麻痹的徐锦波的ultra deep learning模型绝对是alphafold的prior art

[英亲王阿齐格]

“告诉大家用AI 可以大幅度提高蛋白质结构预测精度”

你干啥的？

是不是没什么学历的 炒菜大厨啊

要是没什么学历 就不要谈这种话题

跟你说这些 掉价

[行观曰]

你干啥的？

是不是没什么学历的 炒菜大厨啊

要是没什么学历 就不要谈这种话题

跟你说这些 掉价

我是后厨癫大勺的，所以特别愿意跟你交流技艺，觉得跟你特别谈的来 就像我们后厨洗碗工他们

[行观曰]

你干啥的？

是不是没什么学历的 炒菜大厨啊

要是没什么学历 就不要谈这种话题

跟你说这些 掉价

“告诉大家用AI 可以大幅度提高蛋白质结构预测精度”

而且 John Jumper 是听了徐的讲座的，还和他的博士后有交流

[英亲王阿齐格]

“告诉大家用AI 可以大幅度提高蛋白质结构预测精度”

而且 John Jumper 是听了徐的讲座的，还和他的博士后有交流

The prediction of protein structures has had a long and varied development, which is extensively covered in a number of reviews14,40,41,42,43. Despite the long history of applying neural networks to structure prediction14,42,43, they have only recently come to improve structure prediction10,11,44,45. These approaches effectively leverage the rapid improvement in computer vision systems46 by treating the problem of protein structure prediction as converting an ‘image’ of evolutionary couplings22,23,24 to an ‘image’ of the protein distance matrix and then integrating the distance predictions into a heuristic system that produces the final 3D coordinate prediction. A few recent studies have been developed to predict the 3D coordinates directly47,48,49,50, but the accuracy of these approaches does not match traditional, hand-crafted structure prediction pipelines51. In parallel, the success of attention-based networks for language processing52 and, more recently, computer vision31,53 has inspired the exploration of attention-based methods for interpreting protein sequences54,55,56.

[行观曰]

The prediction of protein structures has had a long and varied development, ... but the accuracy of these approaches does not match traditional, hand-crafted structure prediction pipelines

“告诉大家用AI 可以大幅度提高蛋白质结构预测精度”


显然两个说法不一致。

而跟据2016年之后的A I 进展，上面的英文说法显然在贬低他人的工作

[英亲王阿齐格]

“告诉大家用AI 可以大幅度提高蛋白质结构预测精度”


显然两个说法不一致。

而跟据2016年之后的A I 进展，上面的英文说法显然在贬低他人的工作

你看过paper吗？

没看过去看看，懒得跟你废话吗

你麻痹 哪个paper里 不说大幅改进啊
从你麻痹0.01到0.02也是大幅度改进，必经提高了百分之一百啊。

这他妈的就是套话，你他麻痹的还真信啊
傻逼

[dfc8622]

有些五毛英文都不利索，就是灌纯水带风向的，莫同他们较真

Proc Int Conf Intell Syst Mol Biol
. 1993:1:402-10.
Protein structure prediction system based on artificial neural networks

[行观曰]

你看过paper吗？

没看过去看看，懒得跟你废话吗

你麻痹 哪个paper里 不说大幅改进啊
从你麻痹0.01到0.02也是大幅度改进，必经提高了百分之一百啊。

这他妈的就是套话，你他麻痹的还真信啊
傻逼

傻逼，去跪舔白屌吧

[superdsb]

这是jumper那片nature自己吹牛逼的话
什么几把同行的都lack atomic accuracy
怎么定义的accuracy也没说而且也没比较数据
这种要是换了referee是绝对不会让发出来的
而且AlphaFold自己就没有多高的accuracy

AlphaFold has various limitations:

AlphaFold DB provides monomeric models of proteins, rather than their biologically relevant complexes.[68]
Many protein regions are predicted with low confidence score, including the intrinsically disordered protein regions.[69]
Alphafold-2 was validated for predicting structural effects of mutations with a limited success.[70]
The model relies to some degree upon co-evolutionary information across similar proteins, and thus may not perform well on synthetic proteins or proteins with very low homology to anything in the database.[71]
The ability of the model to produce multiple native conformations of proteins is limited.
AlphaFold 3 version can predict structures of protein complexes with a very limited set of selected cofactors and co- and post-translational modifications.[72] Between 50% and 70% of the structures of the human proteome are incomplete without covalently-attached glycans.[73][68] AlphaFill, a derived database, adds cofactors to AlphaFold models where appropriate.[74]
In the algorithm, the residues are moved freely, without any restraints. Therefore, during modeling the integrity of the chain is not maintained. As a result, AlphaFold may produce topologically wrong results, like structures with an arbitrary number of knots.[75]

The prediction of protein structures has had a long and varied development, which is extensively covered in a number of reviews14,40,41,42,43. Despite the long history of applying neural networks to structure prediction14,42,43, they have only recently come to improve structure prediction10,11,44,45. These approaches effectively leverage the rapid improvement in computer vision systems46 by treating the problem of protein structure prediction as converting an ‘image’ of evolutionary couplings22,23,24 to an ‘image’ of the protein distance matrix and then integrating the distance predictions into a heuristic system that produces the final 3D coordinate prediction. A few recent studies have been developed to predict the 3D coordinates directly47,48,49,50, but the accuracy of these approaches does not match traditional, hand-crafted structure prediction pipelines51. In parallel, the success of attention-based networks for language processing52 and, more recently, computer vision31,53 has inspired the exploration of attention-based methods for interpreting protein sequences54,55,56.

[英亲王阿齐格]

这是jumper那片nature自己吹牛逼的话
什么几把同行的都lack atomic accuracy
怎么定义的accuracy也没说而且也没比较数据
这种要是换了referee是绝对不会让发出来的
而且AlphaFold自己就没有多高的accuracy

AlphaFold has various limitations:

AlphaFold DB provides monomeric models of proteins, rather than their biologically relevant complexes.[68]
Many protein regions are predicted with low confidence score, including the intrinsically disordered protein regions.[69]
Alphafold-2 was validated for predicting structural effects of mutations with a limited success.[70]
The model relies to some degree upon co-evolutionary information across similar proteins, and thus may not perform well on synthetic proteins or proteins with very low homology to anything in the database.[71]
The ability of the model to produce multiple native conformations of proteins is limited.
AlphaFold 3 version can predict structures of protein complexes with a very limited set of selected cofactors and co- and post-translational modifications.[72] Between 50% and 70% of the structures of the human proteome are incomplete without covalently-attached glycans.[73][68] AlphaFill, a derived database, adds cofactors to AlphaFold models where appropriate.[74]
In the algorithm, the residues are moved freely, without any restraints. Therefore, during modeling the integrity of the chain is not maintained. As a result, AlphaFold may produce topologically wrong results, like structures with an arbitrary number of knots.[75]

首先 神经网络用于蛋白质预测 很久了，有几十年的历史。 这个你承认把?
再次，这里面说了， 神经网络有很多种，用于计算机视觉的神经网络也被人拿来预测蛋白质。这个应该就是指卷积网络，卷积网络应该就是这个xu的。
再然后，又说了，基于attention的 transformer 网络 也被用来预测蛋白质。这个就跟xu的卷积网络 不是一个路数。
再然后，他的alphafold属于自己发明的一种网络叫evoformer，这个就是跟之前的不同了。

然后他这个alphfafold网络 预测了不知道多少个蛋白质，据说有几十万个？
跟老徐的这个预测了几百个，那就不是一个数量级了。

对吧，我说的0.02还是高估老徐了，实际上预测几百个蛋白质也就是相当于0.002吧，或者0.0002？

[superdsb]

AlphaFold给的credit太宽了
这种是典型的1到100 不是0到1
而且Xu的模型没有接着做下去恐怕还是因为资源算力的问题
接着做下去优化推广说不定也能有很大影响力
所以这种根本不应该给奖

首先 神经网络用于蛋白质预测 很久了，有几十年的历史。 这个你承认把?
再次，这里面说了， 神经网络有很多种，用于计算机视觉的神经网络也被人拿来预测蛋白质。这个应该就是指卷积网络，卷积网络应该就是这个xu的。
再然后，又说了，基于attention的 transformer 网络 也被用来预测蛋白质。这个就跟xu的卷积网络 不是一个路数。
再然后，他的alphafold属于自己发明的一种网络叫evoformer，这个就是跟之前的不同了。

然后他这个alphfafold网络 预测了不知道多少个蛋白质，据说有几十万个？
跟老徐的这个预测了几百个，那就不是一个数量级了。

对吧，我说的0.02还是高估老徐了，实际上预测几百个蛋白质也就是相当于0.002吧，或者0.0002？

[英亲王阿齐格]

AlphaFold给的credit太宽了
这种是典型的1到100 不是0到1
而且Xu的模型没有接着做下去恐怕还是因为资源算力的问题
接着做下去优化推广说不定也能有很大影响力
所以这种根本不应该给奖

不是啊
alphfafold不是徐的思路啊，徐就是卷积网络属于图像识别的思路


alphfafold是比这个巧妙多了，他用attention的机制 来搜寻类似的序列 在已知蛋白质的构型，然后根据趋同进化的原理 判断 差不多的序列承担差不多的功能，构型应该是差不多的 这个思路 来预测。

[Caravel]

这是jumper那片nature自己吹牛逼的话
什么几把同行的都lack atomic accuracy
怎么定义的accuracy也没说而且也没比较数据
这种要是换了referee是绝对不会让发出来的
而且AlphaFold自己就没有多高的accuracy

AlphaFold has various limitations:

AlphaFold DB provides monomeric models of proteins, rather than their biologically relevant complexes.[68]
Many protein regions are predicted with low confidence score, including the intrinsically disordered protein regions.[69]
Alphafold-2 was validated for predicting structural effects of mutations with a limited success.[70]
The model relies to some degree upon co-evolutionary information across similar proteins, and thus may not perform well on synthetic proteins or proteins with very low homology to anything in the database.[71]
The ability of the model to produce multiple native conformations of proteins is limited.
AlphaFold 3 version can predict structures of protein complexes with a very limited set of selected cofactors and co- and post-translational modifications.[72] Between 50% and 70% of the structures of the human proteome are incomplete without covalently-attached glycans.[73][68] AlphaFill, a derived database, adds cofactors to AlphaFold models where appropriate.[74]
In the algorithm, the residues are moved freely, without any restraints. Therefore, during modeling the integrity of the chain is not maintained. As a result, AlphaFold may produce topologically wrong results, like structures with an arbitrary number of knots.[75]

deepmind估计当时是到处征集科学问题，看看能不能用deep learning解决。Jumper听了徐锦波的报告觉得这也是个好方向就去了。

这玩意其实没有太多idea，就是有一些domain knowledge引进门，后面就是暴力实验，deep mind这方面有非常强大的实力，估计有10-20人的团队。其实这里面的meat肯定是在deep learning算法方面。 Jumper到底有多大贡献我很怀疑。

当初alphago，也是脸书的tianyuandong，先出了成果，在一个网站上下到了5段，同时deep mind发力，上了20个人，搞出alphago。tianyuandong基本就是单枪匹马。

[Caravel]

不是啊
alphfafold不是徐的思路啊，徐就是卷积网络属于图像识别的思路


alphfafold是比这个巧妙多了，他用attention的机制 来搜寻类似的序列 在已知蛋白质的构型，然后根据趋同进化的原理 判断 差不多的序列承担差不多的功能，构型应该是差不多的 这个思路 来预测。

没啥巧妙，就是用最新的deep learning的架构，徐锦波估计只懂很初级的全联通和cnn.

[huangchong]

而且他只预测了几百个蛋白质把

跟alphafold那种预测了几乎所有的，还是不一样的，alphafold的论文你看了吗？ 你但凡看过 就不会觉得 ，跟徐有什么关系。 alphafold的网络用的是 attention，所以他叫evoformer（这个显然是取了transformer，外加evolution）

预测多少不说明什么问题。关键是alphafold2一出来 准确度让蛋白质晶体学界非常震惊。这就造成了轰动效应。

前面说了 拿dl去学pdb 这种主意我2012都有 并不是什么难点 难的是真的得到预测准确的网络

[英亲王阿齐格]

预测多少不说明什么问题。关键是alphafold2一出来 准确度让晶体学界非常震惊。

在论文里预测了几百个的意思是
只有这几百个对得上

如果能对得上几千个，他肯定不会说预测了几百个，肯定会说预测了几千个

[huangchong]

deepmind 的alphafold2 比赛结果出来 整个蛋白质结构界是非常震惊的 然后不知怎么 David Baker的韩国女博后也做了个类似的网络 叫rosettafold 跟 alphafold2背靠背发了nature 估计是他们通过什么办法 把alphafold的方法给猜出来了一些

[行观曰]

在论文里预测了几百个的意思是
只有这几百个对得上

如果能对得上几千个，他肯定不会说预测了几百个，肯定会说预测了几千个

那是能拿到的数据量和算力资源所决定的

请继续换姿势跪舔

－－－－－
诺贝尔一直标榜奖励的原创性，

徐是第一个人明确“告诉大家用AI 可以大幅度提高蛋白质结构预测精度”

后面跟上的人属于技术改进

[huangchong]

“告诉大家用AI 可以大幅度提高蛋白质结构预测精度”

而且 John Jumper 是听了徐的讲座的，还和他的博士后有交流

那看来以后是不能听讲座了

[行观曰]

那看来以后是不能听讲座了

是徐不应该给讲座：

原理已经走通了

viewtopic.php?t=611526

徐当时应该去找算力最大的公司，而不是在学术界瞎BB去给讲座