計(jì)算設(shè)計(jì)是未來(lái)發(fā)展新趨勢(shì)，目前在國(guó)內(nèi)外基本上都處在起步階段。因此，這是我國(guó)在蛋白質(zhì)設(shè)計(jì)改造領(lǐng)域比肩世界先進(jìn)水平的難得機(jī)會(huì)。期待我們能夠把握好這一發(fā)展機(jī)遇，處理好人工智能在蛋白質(zhì)設(shè)計(jì)改造領(lǐng)域的應(yīng)用，通過(guò)開(kāi)發(fā)具有自主知識(shí)產(chǎn)權(quán)的蛋白質(zhì)計(jì)算設(shè)計(jì)新技術(shù)，解決“卡脖子”技術(shù)難題，滿足工業(yè)界綠色、節(jié)能、環(huán)保轉(zhuǎn)型升級(jí)需求，共同譜寫出該領(lǐng)域新的光輝篇章。

[1] Romero PA, Arnold FH. Exploring protein fitness landscapes by directed evolution. Nat Rev Mol Cell Biol, 2009, 10(12): 866–876.

[2] Lynch M, Conery JS. The evolutionary fate and consequences of duplicate genes. Science, 2000, 290(5494): 1151–1155.

[3] Sheldon RA, Pereira PC. Biocatalysis engineering: the big picture. Chem Soc Rev, 2017, 46(10): 2678–2691.

[4] Arnold FH. Directed evolution: bringing new chemistry to life. Angew Chem Int Ed Engl, 2018, 57(16): 4143–4148.

[5] Kan SBJ, Lewis RD, Chen K, et al. Directed evolution of cytochrome c for carbon-silicon bond formation: bringing silicon to life. Science, 2016, 354(6315): 1048–1051.

[6] Kan SBJ, Huang XY, Gumulya Y, et al. Genetically programmed chiral organoborane synthesis. Nature, 2017, 552(7683): 132–136.

[7] Hammer SC, Kubik G, Watkins E, et al. Anti-Markovnikov alkene oxidation by metal-oxo-mediated enzyme catalysis. Science, 2017, 358(6360): 215–218.

[8] Zhang RK, Chen K, Huang XY, et al. Enzymatic assembly of carbon-carbon bonds via iron-catalysed3C-H functionalization. Nature, 2019, 565(7737): 67–72.

[9] Cho I, Jia ZJ, Arnold FH. Site-selective enzymatic C?H amidation for synthesis of diverse lactams. Science, 2019, 364(6440): 575–578.

[10] Ma SK, Gruber J, Davis C, et al. A green-by-design biocatalytic process for atorvastatin intermediate. Green Chem, 2010, 12(1): 81–86.

[11] Savile CK, Janey JM, Mundorff EC, et al. Biocatalytic asymmetric synthesis of chiral amines from ketones applied to sitagliptin manufacture. Science, 2010, 329(5989): 305–309.

[12] Zhang YJ, Zhang WX, Zheng GW, et al. Identification of an ε-keto ester reductase for the efficient synthesis of an ()-α-lipoic acid precursor. Adv Synth Catal, 2015, 357(8): 1697–1702.

[13] Reetz MT, Bocola M, Carballeira JD, et al. Expanding the range of substrate acceptance of enzymes: combinatorial active-site saturation test. Angew Chem Int Ed Engl, 2005, 44(27): 4192–4196.

[14] Reetz MT, Carballeira JD. Iterative saturation mutagenesis (ISM) for rapid directed evolution of functional enzymes. Nat Protoc, 2007, 2(4): 891–903.

[15] Sun ZT, Liu Q, Qu G, et al. Utility of B-factors in protein science: interpreting rigidity, flexibility, and internal motion and engineering thermostability. Chem Rev, 2019, 119(3): 1626–1665.

[16] Li AT, Ilie A, Sun ZT, et al. Whole-cell-catalyzed multiple regio- and stereoselective functionalizations in cascade reactions enabled by directed evolution. Angew Chem Int Ed Engl, 2016, 55(39): 12026–12029.

[17] Yu D, Wang JB, Reetz MT. Exploiting designed oxidase-peroxygenase mutual benefit system for asymmetric cascade reactions. J Am Chem Soc, 2019, 141(14): 5655–5658.

[18] Xu J, Cen YX, Singh W, et al. Stereodivergent protein engineering of a lipase to access all possible stereoisomers of chiral esters with two stereocenters. J Am Chem Soc, 2019, 141(19): 7934–7945.

[19] Sun ZT, Lonsdale R, Ilie A, et al. Catalytic asymmetric reduction of difficult-to-reduce ketones: triple-code saturation mutagenesis of an alcohol dehydrogenase. ACS Catal, 2016, 6(3): 1598–1605.

[20] Sun ZT, Lonsdale R, Wu L, et al. Structure-guided triple-code saturation mutagenesis: efficient tuning of the stereoselectivity of an epoxide hydrolase. ACS Catal, 2016, 6(3): 1590–1597.

[21] Fox RJ, Davis SC, Mundorff EC, et al. Improving catalytic function by ProSAR-driven enzyme evolution. Nat Biotechnol, 2007, 25(3): 338–344.

[22] Gonzalez-Perez D, Molina-Espeja P, Garcia-Ruiz E, et al. Mutagenic organized recombination process by homologousgrouping (MORPHING) for directed enzyme evolution. PLoS ONE, 2014, 9(3): e90919.

[23] Qu G, Zhao J, Zheng P, et al. Recent advances in directed evolution. Chin J Biotech, 2018, 34(1): 1–11 (in Chinese). 曲戈, 趙晶, 鄭平, 等. 定向進(jìn)化技術(shù)的最新進(jìn)展. 生物工程學(xué)報(bào), 2018, 34(1): 1–11.

[24] Samish I. Computational protein design. New York, NY: Humana Press, 2017.

[25] Huang PS, Boyken SE, Baker D. The coming of age ofprotein design. Nature, 2016, 537(7620): 320–327.

[26] Romero-Rivera A, Garcia-Borràs M, Osuna S. Computational tools for the evaluation of laboratory-engineered biocatalysts. Chem Commun, 2016, 53(2): 284–297.

[27] Cui YL, Wu B. Biological components design for engineering requirements. Bull Chin Acad Sci, 2018, 33(11): 1150–1157 (in Chinese). 崔穎璐, 吳邊. 符合工程化需求的生物元件設(shè)計(jì). 中國(guó)科學(xué)院院刊, 2018, 33(11): 1150–1157.

[28] Li RF, Wijma HJ, Song L, et al. Computational redesign of enzymes for regio- and enantioselective hydroamination. Nat Chem Biol, 2018, 14(7): 664–670.

[29] The runners-up. Science, 2016, 354(6319): 1518–1523.

[30] Azoitei ML, Correia BE, Ban YEA, et al. Computation-guided backbone grafting of a discontinuous motif onto a protein scaffold. Science, 2011, 334(6054): 373–376.

[31] Correia BE, Bates JT, Loomis RJ, et al. Proof of principle for epitope-focused vaccine design. Nature, 2014, 507(7491): 201–206.

[32] Fleishman SJ, Whitehead TA, Ekiert DC, et al. Computational design of proteins targeting the conserved stem region of influenza hemagglutinin. Science, 2011, 332(6031): 816–821.

[33] Procko E, Berguig GY, Shen BW, et al. A computationally designed inhibitor of an Epstein-Barr viral Bcl-2 protein induces apoptosis in infected cells. Cell, 2014, 157(7): 1644–1656.

[34] Dahiyat BI, Mayo SL.protein design: fully automated sequence selection. Science, 1997, 278(5335): 82–87.

[35] Kuhlman B, Dantas G, Ireton GC, et al. Design of a novel globular protein fold with atomic-level accuracy. Science, 2003, 302(5649): 1364–1368.

[36] Leaver-Fay A, Tyka M, Lewis SM, et al. ROSETTA3: an object-oriented software suite for the simulation and design of macromolecules. Methods Enzymol, 2011, 487: 545–574.

[37] Das R, Baker D. Macromolecular modeling with rosetta. Annu Rev Biochem, 2008, 77: 363–382.

[38] Alford RF, Leaver-Fay A, Jeliazkov JR, et al. The Rosetta all-atom energy function for macromolecular modeling and design. J Chem Theory Comput, 2017, 13(6): 3031–3048.

[39] Cui YL, Wu B. A brief overview of computational protein structure prediction and enzyme design. Guangxi Sci, 2017, 24(1): 1–6 (in Chinese). 崔穎璐, 吳邊. 計(jì)算機(jī)輔助蛋白結(jié)構(gòu)預(yù)測(cè)及酶的計(jì)算設(shè)計(jì)研究進(jìn)展. 廣西科學(xué), 2017, 24(1): 1–6.

[40] Kiss G, ?elebi–?l?üm N, Moretti R, et al. Computational enzyme design. Angew Chem Int Ed Engl, 2013, 52(22): 5700–5725.

[41] Jiang L, Althoff EA, Clemente FR, et al.computational design of retro-aldol enzymes. Science, 2008, 319(5868): 1387–1391.

[42] R?thlisberger D, Khersonsky O, Wollacott AM, et al. Kemp elimination catalysts by computational enzyme design. Nature, 2008, 453(7192): 190–195.

[43] Siegel JB, Zanghellini A, Lovick HM, et al. Computational design of an enzyme catalyst for a stereoselective bimolecular Diels-Alder reaction. Science, 2010, 329(5989): 309–313.

[44] Grigoryan G, Reinke AW, Keating AE. Design of protein-interaction specificity gives selective bZIP-binding peptides. Nature, 2009, 458(7240): 859–864.

[45] Fry HC, Lehmann A, Saven JG, et al. Computational design and elaboration of aheterotetrameric α-helical protein that selectively binds an emissive abiological (porphinato)zinc chromophore. J Am Chem Soc, 2010, 132(11): 3997–4005.

[46] Korendovych IV, Senes A, Kim YH, et al.design and molecular assembly of a transmembrane diporphyrin-binding protein complex. J Am Chem Soc, 2010, 132(44): 15516–15518.

[47] Grigoryan G, Kim YH, Acharya R, et al. Computational design of virus-like protein assemblies on carbon nanotube surfaces. Science, 2011, 332(6033): 1071–1076.

[48] Karanicolas J, Corn JE, Chen I, et al. Aprotein binding pair by computational design and directed evolution. Mol Cell, 2011, 42(2): 250–260.

[49] Fry HC, Lehmann A, Sinks LE, et al. Computationaldesign and characterization of a protein that selectively binds a highly hyperpolarizable abiological chromophore. J Am Chem Soc, 2013, 135(37): 13914–13926.

[50] Tinberg CE, Khare SD, Dou JY, et al. Computational design of ligand-binding proteins with high affinity and selectivity. Nature, 2013, 501(7466): 212–216.

[51] Huang PS, Oberdorfer G, Xu CF, et al. High thermodynamic stability of parametrically designed helical bundles. Science, 2014, 346(6208): 481–485.

[52] Thomson AR, Wood CW, Burton AJ, et al. Computational design of water-soluble α-helical barrels. Science, 2014, 346(6208): 485–488.

[53] Joh NH, Wang T, Bhate MP, et al.design of a transmembrane Zn2+-transporting four-helix bundle. Science, 2014, 346(6216): 1520–1524.

[54] Wang RY, Kudryashev M, Li XM, et al.protein structure determination from near-atomic-resolution cryo-EM maps. Nat Methods, 2015, 12(4): 335–338.

[55] Lin YR, Koga N, Tatsumi-Koga R, et al. Control over overall shape and size indesigned proteins. Proc Natl Acad Sci USA, 2015, 112(40): E5478–E5485.

[56] Huang PS, Feldmeier K, Parmeggiani F, et al.design of a four-fold symmetric TIM-barrel protein with atomic-level accuracy. Nat Chem Biol, 2016, 12(1): 29–34.

[57] Boyken SE, Chen ZB, Groves B, et al.design of protein homo-oligomers with modular hydrogen-bond network-mediated specificity. Science, 2016, 352(6286): 680–687.

[58] Bhardwaj G, Mulligan VK, Bahl CD, et al. Accuratedesign of hyperstable constrained peptides. Nature, 2016, 538(7625): 329–335.

[59] Dang BB, Wu HF, Mulligan VK, et al.design of covalently constrained mesosize protein scaffolds with unique tertiary structures. Proc Natl Acad Sci USA, 2017, 114(41): 10852–10857.

[60] Polizzi NF, Wu YB, Lemmin T, et al.design of a hyperstable non-natural protein-ligand complex with sub-? accuracy. Nat Chem, 2017, 9(12): 1157–1164.

[61] Chevalier A, Silva DA, Rocklin GJ, et al. Massively parallelprotein design for targeted therapeutics. Nature, 2017, 550(7674): 74–79.

[62] Geiger-Schuller K, Sforza K, Yuhas M, et al. Extreme stability in-designed repeat arrays is determined by unusually stable short-range interactions. Proc Natl Acad Sci USA, 2018, 115(29): 7539–7544.

[63] Dou JY, Vorobieva AA, Sheffler W, et al.design of a fluorescence-activating β-barrel. Nature, 2018, 561(7724): 485–491.

[64] Marcos E, Chidyausiku TM, McShan AC, et al.design of a non-local β-sheet protein with high stability and accuracy. Nat Struct Mol Biol, 2018, 25(11): 1028–1034.

[65] Shen H, Fallas JA, Lynch E, et al.design of self-assembling helical protein filaments. Science, 2018, 362(6415): 705–709.

[66] Silva DA, Yu S, Ulge UY, et al.design of potent and selective mimics of IL-2 and IL-15. Nature, 2019, 565(7738): 186–191.

[67] Xiong P, Wang M, Zhou XQ, et al. Protein design with a comprehensive statistical energy function and boosted by experimental selection for foldability. Nat Commun, 2014, 5: 5330.

[68] Zhou XQ, Xiong P, Wang M, et al. Proteins of well-defined structures can be designed without backbone readjustment by a statistical model. J Struct Biol, 2016, 196(3): 350–357.

[69] Liu HY, Chen Q. Computational protein design for given backbone: recent progresses in general method-related aspects. Curr Opin Struct Biol, 2016, 39: 89–95.

[70] Chu HY, Liu HY. TetraBASE: a side chain-independent statistical energy for designing realistically packed protein backbones. J Chem Inf Model, 2018, 58(2): 430–442.

[71] Murphy PM, Bolduc JM, Gallaher JL, et al. Alteration of enzyme specificity by computational loop remodeling and design. Proc Natl Acad Sci USA, 2009, 106(23): 9215–9220.

[72] Siegel JB, Smith AL, Poust S, et al. Computational protein design enables a novel one-carbon assimilation pathway. Proc Natl Acad Sci USA, 2015, 112(12): 3704–3709.

[73] Wijma HJ, Floor RJ, Jekel PA, et al. Computationally designed libraries for rapid enzyme stabilization. Protein Eng Des Sel, 2014, 27(2): 49–58.

[74] Wijma HJ, Fürst MJLJ, Janssen DB. A computational library design protocol for rapid improvement of protein stability: FRESCO//Bornscheuer UT, H?hne M, Eds. Protein Engineering. New York, NY:Humana Press, 2018: 69–85.

[75] Arabnejad H, Dal Lago M, Jekel PA, et al. A robust cosolvent-compatible halohydrin dehalogenase by computational library design. Protein Eng Des Sel, 2017, 30(3): 175–189.

[76] Floor RJ, Wijma HJ, Colpa DI, et al. Computational library design for increasing haloalkane dehalogenase stability. ChemBioChem, 2014, 15(11): 1660–1672.

[77] Bu YF, Cui YL, Peng Y, et al. Engineering improved thermostability of the GH11 xylanase fromvia computational library design. Appl Microbiol Biotechnol, 2018, 102(8): 3675–3685.

[78] Jordan MI, Mitchell TM. Machine learning: trends, perspectives, and prospects. Science, 2015, 349(6245): 255–260.

[79] Yang KK, Wu Z, Arnold FH. Machine-learning-guided directed evolution for protein engineering. Nat Methods, 2019, 16(8): 687–694.

[80] Li GY, Dong YJ, Reetz MT. Can machine learning revolutionize directed evolution of selective enzymes? Adv Synth Catal, 2019, 361(11): 2377–2386.

[81] Zhou XX, Wang YB, Pan YJ, et al. Differences in amino acidscomposition and coupling patterns between mesophilic and thermophilic proteins. Amino Acids, 2008, 34(1): 25–33.

[82] Kawashima S, Pokarowski P, Pokarowska M, et al. AAindex: amino acid index database, progress report 2008. Nucleic Acids Res, 2008, 36(S1): D202–D205.

[83] Ofer D, Linial M. ProFET: feature engineering captures high-level protein functions. Bioinformatics, 2015, 31(21): 3429–3436.

[84] Wolpert DH. The lack of A priori distinctions between learning algorithms. Neural Comput, 1996, 8(7): 1341–1390.

[85] Wu Z, Kan SBJ, Lewis RD, et al. Machine learning-assisted directed protein evolution with combinatorial libraries. Proc Natl Acad Sci USA, 2019, 116(18): 8852–8858.

[86] Fox R. Directed molecular evolution by machine learning and the influence of nonlinear interactions. J Theor Biol, 2005, 234(2): 187–199.

[87] Fox R, Roy A, Govindarajan S, et al. Optimizing the search algorithm for protein engineering by directed evolution. Protein Eng Des Sel, 2003, 16(8): 589–597.

[88] Liao J, Warmuth MK, Govindarajan S, et al. Engineering proteinase K using machine learning and synthetic genes. BMC Biotechnol, 2007, 7: 16.

[89] Govindarajan S, Mannervik B, Silverman JA, et al. Mapping of amino acid substitutions conferring herbicide resistance in wheat glutathione transferase. ACS Synth Biol, 2015, 4(3): 221–227.

[90] Musdal Y, Govindarajan S, Mannervik B. Exploring sequence-function space of a poplar glutathione transferase using designed information-rich gene variants. Protein Eng Des Sel, 2017, 30(8): 543–549.

[91] Li YQ, Fang JW. PROTS-RF: a robust model for predicting mutation-induced protein stability changes. PLoS ONE, 2012, 7(10): e47247.

[92] Capriotti E, Fariselli P, Calabrese R, et al. Predicting protein stability changes from sequences using support vector machines. Bioinformatics, 2005, 21(S2): ii54–ii58.

[93] Buske FA, Their R, Gillam EMJ, et al.characterization of protein chimeras: relating sequence and function within the same fold. Proteins, 2009, 77(1): 111–120.

[94] Zaugg J, Gumulya Y, Malde AK, et al. Learning epistatic interactions from sequence-activity data to predict enantioselectivity. J Comput Aided Mol Des, 2017, 31(12): 1085–1096.

[95] Saladi SM, Javed N, Müller A, et al. A statistical model for improved membrane protein expression using sequence-derived features. J Biol Chem, 2018, 293(13): 4913–4927.

[96] Pires DEV, Ascher DB, Blundell TL. mCSM: predicting the effects of mutations in proteins using graph-based signatures. Bioinformatics, 2014, 30(3): 335–342.

[97] Romero PA, Krause A, Arnold FH. Navigating the protein fitness landscape with Gaussian processes. Proc Natl Acad Sci USA, 2013, 110(3): E193–E201.

[98] Saito Y, Oikawa M, Nakazawa H, et al. Machine-learning-guided mutagenesis for directed evolution of fluorescent proteins. ACS Synth Biol, 2018, 7(9): 2014–2022.

[99] Bedbrook CN, Yang KK, Rice AJ, et al. Machine learning to design integral membrane channelrhodopsins for efficient eukaryotic expression and plasma membrane localization. PLoS Comput Biol, 2017, 13(10): e1005786.

[100] Bedbrook CN, Yang KK, Robinson JE, et al. Machine learning-guided channelrhodopsin engineering enables minimally-invasive optogenetics. bioRxiv, 2019, doi:10.1101/565606.

[101] Zhang S, Zhou JT, Hu HL, et al. A deep learning framework for modeling structural features of RNA-binding protein targets. Nucleic Acids Res, 2016, 44(4): e32.

[102] Alipanahi B, Delong A, Weirauch MT, et al. Predicting the sequence specificities of DNA- and RNA-binding proteins by deep learning. Nat Biotechnol, 2015, 33(8): 831–838.

[103] Zeng HY, Edwards MD, Liu G, et al. Convolutional neural network architectures for predicting DNA-protein binding. Bioinformatics, 2016, 32(12): i121–i127.

[104] Hu JJ, Liu ZH. DeepMHC: deep convolutional neural networks for high-performance peptide-mhc binding affinity prediction. bioRxiv, 2017, doi:10.1101/239236.

[105] Jiménez J, Doerr S, Martínez-Rosell G, et al. DeepSite: protein-binding site predictor using 3D-convolutional neural networks. Bioinformatics, 2017, 33(19): 3036–3042.

[106] Mazzaferro C. Predicting protein binding affinity with word embeddings and recurrent neural networks. bioRxiv, 2017, doi:10.1101/128223.

[107] Khurana S, Rawi R, Kunji K, et al. DeepSol: a deep learning framework for sequence-based protein solubility prediction. Bioinformatics, 2018, 34(15): 2605–2613.

[108] Giollo M, Martin AJ, Walsh I, et al. NeEMO: a method using residue interaction networks to improve prediction of protein stability upon mutation. BMC Genomics, 2014, 15(S4): S7.

[109] Almagro Armenteros JJ, S?nderby CK, S?nderby SK, et al. DeepLoc: prediction of protein subcellular localization using deep learning. Bioinformatics, 2017, 33(21): 3387–3395.

[110] S?nderby SK, Winther O. Protein secondary structure prediction with long short term memory networks. arXiv preprint arXiv:1412.7828, 2014.

[111] Szalkai B, Grolmusz V. Near perfect protein multi-label classification with deep neural networks. Methods, 2018, 132: 50–56.

[112] Hopf TA, Colwell LJ, Sheridan R, et al. Three-dimensional structures of membrane proteins from genomic sequencing. Cell, 2012, 149(7): 1607–1621.

[113] Bengio Y. Learning deep architectures for AI. Found Trends Mach Learn, 2009, 2(1): 1–127.

[114] LeCun Y, Bengio Y, Hinton G. Deep learning. Nature, 2015, 521(7553): 436–444.

[115] Qu G, Li AT, Acevedo-Rocha CG, et al. The crucial role of methodology development in directed evolution of selective enzymes. Angew Chem Int Ed Engl, 2019, 58, doi: 10.1002/anie.201901491.

[116] Qu G, Zhang K, Jiang YY, et al. The Nobel Prize in chemistry 2018: the directed evolution of enzymes and the phage display technologies. J Biol, 2019, 36(1): 1–6 (in Chinese).曲戈，張錕，蔣迎迎，等. 2018諾貝爾化學(xué)獎(jiǎng)：酶定向進(jìn)化與噬菌體展示技術(shù). 生物學(xué)雜志, 2019, 36(1): 1–6.

Protein engineering: from directed evolution to computational design

Ge Qu1*, Tong Zhu2*, Yingying Jiang1, Bian Wu2, and Zhoutong Sun1

1 Tianjin Institute of Industrial Biotechnology, Chinese Academy of Sciences, Tianjin 300308, China 2 Institute of Microbiology, Chinese Academy of Sciences, Beijing 100101, China

By constructing mutant libraries and utilizing high-throughput screening methods, directed evolution has emerged as the most popular strategy for protein design nowadays. In the past decade, taking advantages of computer performance and algorithms, computer-assisted protein design has rapidly developed and become a powerful method of protein engineering. Based on the simulation of protein structure and calculation of energy function, computational design can alter the substrate specificity and improve the thermostability of enzymes, as well asdesign of artificial enzymes with expected functions. Recently, machine learning and other artificial intelligence technologies have also been applied to computational protein engineering, resulting in a series of remarkable applications. Along the lines of protein engineering, this paper reviews the progress and applications of computer-assisted protein design, and current trends and outlooks of the development.

protein engineering, computational design, simulation of structure, energy function, artificial intelligence

10.13345/j.cjb.190221

孫周通 博士，博士生導(dǎo)師，研究員。2012年博士畢業(yè)于中國(guó)科學(xué)院上海生命科學(xué)研究院；2012––2016年分別在新加坡南洋理工大學(xué)和德國(guó)馬普煤炭化學(xué)研究所/馬爾堡大學(xué)從事博士后研究；2016年至今，中國(guó)科學(xué)院天津工業(yè)生物技術(shù)研究所研究員，課題組長(zhǎng)(PI)，主要從事酶分子工程與工業(yè)生物催化研究。申請(qǐng)中國(guó)發(fā)明專利12項(xiàng)，獲中國(guó)產(chǎn)學(xué)研創(chuàng)新成果獎(jiǎng)一項(xiàng)，在國(guó)際主流期刊、、、、、、等發(fā)表SCI文章20余篇。

曲戈, 朱彤, 蔣迎迎, 等. 蛋白質(zhì)工程：從定向進(jìn)化到計(jì)算設(shè)計(jì). 生物工程學(xué)報(bào), 2019, 35(10): 1843–1856.

Qu G, Zhu T, Jiang YY, et al. Protein engineering: from directed evolution to computational design. Chin J Biotech, 2019, 35(10): 1843–1856.

May29, 2019;

July17, 2019

Supported by: CAS Pioneer Hundred Talents Program (No. 2016-053).

Zhoutong Sun. Tel/Fax: +86-22-84861981; E-mail: sunzht@tib.cas.cn

*These authors contributed equally to this study.

中國(guó)科學(xué)院率先行動(dòng)“百人計(jì)劃”項(xiàng)目 (No. 2016-053) 資助。

2019-08-20

http://kns.cnki.net/kcms/detail/11.1998.Q.20190819.1802.002.html

(本文責(zé)編 陳宏宇)