gan學到的是什么

一個介紹 (An Introduction)

Generative Networks like GANs are unique to other deep learning models in that they generate a sample instead of optimizing for an output. This allows for a measure of creativity; scientists can analyze the output of a generative model to understand a biological system.

諸如GAN之類的生成網絡對于其他深度學習模型而言是獨特的，因為它們生成樣本而不是針對輸出進行優化。 這可以衡量創造力； 科學家可以分析生成模型的輸出以了解生物系統。

The purpose of this article is to detail the potential applications of GANs to scientific research, so I will assume a preliminary understanding of GANs. Basically, GANs contains a generator which learns the distribution of a dataset with the help of a discriminator, resulting in a model capable of outputting new samples. The architecture is as follows:

本文的目的是詳細介紹GAN在科學研究中的潛在應用，因此，我將對GAN進行初步的了解。 基本上，GANs包含一個生成器，該生成器借助鑒別器來學習數據集的分布，從而生成能夠輸出新樣本的模型。 架構如下：

source來源

I divide the contribution of GANs to scientific research into 3 major categories: preparation, direction, and modeling.

我將GAN對科學研究的貢獻分為三個主要類別：準備，指導和建模。

制備 (Preparation)

Preparing samples to study is probably one of the most unattractive parts of biological research. GANs are useful when we have some data, and we want to convert that into an image (when we input data, we use conditional GANs, or cGANS). Thus, we need to analyze where we might need to produce an image in the preparation phase.

準備樣本進行研究可能是生物學研究中最缺乏吸引力的部分之一。 當我們有一些數據并且想要將其轉換為圖像時(當我們輸入數據時，我們使用條件GAN或cGANS)，GAN很有用。 因此，我們需要分析在準備階段可能需要在何處產生圖像。

One paper [1] has used cGANs as a method for normalizing stained tissue cells for computational analysis. Stain normalization is done to reduce inconsistencies in stained tissues (e.g. some samples may be more darkly stained than others) and prepare those tissues for computer-aided detection systems.

一篇論文[1]已將cGANs用作歸一化染色組織細胞以進行計算分析的方法。 進行染色歸一化以減少染色組織中的不一致(例如，某些樣品可能比其他樣品更暗)，并為計算機輔助檢測系統準備這些組織。

However, performing standard techniques of stain normalization is often distorts subtleties in tissue structure. The authors utilized cGAN to address this problem, since the cGANs better learn the underlying structure of the tissue samples, thereby preserving its structure. cGANs can thus be used as a preprocessing step in a tissue analysis pipeline.

但是，執行標準的污漬歸一化技術通常會扭曲組織結構中的細微差別。 作者利用cGAN解決了這個問題，因為cGAN可以更好地了解組織樣本的基礎結構，從而保留其結構。 因此，cGAN可用作組織分析流程中的預處理步驟。

[source][來源]

方向 (Direction)

GANs is particularly useful for establishing potential directions in scientific study: we can generate molecules or try out potential protein structures using GANs.

GAN對于建立科學研究的潛在方向特別有用：我們可以使用GAN生成分子或嘗試潛在的蛋白質結構。

Molecules that GANs output are rarely stable or potentially useful, but we can subsequently use other deep learning models to screen the few promising molecules in a dataset. This will advance drug discovery by outputting far more viable drugs than we can produce with standard techniques (standard techniques for drug discovery means just relying on the imagination of senior chemists) [2].

GANs輸出的分子很少穩定或可能有用，但是我們可以隨后使用其他深度學習模型來篩選數據集中的一些有前途的分子。 這將通過輸出比我們用標準技術(標準的藥物發現技術僅依賴于高級化學家的想象力)生產的可行性更高的可行藥物來促進藥物開發[2]。

[source][來源]

The process of drug discovery is more than just discovering possible drugs, though. Testing is rigorous and the probability that a drug passes these standards is extremely low. For a drug to be useful, it must react with the intended protein or pathway to produce the intended effect while simultaneously not reacting to our countless other bodily systems. Such a feat is necessarily difficult; GANs just allows us to fail faster.

但是，藥物發現的過程不僅僅是發現可能的藥物。 測試非常嚴格，藥物通過這些標準的可能性極低。 為了使一種藥物有用，它必須與預期的蛋白質或途徑發生React以產生預期的效果，同時又不與我們無數的其他身體系統發生React。 這樣的壯舉一定是困難的。 GAN只是讓我們更快地失敗。

GANs can also suggest new scientific directions by generating potential protein designs. Outputting new proteins, however, is a much harder task than generating small molecules by virtue of the countless interactions contributing to the protein’s complex structure. Therefore, using GANs to design new proteins is not as developed as for generating smaller molecules. However, because new proteins are used for industry (e.g. laundry detergent uses enzymes), the task of protein design is not limited by the downstream effects of molecules in a biological system.

GAN還可以通過產生潛在的蛋白質設計來提出新的科學方向。 但是，由于產生了許多復雜的蛋白質，因此與產生小分子相比，輸出新蛋白質要困難得多。 因此，使用GAN設計新蛋白的能力還不如生成較小的分子。 但是，由于新的蛋白質被用于工業(例如洗衣粉使用酶)，因此蛋白質設計的任務不受生物系統中分子的下游作用的限制。

造型 (Modeling)

A GAN is able to output a new image because it learns the distribution of a particular kind of imaging. For instance, if the GAN trains on a dataset of cats, it learns the distribution of cat images, and hence is able to output images (which look like cats) based on that distribution. We can use the same process to model biological systems.

GAN能夠輸出新圖像，因為它可以學習特定種類的成像的分布。 例如，如果GAN在貓的數據集上訓練，它會學習貓的圖像分布，因此能夠基于該分布輸出圖像(看起來像貓)。 我們可以使用相同的過程對生物系統進行建模。

If one protein affects the structure of the entire cell, then we can modify the feature vector representing that protein to generate an image of the cell’s structure. Using this GANs, we can study changes of the cell structure by taking samples at discrete phases in its development. Then, we can interpolate the feature vectors of the input protein, resulting in a continuous model of the cell’s development [4].

如果一種蛋白質影響整個細胞的結構，那么我們可以修改代表該蛋白質的特征向量，以生成細胞結構的圖像。 使用這種GAN，我們可以通過在發育過程中的不同階段取樣來研究細胞結構的變化。 然后，我們可以內插輸入蛋白質的特征向量，從而形成細胞發育的連續模型[4]。

[source][來源]

We can apply the same process to a plethora of systems, such as modeling tissue development [5]. Subsequently, when we further study the system, we can derive additional variables affecting the system. We can retrain our GAN with the addition of that new discovered variable, then the GAN will be an even more accurate model of the biological system. Then, when we discover all significant variables, the GAN will practically be a perfect representation of the biological system: we input some variables, then it generates the exact conditions of the biological system.

我們可以將相同的過程應用于多種系統，例如對組織發育進行建模[5]。 隨后，當我們進一步研究系統時，我們可以得出影響系統的其他變量。 我們可以通過添加新發現的變量來重新訓練GAN，然后GAN將成為生物系統的更準確模型。 然后，當我們發現所有重要變量時，GAN實際上將是生物系統的完美代表：我們輸入一些變量，然后它會生成生物系統的確切條件。

We are far from achieving this ideal given the complexity of practically any biological system, GANs can potentially represent perfect mathematical models of biological systems as we investigate further into them.

考慮到幾乎任何生物系統的復雜性，我們都遠未達到這一理想，隨著我們對它們的進一步研究，GAN可以潛在地代表生物系統的完美數學模型。

翻譯自: https://medium.com/@adam.mehdi23/gans-for-driving-biological-research-d1c2d678036c

gan學到的是什么

總結

以上是生活随笔為你收集整理的gan学到的是什么_GAN推动生物学研究的全部內容，希望文章能夠幫你解決所遇到的問題。

如果覺得生活随笔網站內容還不錯，歡迎將生活随笔推薦給好友。