跟讀練習: CNV analysis with Daniel Howrigan - 通過影片學習英語口語

選擇練習模式

每種模式都會開啟全屏練習室。

196 句

如果句子過短或過長，請點選 Edit 進行調整。

Hello.

My name is Daniel Harrigan, and today I'll be talking about how to run copy number variation analysis.

Today's talk, I'll address a couple of questions, namely, what is a copy number variant and how do we detect it with genotype data?

What does the CNV file format look like?

And what does CNV analysis output look like?

Finally, how do I use this to run CNV burden and association tests?

So, what is a copy number variant?

Well, I define it here as a subset of structural variation involving a gain

or duplication or a loss slash deletion of genomic sequence.

Now, I call it a subset of structural variation because structural variation

can encompass basically any change in the length of a genomic sequence

that can be as small as a single base pair insertion

or deletion now most snips that we we think of are just substitutions

so the actual size of the genome isn't changed with a single nucleotide polymorphism

but a structural variant can encompass anything from very small all the way to entire chromosomes.

What we call CNVs, while it being kind of a general name,

falls into the category of roughly at least defined here at minimum one kilobase to submicroscopic.

With array data, we generally say something around 10 to 100 kilobases, given our sensitivity to detect these CNVs,

whereas larger CNVs usually are greater than 500 kilobases up to multiple megabases.

And then we get into much larger events.

Now, I'm not going to get into the details of the mechanisms that cause it.

I've listed a few here.

You can look them up.

But there are different hotspots in the genome

that are more prone to these copy number variants due to repeat regions

or often ways in the machinery that can make mistakes and lead to these gains or losses of genomic sequence.

How do we detect it using genotype data?

So when we run across the genome collecting a bunch of SNPs, I show here kind of the basic mechanism

or at least a figure showing how we take light intensity from different experiments looking at capturing either

you know allele a i have aa as allele a

or the b allele being tt here

and then a heterozygous you can see a mix of both red

and green now we can leverage these allele frequencies

and what we call the b allele frequency

or at least the frequency of of the t in this case now normally

when there's no copy number variant these will be roughly equal

and so will be around 50 percent

when we see something like a large deletion or any sort of deletion we should see a loss of heterozygosity,

and therefore we will see a gap in these B alleles,

at least at any site that you would normally be heterozygote at.

With the duplication, these would be a little bit off, so it would be more like two -thirds,

one -third, and you would see maybe a movement kind of akin to where these pink lines may be going,

even though they don't quite to find that you would see in a duplication

that the b allele frequencies move off 50 but not to one

and zero now that's looking at the b alleles the other thing

we use to detect it is the log r ratio

and this is basically measuring the light intensity and when you see a drop here

where we would see maybe in in all of these calls you should see a drop of say 50 percent

when we have a deletion and you will see subsequently a rise of about 33 percent

for a duplication now i'm not going to get into the details of the colors

that are used to detect these copy number variants suffice to say this is kind of the foundation that it's built upon

What I'm going to talk about next is given

that you've run copy number variant callers here are the kind of QC calls that you're moving into the analysis stage.

Now, I use Plink to analyze the CNV data so you can get a bunch of different file formats,

copy number variants from different callers.

But you would let, you know, if you want to use Plink, you would convert this into this particular file format, which is the .cnv file.

And it's basically individual identifiers, the chromosome, the start and end position of this CNV,

the type, one being a deletion, three being a duplication.

and then also there's a few other fields here the score

and sites field in this example i use the score being say the number of copy number variant callers

that agreed upon this given call

and it could range up to six sites here being the number of snips that are used to call the cnv

Now, I note here that score and sites are not forced into a particular convention.

You could say replace score with the number of genes overlapping the CNV,

or the sites being some other variable that you would be interested in measuring.

Now along with this file, plink creates a .cnv .map file,

and this is basically breaking down the break points of each

cnv into akin to a map file similar to what we have for snip data

and you can see here i note

that every different position is mapped even the the end position

but also maybe a single position after that end

because you may want to do a test an additional final

test after the end of a CNB to see how things have changed.

Note that the CNB file format commands are not available in Plink 1 .9, but the initial version granted the speed ups

that you get using the newer version of Plink aren't all

that applicable here because we're generally dealing with rare variation and

so these file sizes are generally not too large and the sorts of computing that you use is not too heavy.

So what does CNV analysis output look like?

Usually whenever you run a command in Flink looking at your CNV files, you get out .cnv .indiv file.

So this is per sample file where you say the number of cnv segments

that this individual has the number of kilobases that these segments cover

and the average kilobases covered per segment you also get a cnv .summary file

this is akin to the map file that is summarizing the number of affected

and unaffected individuals at any given break point or start, end, and plus one of a CMB.

And this is basically, you know, the output here, while it seems quite simple,

I'll show how with these files you can do quite sophisticated analysis by using a lot of different filters.

So a lot of the magic in Plink is all the flags that you can use to subset your list of CNVs.

So what I have here is a kind of verbose command just to show the optionality available in Plink.

And for each of these, I describe what that function is doing.

So with Plink, you have dash dash C file.

This reads in the CNV and CNV map file.

I want to select only deletions.

I want to select CNVs that are at least 100 kilobases in length.

I want CNVs with a score of four or higher and at least 50 sites.

I want to exclude CNVs that overlap a particular region.

So I can insert a different text file with a list of chromosome and start -end positions here.

And I want to make sure in this exclusion that the CNVs must overlap by at least 50 % to be excluded.

I can also look at frequency where at least 10 CNVs overlap.

I would like to exclude those because maybe I'm more interested in very rare.

100

uh cnvs i can also write out the frequencies of of these of these cnvs as well just

101

so i can guarantee which cnvs are being dropped

102

which cnvs are being kept

103

and then i can run a basic burden test using a permutation model

104

and here i just set the number of permutations these 10 000

105

so you can see there's a lot of different flags here and it's manipulating a lot of these flags

106

that can give you just what you would like in terms of your analysis now granted

107

using the burden tests in plank doesn't handle covariates it basically just looks at something like case control status

108

and so what i recommend is taking the output particularly the cnd cnv .indiv file

109

and reading that into say python or r i prefer r to run more sophisticated models

110

and so you can see when you put more filters here

111

if i go back the number of segments will change depending on what filters,

112

and obviously, subsequently, the number of kilobases covered by these segments will change as a function.

113

And it's kind of iteratively reading in these files at different filtering steps that can produce a wide range of tests.

114

So I've shown a couple figures here that we published when looking at CNV burden in the PGC schizophrenia.

115

So I took in those .CNV .indiv files into R.

116

I ran a logistic regression predicting schizophrenia status and adding a number of covariates,

117

such as principal components, genotype platforms, and then basically as I iteratively ran different

118

uh commands in plank to look at kv cnv counts lengths frequencies whether

119

or not they're in

120

or not in particular regions you can build up a number of tests of overall burn

121

so in this example here this would be all cnvs deletions

122

and duplications stratified by different genotyping platforms and then all together and then b here i'm stratifying by different frequencies

123

and say previously implicated cnvs as a region

124

i want to say in the blue bars we're looking at enrichment here in the green

125

and blue bars but you can see the big deviation here there's a lot of enrichment

126

when we talk about cnvs at this size or at least this frequency i mean

127

and they go away because most of these have been implicated

128

so i've excluded these regions rerun the burden test

129

and you can see that we've captured much of the signal already with previously implicated cnvs

130

so how do we use this to run cnv association tests at individual cnv loci or at individual break points of CNVs.

131

Basically, I would run a very similar command.

132

I could use all the same filters, basically get rid of a number of commands.

133

In particular, you're getting rid of this CNV indiv .perm step, but you still run a permutation test,

134

and what you'll get out is a .cnv .summary .m perm file

135

and at each base position you can run an association test here using permutation

136

this would be the point wise permutation value

137

but there's also a family wise permutation p -value

138

that corrects for all the tests here

139

so this association is run at all possible start

140

and in n plus one positions and one of the things you can do

141

if you like to include covariates in your data at least

142

what i've done in the past is maybe run your logistic

143

regression model with a lot of your covariates pull out the residuals

144

and then use this as a quantitative trait and you can run association mapping in Plink to get p -values that way.

145

So what does this look like?

146

I think having a figure kind of is instructive here.

147

So I plotted

148

through a browser i'll break this down this is our signal at the norexin1 gene

149

so in red we have our deletions uh light red deletions in our schizophrenia cases

150

dark red deletions in our schizophrenia controls i also have duplications in blue

151

which don't make up much of the signal here

152

but as you can see i've also plotted the log negative log 10 p -value.

153

And if you kind of look closely, you can see each little breakpoint, you can see a different test being run.

154

And you can see at, you know, a spot like this where there's many different breakpoints,

155

a lot of granularity, you can run many different tests and kind of get a shape of the association around this gene.

156

Now you can also collapse.

157

Another test that we've done is collapsing across say all the exons of this gene

158

and so this would be akin more to like a gene burden test where you collapse this region

159

and then you test for overlap at that region run a similar model

160

and then you can kind of aggregate all the cnvs and cases and controls to report say a gene -based p -value.

161

So that's a very quick overview of how to run burden and association with CNV data.

162

And a few considerations is that, you know, one of the things that you don't have access to

163

when looking at cnvs is imputation and so it is a consideration to think about you know

164

there isn't kind of a reference uh you know reference haplotypes

165

or a larger data set to do additional qc

166

so there can be additional challenges particularly with subpar data

167

that they can't be rescued in the way

168

that imputation uh can rescue uh snip genotypes and on

169

that note you know genotyping chip does matter very much

170

because you can't impute a bunch of different new sites the

171

variability in terms of the number of snips particularly for smaller cnvs is a very kind of important consideration

172

and so you know you can think of in a particular genotyping chip

173

if you don't have a good case control balance there you may not have the sensitivity to properly detect CNVs.

174

And so there's a lot of work

175

that goes into determining at what length of CNVs or at least at what case control balances given your genotyping chip,

176

do you have the right amount of power and sensitivity to do a proper case control association

177

Another thing too, ancestry PCs,

178

most of the CNV hotspots and associated CNV that we see in psychiatric disease aren't very impacted,

179

and mainly that's because these are recurrent de novo CNV areas where there's a higher mutation rate,

180

but this is not really ancestrally defined in terms of the fact

181

that there's not a large difference in the allele frequency of the CNV across different ancestries, but it is still useful to include,

182

particularly as you get to higher frequencies and you get more inherited CNVs. And finally,

183

multiple testing correction biopermutation is one of the more robust ways to account for the multiple tests

184

because the nature of cnv data is such

185

that given the type of genotyping chip

186

that you use the size of your data set no no

187

particular study is going to be very similar to similar to another study and

188

that kind of leveraging the correlation structure within your own data set given

189

you know your ability to detect CNVs of various sizes

190

and frequencies using permutation is usually the best way to go about properly testing for association.

191

So if you have any questions, feel free to email me.

192

I've also put down some sites.

193

You could search here for the PGC CNV paper that we did in 2017.

194

And also, I think the write -up in PLINK on how to do this is really good

195

and very descriptive of all its functionality.

196

Thank you.

下載應用

流利口語所需的一切

AI 發音評分逐句打分

音標練習掌握每個音

詞彙學習建立你的詞庫

詞彙遊戲邊玩邊學

掃碼

Download on theApp Store Get it onGoogle Play

FOR YOU

本課的詞彙與口語要點

這節口語課以影片“CNV analysis with Daniel Howrigan”為素材。影片中反覆出現的詞有：cnv, file, test, deletion, plink。這段影片共有 196 個句子、2593 個單詞可供跟讀。講話部分時長為 19:53。說話人語速平穩，每分鐘約 130 個詞，很適合跟讀。只有 77% 的單詞屬於英語最常用的 3,000 詞，詞彙難度較高。

影片中的重點詞彙

影片中 15 個值得學習的單詞，附發音和釋義：

單詞	發音	釋義
frequency 名詞	/ˈfɹiːkwənsi/	頻率 /頻率, 頻度 /頻度
variant 名詞	/ˈvæɹi.ənt/	變體 /變體, 異體 /異體
burden 名詞	/ˈbɝdn̩/	負荷 /負荷
detect 動詞	/dɪˈtɛkt/	察覺 /察覺, 發覺 /發覺
gene 名詞	/d͡ʒiːn/	基因
filter 名詞	/ˈfɪltɚ/	過濾器 /過濾器, 濾過器 /濾過器
segment 名詞	/ˈsɛɡmɛnt/	片斷 /片斷, 段
variation 名詞	/ˌvɛəɹiˈeɪʃn̩/	變化 /變化, 變動 /變動
overlap 動詞	/ˌəʊvəˈlæp/	重疊
format 名詞	/ˈfɔːɹ.mæt/	格式, 開本 /開本
output 名詞	/ˈaʊtpʊt/	輸出 /輸出
chip 名詞	/t͡ʃɪp/	籌碼
akin 形容詞	/ʌˈkɪn/	與（……）相似
schizophrenia 名詞	/ˌskɪtsəˈfɹiniə/	精神分裂症, 思覺失調症 /思覺失調症
sequence 名詞	/ˈsiː.kwəns/	序列, 順序 /順序

影片中出現的短語動詞

單詞	發音	釋義
break down 動詞		壞了 /壞了, 出故障
go away 動詞		離開 /離開, 走
go back 動詞		回去, 返回
make up 動詞	/ˌmeɪk ˈʌp/	組成 /組成
put down 動詞		放下

值得反覆跟讀的句子

影片中簡短而完整的句子，可以直接用在日常對話裡：

I've listed a few here.
How do we detect it using genotype data?
I've also put down some sites.

影片中的語法

說話人最常用的結構，並附上影片中的原話：

結構	影片中的用法
現在完成時 have/has + 過去分詞 — 過去發生但與現在仍有關聯的事	I've listed · you've run · have changed
被動語態 be + 過去分詞 — 強調發生了什麼，而不是誰做的	isn't changed · is given · are not forced

需要注意的發音

說話人用了 35 次縮略和弱讀形式，例如 I've, I'm, don't。請按聽到的簡短形式來說。

“sh” 和 “zh” 音: variation /ˌvɛəɹiˈeɪʃn̩/, structural /ˈstɹʌk(t͡)ʃəɹəl/, consideration /kənˌsɪd.əˈɹeɪ.ʃən/, capture /ˈkæp.(t)ʃɚ/, dash /ˈdæʃ/
長單詞——注意重音位置: schizophrenia /ˌskɪtsəˈfɹiniə/, consideration /kənˌsɪd.əˈɹeɪ.ʃən/, sensitivity /ˌsɛn.səˈtɪv.ə.ti/, mechanism /ˈmɛk.ə.nɪ.zəm/, subsequently /ˈsʌb.sɪ.kwənt.li/

中文母語者容易讀錯的音：

/v/ — 不要讀成 /w/，上齒輕觸下唇: variant /ˈvæɹi.ənt/, variation /ˌvɛəɹiˈeɪʃn̩/, overlap /ˌəʊvəˈlæp/, sensitivity /ˌsɛn.səˈtɪv.ə.ti/, leverage /ˈlɛv.(ə.)ɹɪd͡ʒ/
詞尾子音 — 要讀清楚，後面不要加母音: variant /ˈvæɹi.ənt/, detect /dɪˈtɛkt/, segment /ˈsɛɡmɛnt/, overlap /ˌəʊvəˈlæp/, format /ˈfɔːɹ.mæt/
/r/ 和 /l/ 的區別: structural /ˈstɹʌk(t͡)ʃəɹəl/, roughly /ˈɹʌf.li/, leverage /ˈlɛv.(ə.)ɹɪd͡ʒ/, variable [ˈvæɹ.i.ə.bl̩], correlation /ˌkɔɹ.əˈleɪ.ʃən/

如何用這段影片練習

先完整聽一遍影片，不要開口，記下不認識的單詞。
用正常速度逐句跟讀，每句重複到你的節奏與說話人一致為止。
錄下自己的聲音並與原聲對比，特別注意 frequency, variant, burden 這類單詞。

什麼是跟讀法？

跟讀法 (Shadowing) 是一種有科學依據的語言學習技巧，最初開發用於專業口譯員的培訓，並由多語言者Alexander Arguelles博士普及。這個方法簡單而強大：您在聽英語母語原聲的同時立即大聲重複——就像是一個延遲1-2秒緊跟說話者的影子。與被動聽力或語法練習不同，跟讀法強迫您的大腦和口腔肌肉同時處理並模模擬實的講話模式。研究表明它能顯著提高發音準確性，語調，節奏，連讀，聽力理解和口語流利度——使其成為雅思口語備考和真實英語交流最有效的方法之一。

影子跟讀法: 閱讀完整分步指南 →

跟讀練習: CNV analysis with Daniel Howrigan - 通過影片學習英語口語

選擇練習模式

下載應用

推薦影片

4 How to use PCR and qPCR

Other Corneal Disorders - CRASH! Medical Review Series

We can cure almost all human diseases. Here’s how. | Albert-László Barabási

SANA-Video: Efficient Video Generation with Block Linear Diffusion Transformer, [ICLR 2026, Oral]

Sequential, Multiple Assignment, Randomized Trial Designs With Dr Kidwell

RE2 Unit 9B

Gaussian Naive Bayes, Clearly Explained!!!

9 Insane Claude Mods I Already Can't live Without

Interview with Dr. Peter Schwartz | SADS Live Episode 23 | July 31, 2020

SVINutes: VAN Score

VETERINARY BASICS: How to read a blood test! CBC & Haematology

What it's like to have schizophrenia