Speaking to nature: a deep learning representational model of proteins ushers in protein linguistics

作者
Daniel Bojar
出处
期刊:Synthetic biology [Oxford University Press]
卷期号:4 (1): ysz013-ysz013 被引量:1
标识
DOI:10.1093/synbio/ysz013
摘要

Understanding, modifying and designing proteins require an intimate knowledge of their 3D structure. Even structure-agnostic protein engineering approaches, such as directed evolution, are limited in scope because of the vast potential sequence space and the epistatic effects that multiple mutations have on protein function. To overcome these difficulties, a holistic understanding of sequence–structure–function relationships has to be established. In their recent preprint, members of the Church Group at the Wyss Institute and collaborators describe a novel approach to predicting protein stability and functionality from raw sequence (1). Their representational model UniRep (unified representation), for the first time, demonstrates an advanced understanding of protein features by means of language modeling. Using deep learning techniques, which were recently recognized with the prestigious Turing Award, Alley et al. built a language model for proteins with amino acids as characters based on natural language processing (NLP) techniques. NLP has not only revolutionized our computational understanding of language—think for instance voice-to-text software—but has been coopted for exciting applications in synthetic biology. The recurrent neural network (RNN; a type of neural network which can process sequential inputs such as text) used by Alley et al. was trained by iteratively predicting the next amino acid given the preceding amino acids for the ∼24 million protein sequences contained in the UniRef50 database. The RNN thus gathered implicit knowledge about the context of a given amino acid and higher-level features such as secondary structure. The authors then averaged the protein representation of their RNN at every sequence position to yield a protein language representation they call UniRep. They then extended UniRep by adding representations of the final sequence position of their RNN to generate the more complete representation called ‘UniRep Fusion’, which serves as an overview of the entire protein sequence. UniRep Fusion was then used as an input for a machine learning model to predict protein stability. Notably, this architecture was more accurate than Rosetta, the de facto state-of-the-art for predicting protein stability. Their protein language representation allowed the authors to predict the relative brightness of 64 800 GFP mutants differing in as few as one amino acids. Remarkably, their predicted relative brightness values correlated strongly with experimental observation (r = 0.98). UniRep, as the representation of ∼24 million proteins, captures many phenomena of general importance for protein structure and function. These general features can be complemented by dataset-specific attributes when training on a subset of protein mutants or de novo designed proteins. This approach could for instance be adopted for screening novel proteins generated by deep learning models. Analogous to de novo designed proteins by Rosetta, generating proteins through protein language models might be most advantageous for proteins with radically new functionalities, which are unlikely to be generated by incremental directed evolution. To arrive in this virtual world of protein engineering though, more advances have to be made. It required the authors of UniRep 1 week of GPU usage to train their large model for one epoch (seeing every protein sequence in UniRef50 once). Switching from the redundancy-ridden UniRef50 database (∼24 million sequences) to preUEP (2), a redundancy-reduced protein sequence database (∼8 million sequences), might enable faster training. This reductionist approach might allow for the ‘vocabulary’ of the model to be extended from single amino acids to larger protein fragments, capturing more structural properties. In general, there are a plethora of NLP techniques developed for written languages which might be useful in protein linguistics. One particularly promising concept would be attention (3), the selective focus on sequence stretches far away from each other, which dramatically improves language models. Given that protein language may be considered one of the most natural languages by definition, modern NLP techniques could transform protein linguistics into a potent tool for the study as well as engineering of proteins for the purposes of synthetic biology.

科研通智能强力驱动
Strongly Powered by AbleSci AI
科研通是完全免费的文献互助平台,具备全网最快的应助速度,最高的求助完成率。 对每一个文献求助,科研通都将尽心尽力,给求助人一个满意的交代。
实时播报
汉堡包的应助被幸运的成仁采纳,获得10
刚刚
科研小白发布了新的文献求助10
刚刚
刚刚
上岸发布了新的文献求助10
刚刚
云槿完成签到,获得积分10
1秒前
橙子发布了新的文献求助10
2秒前
xx发布了新的文献求助10
2秒前
2秒前
所所的应助被大气的fgyyhjj采纳,获得10
2秒前
xiaot完成签到,获得积分10
2秒前
徐长硕完成签到,获得积分20
2秒前
大模型的应助被故渊丶采纳,获得10
3秒前
3秒前
领导范儿的应助被欢喜若灵采纳,获得10
4秒前
大模型的应助被红雨灰衣采纳,获得10
4秒前
甜美山灵完成签到 ,获得积分10
5秒前
丘比特的应助被纯情的秋天采纳,获得10
5秒前
聪慧雪糕完成签到,获得积分20
5秒前
阳光的山水完成签到 ,获得积分10
5秒前
徐长硕发布了新的文献求助10
6秒前
ping发布了新的文献求助10
6秒前
LuXiaoyu的应助被oleskarabach采纳,获得10
6秒前
疯狂加载ing的应助被oleskarabach采纳,获得10
7秒前
牧百川发布了新的文献求助10
7秒前
8秒前
打打的应助被Felix0917采纳,获得10
8秒前
xiaot发布了新的文献求助10
8秒前
zzz发布了新的文献求助10
8秒前
9秒前
田様的应助被科研小白采纳,获得10
10秒前
tjzbw发布了新的文献求助10
11秒前
11秒前
研二就毕业完成签到,获得积分10
12秒前
晚霞跌落星河的应助被master采纳,获得30
13秒前
4yjmngbfd发布了新的文献求助10
13秒前
星辰大海的应助被稳重的馒头采纳,获得10
14秒前
16秒前
李健的应助被牧百川采纳,获得10
18秒前
风吹麦浪的应助被ping采纳,获得10
19秒前
天青色等烟雨完成签到,获得积分10
19秒前
高分求助中
(应助此贴封号)【重要!!请各用户(尤其是新用户)详细阅读】【科研通的精品贴汇总】 10000
拉伸高分子熔体中缠结行为的模拟和小角中子散射研究 1700
The World Atlas of Wine 9th edition 1000
奈特基础组织学:配套组织病理学(第4版)(Netter's Essential Histology With Correlated Histopathology) 800
八旗制度の研究 600
Improving the wear resistance of vehicle gear oil by oil soluble LaF3 nanoparticles and typical additives of gear oil 500
Boundaries and Bodhisattvas 500
热门求助领域 (近24小时)
化学 材料科学 医学 生物 纳米技术 计算机科学 工程类 有机化学 化学工程 生物化学 物理 复合材料 内科学 催化作用 光电子学 人工智能 细胞生物学 无机化学 物理化学 数学
热门帖子
关注 科研通微信公众号,转发送积分 8183621
求助须知:如何正确求助?哪些是违规求助? 9572918
关于积分的说明 21828010
捐赠科研通 7780809
什么是DOI,文献DOI怎么找? 3434031
关于科研通互助平台的介绍 2558243
邀请新用户注册赠送积分活动 2480496