exploratory_plots_dc.Rmd

---
title: "R Notebook"
output:
  html_document: default
  html_notebook: default
---


```{r}
source("global.R")
source("shinyCache2.R")
#load("callmat.RData")
#callgr <- callmat
datadir <- "../data.chm"
ids <- c("16c58fbcc5633564b10ebe8f78d87883",
		"40c68f29b6d7cb2358f31a7073250406",
		"43a13d07730deb934e9fc01e3b3cd26f",
		"8dcad8fe04f4ebc0ad3254ab4420cdc8",
		"acd889cc16741fb0fba62faa4f7005f3",
		"b1112f1c3cbd28c464f58fc5c5c02f9b",
		"9d134f160ac68c0445002fbb78db4a5e")
callmat <- .LoadCallMatrixForIds(
	datadir=datadir,
	metadata=LoadCachedMetadata(datadir),
	ids=c("00000000000000000000000000000001", ids),
	ignore.interchromosomal=TRUE, mineventsize=51, maxeventsize=NULL,
	maxgap=200,
	sizemargin=0.25,
	ignore.strand=TRUE,
	grtransform=.primaryHumanOnly,
	grtransformName="test",
	nominalPosition=FALSE
)

datadir <- "../data.na12878"
ids <- c(
		#"2f4d15f5f6e428fb91e12dac971571b9", #bcftools
		#"5cdeeb9fd824f642f0ebee04627adf6e", gridss 1.2.1
		"6ae03359fcf0a39aa236a0aaea7ea915",
		"8aaf2886ffe782e666661d6b890b330a",
		"80dd0c2aa34964f330702ee3d0bfa53f",
		#"1139a1e6ef2d598098fe5c2ff609052a", manta/0.29.6
		"26511afa5055a939fdf99a2ac24938cc",
		"a6cbc4fc76337d871ef899d51fbae0d9",
		"b6aeda0d81ff8c839e381b57898dc3d8",
		#"e77293fb821458c3fd04c58ea88b7b16", #gasv
		#"ee44a2b21004af249f115bb5d9508ceb", gridss 1.3.0
		#"f75b305c5579449e347c1b87832e65d8", #gasv
		"f849b9493f2dd5dc20b6b7e49e1c89d7",
		"fa5f9a52d3b67a206cb6e581053a9269")

callgr <- .LoadCallMatrixForIds(
	datadir=datadir,
	metadata=LoadCachedMetadata(datadir),
	ids=c(
		"00000000000000000000000000000000",
		"00000000000000000000000000000001",
		"00000000000000000000000000000002",
		"00000000000000000000000000000003",
		ids),
	ignore.interchromosomal=TRUE, mineventsize=51, maxeventsize=NULL,
	maxgap=200,
	sizemargin=0.25,
	ignore.strand=TRUE,
	grtransform=.primaryHumanOnly,
	grtransformName="test",
	nominalPosition=FALSE
)
# callgr <- callgr[simpleEventType(callgr) == "DEL"]

source("sv_benchmark.R")

metadata <- 
    LoadCachedMetadata(datadir) %>% 
    as.tbl()

callgr$simpleEvent <- simpleEventType(callgr)
callgr <- callgr[!is.na(callgr$svLen) & abs(callgr$svLen) < 50000]
callgr$QUAL[is.na(callgr$QUAL)] <- 1
callgr$CallSet <- ifelse(callgr$ignore.filtered, PASS_CALLS, ALL_CALLS)

```

PCA plot of calls made

```{r}

calldf <- mcols(callgr)
calldf <- calldf[,(1+which("Id" == names(calldf))):length(names(calldf))]
callm <- as.matrix(calldf)

```

We now want to rescale ouur QUALs so they are comparable.
since we make no assumptions on the shape of the QUAL, the transform
can be arbitrary, we only require it to be monotonic

Lets try CDF of precision

Rolling window of precision
```{r}
require(tidyverse)
require(zoo)
require(ggplot2)

data.frame(Id=callgr$Id, CallSet=callgr$CallSet, qual=callgr$QUAL, tp=callgr$Id00000000000000000000000000000001 != -1) %>%
	group_by(Id, CallSet) %>%
	arrange(qual) %>%
	mutate(
		windowprec100=rollmean(tp, k=100, fill=FALSE, align="right"),
		windowprec250=rollmean(tp, k=250, fill=FALSE, align="right"),
		windowprec1000=rollmean(tp, k=1000, fill=FALSE, align="right")
		) %>%
	ggplot(aes(x=qual)) +
	geom_point(aes(y=windowprec100), colour="red") +
	geom_point(aes(y=windowprec250), colour="green") +
	geom_point(aes(y=windowprec1000), colour="blue") +
	coord_cartesian(ylim=c(0,1))
	facet_wrap(~ Id + CallSet, scales="free")
```

Leon suggested that a lowess fit was more appropriate as it weights nearby points according to their distance in QUAL space

Lachlan suggested smooth.spline to fit the QUAL-precision curve

```{r}
for (id in ids) {
	ssz <- smooth.spline(x=callgr[callgr$Id==id]$QUAL, y=as.numeric(callgr[callgr$Id==id]$Id00000000000000000000000000000001 != -1), spar=0.5, tol=0.5)
	plot(ssz)
}
```

Log transform

```{r}
for (id in ids) {
	df <- data.frame(
			logq=log10(callgr[callgr$Id==id]$QUAL + 2),
			tp=callgr[callgr$Id==id]$Id00000000000000000000000000000001 != -1) %>%
		group_by(logq) %>%
		summarise(
			prec=sum(tp)/n(),
			w=n()
			)
	ssz <- smooth.spline(x=df$logq, y=df$prec, w=df$w, tol=1e-3)
	plot(ssz)
}
```

Plot QUAL bins

```{r}
require(binom)
for (id in ids) {
	df <- data.frame(
			q=callgr[callgr$Id==id]$QUAL,
			logq=log10(callgr[callgr$Id==id]$QUAL + 1),
			tp=callgr[callgr$Id==id]$Id00000000000000000000000000000001 != -1) %>%
		mutate(
			qbin=cut(q, breaks=(0:100)/100*max(q), labels=(0:99)/100*max(q)),
			logqbin=cut(logq, breaks=(0:100)/100*max(logq), labels=(0:99)/100*max(logq))) %>%
		group_by(qbin) %>%
		summarise(
			qmean=mean(q),
			logqmean=mean(logq),
			n=n(),
			prec=sum(tp)/n(),
			prec_lower=binom.confint(sum(tp), n(), 0.95, methods="exact")$lower,
			prec_upper=binom.confint(sum(tp), n(), 0.95, methods="exact")$upper)
	myplot <- ggplot(df) +
		aes(x=qmean,
			ymin=prec_lower,
			y=prec,
			ymax=prec_upper) +
		geom_errorbar() +
		geom_point(colour="blue")
		#stat_smooth(method="gam", method.args=list(family=binomial()))
	print(myplot)
}
```

```{r}
require(binom)
for (id in ids) {
	df <- data.frame(
			q=callgr[callgr$Id==id]$QUAL,
			logq=log10(callgr[callgr$Id==id]$QUAL + 1),
			tp=callgr[callgr$Id==id]$Id00000000000000000000000000000001 != -1) %>%
		mutate(
			qbin=cut(q, breaks=(0:100)/100*max(q), labels=(0:99)/100*max(q)),
			logqbin=cut(logq, breaks=(0:100)/100*max(logq), labels=(0:99)/100*max(logq))) %>%
		group_by(logqbin) %>%
		summarise(
			qmean=mean(q),
			logqmean=mean(logq),
			n=n(),
			prec=sum(tp)/n(),
			prec_lower=binom.confint(sum(tp), n(), 0.95, methods="exact")$lower,
			prec_upper=binom.confint(sum(tp), n(), 0.95, methods="exact")$upper)
	myplot <- ggplot(df) +
		aes(x=logqmean,
			ymin=prec_lower,
			y=prec,
			ymax=prec_upper) +
		geom_errorbar(aes(alpha=pmin(1, n/20))) +
		geom_point(colour="blue") #aes(size=log10(n)), 
		#stat_smooth(method="gam", method.args=list(family=binomial()))
	print(myplot)
}
```

```{r}
require(binom)
df <- data.frame(
		Id=callgr$Id,
		q=callgr$QUAL,
		logq=log10(callgr$QUAL + 1),
		tp=callgr$Id00000000000000000000000000000001 != -1) %>%
	group_by(Id) %>%
	mutate(
		qbin=as.numeric(as.character(cut(q, breaks=(0:100)/100*max(q), labels=(0:99)/100*max(q)))),
		logqbin=as.numeric(as.character(cut(logq, breaks=(0:100)/100*max(logq), labels=(0:99)/100*max(logq))))) %>%
	group_by(Id, logqbin) %>%
	summarise(
		qmean=mean(q),
		logqmean=mean(logq),
		n=n(),
		prec=sum(tp)/n(),
		prec_lower=binom.confint(sum(tp), n(), 0.95, methods="exact")$lower,
		prec_upper=binom.confint(sum(tp), n(), 0.95, methods="exact")$upper)

ggplot(df) +
	aes(x=logqmean,
		ymin=prec_lower,
		y=prec,
		ymax=prec_upper) +
	coord_cartesian(ylim=c(0,1)) + 
	geom_errorbar(aes(alpha=pmin(1, n/20))) +
	geom_point(colour="blue") + #aes(size=log10(n)), 
	facet_grid(
	    ~ Id # + CallSet
	    , scales="free")
	#stat_smooth(method="gam", method.args=list(family=binomial()))
```

Trying a stacked plot w/ a histogram:

```{r}
require(cowplot)

test_id <- "26511afa5055a939fdf99a2ac24938cc"

test_df <- 
    df %>%
    ungroup() %>%
    filter(Id == test_id)

ci_plot <-
    test_df %>%
    ggplot() +
	aes(x = logqmean,
		ymin = prec_lower,
		y = prec,
		ymax = prec_upper) +
	coord_cartesian(ylim = c(0,1)) + 
	geom_linerange(aes(alpha = n > 10), color = "grey70") +
    scale_alpha_manual(values = c(0, 1)) +
	geom_point(aes(color = n > 10)) +
    scale_color_manual(values = c("grey70", "#396AB1")) +
    cowplot::theme_cowplot() +
    cowplot::background_grid(major = "xy", minor = "none") +
    theme(
        legend.position = "none",
        axis.text.x = element_blank(),
        axis.title.x = element_blank(),
        axis.ticks.x = element_blank(),
        plot.margin = unit(c(1, 1, -0.2, 1), "lines")) +
    ggtitle(test_id)

hist_plot <-
    test_df %>%
    ggplot() +
    aes(x = logqmean, y = log10(n)) +
    scale_y_reverse(
        expand = c(0,0),
        breaks = c(1,2,3,4)
    ) +
    geom_col(
        # Width manually in here?
    ) +
    # cowplot::theme_nothing() +
    cowplot::background_grid(major = "xy", minor = "none") +
    theme(
        axis.line = element_blank(),
        axis.ticks = element_blank(),
        plot.margin = unit(c(0, 1, 1, 1), "lines"))
```

```{r}
ci_plot
```

```{r}
hist_plot
```

```{r}
hist_grob <-
    ggplotGrob(hist_plot)

ci_grob <-
    ggplotGrob(ci_plot)
```

```{r}
require(gtable)
require(gridExtra)
require(grid)

g <- gridExtra::rbind.gtable(
    ci_grob, hist_grob, 
    size = "first")

grid::grid.newpage()
grid::grid.draw(g)
```

Rearrange heights above via the approach at ```http://felixfan.github.io/stacking-plots-same-x/#different-heights-same-widths```

```{r}
layout_indices <- grep("panel", g$layout$name)

panels <- g$layout$t[layout_indices]

g$heights[panels] <- g$heights[panels] * c(3,1)

grid.newpage()
grid.draw(g)
```

Package this up into a function:

```{r}
get_binned_qual_data <- function(callgr, bin_by) {
    
    df <- data.frame(
		Id=callgr$Id,
		q=callgr$QUAL,
		logq=log10(callgr$QUAL + 1),
		tp=callgr$Id00000000000000000000000000000001 != -1) %>%
	group_by(Id) %>%
	mutate(
		qbin=as.numeric(as.character(cut(q, breaks=(0:100)/100*max(q), labels=(0:99)/100*max(q)))),
		logqbin=as.numeric(as.character(cut(logq, breaks=(0:100)/100*max(logq), labels=(0:99)/100*max(logq))))) %>%
        group_by(Id, .dots = bin_by) %>%
	summarise(
		qmean=mean(q),
		logqmean=mean(logq),
		n=n(),
		prec=sum(tp)/n(),
		prec_lower=binom.confint(sum(tp), n(), 0.95, methods="exact")$lower,
		prec_upper=binom.confint(sum(tp), n(), 0.95, methods="exact")$upper)

}


ci_plot <- function(test_id, test_df, qual_column) {
    
    caller_name <- (metadata %>% filter(Id == test_id))$CX_CALLER
    
    ci_ggplot <-
        test_df %>%
        ggplot() +
    	aes_string(x = qual_column) +
        aes(
    		ymin = prec_lower,
    		y = prec,
    		ymax = prec_upper) +
    	coord_cartesian(ylim = c(0,1)) + 
    	geom_linerange(
    	    aes(alpha = factor(n <= 10, levels = c(F, T))), 
    	    color = "grey70") +
        scale_alpha_manual(values = c(1, 0)) +
    	geom_point(aes(color = factor(n <= 10, levels = c(F, T)))) +
        scale_color_manual(values = c("#396AB1", "grey70")) +
        cowplot::theme_cowplot() +
        cowplot::background_grid(major = "xy", minor = "none") +
        theme(
            legend.position = "none",
            axis.text.x = element_blank(),
            axis.title.x = element_blank(),
            axis.ticks.x = element_blank(),
            plot.margin = unit(c(1, 1, -0.2, 1), "lines")) +
        ggtitle(StripCallerVersion(caller_name)) +
        ylab("precision")
    
    return(ci_ggplot)
}

flipped_hist_plot <- function(test_df, qual_column) {

    hist_ggplot <- 
        test_df %>%
        ggplot() +
        aes_string(
            x = qual_column,
            xend = qual_column
        ) +
        aes(y = log10(n)) +
        scale_y_reverse(
            expand = c(0,0),
            breaks = 1:6,
            labels = c("10", "", "1k", "", "100k", ""),
            limits = c(7,0)) +
        geom_segment(
            yend = 0, 
            color = "grey50") +
        # cowplot::theme_nothing() +
        cowplot::background_grid(major = "xy", minor = "none") +
        theme(
            axis.line = element_blank(),
            axis.ticks = element_blank(),
            plot.margin = unit(c(0, 1, 1, 1), "lines")) +
        ylab("# calls") +
        xlab(str_replace(qual_column, "qmean", " quality score"))
    
    return(hist_ggplot)
}

stacked_precision_plot <- function(
    test_id, callgr, qual_column = "logqmean") {
    
    bin_by <- str_replace(qual_column, "mean", "bin")
    
    test_df <- 
        get_binned_qual_data(callgr, bin_by) %>%
        filter(Id == test_id)
    
    hist_grob <-
        flipped_hist_plot(test_df, qual_column) %>%
        ggplotGrob()

    ci_grob <-
        ci_plot(test_id, test_df, qual_column) %>%
        ggplotGrob()
    
    combined_grob <- 
        gridExtra::rbind.gtable(
        ci_grob, hist_grob, 
        size = "first")
    
    layout_indices <- grep("panel", combined_grob$layout$name)

    panels <- combined_grob$layout$t[layout_indices]

    combined_grob$heights[panels] <- 
        combined_grob$heights[panels] * c(2,1)
    
    return(combined_grob)
}

grid.newpage()
grid.draw(stacked_precision_plot(test_id, callgr, "qmean"))
```

All the ids:

```{r}
all_id_precision_plot_grobs <-
    map(
        ids, 
        (function(caller_id) {
            caller_name <- (metadata %>% filter(Id == caller_id))$CX_CALLER
            stacked_precision_plot(
                caller_id, callgr,
                ifelse(
                    StripCallerVersion(caller_name) %in% 
                        c("manta", "hydra"),
                    "qmean",
                    "logqmean"
                ))}))
```

```{r}
grid.newpage()

combined_grobs <-
    do.call(
        function(...) grid.arrange(..., ncol = 3), 
        all_id_precision_plot_grobs)

grid.draw(combined_grobs)
```


Try raw smoothing
```{r}
for (id in ids) {
	df <- data.frame(
			q=callgr[callgr$Id==id]$QUAL,
			logq=log10(callgr[callgr$Id==id]$QUAL + 1),
			tp=as.numeric(callgr[callgr$Id==id]$Id00000000000000000000000000000001 != -1))
	myplot <- ggplot(df) +
		aes(x=q, y=tp) +
		geom_smooth() +
		coord_cartesian(ylim=c(0,1))
	print(myplot)
}
```



Precision/recall plot

```{r}

callgr$caller_hits_ex_truth <- rowSums(as.matrix(as.data.frame(mcols(callgr)[,
	str_detect(names(mcols(callgr)), "^fId[a-f0-9]+") & !(names(mcols(callgr)) %in% c(
			"fId00000000000000000000000000000000",
			"fId00000000000000000000000000000001",
			"fId00000000000000000000000000000002",
			"fId00000000000000000000000000000003"))])) != -1)

rocby <- function(callgr, ..., rocSlicePoints=100, truthid="00000000000000000000000000000001", ignore.duplicates=TRUE) {
	groupingCols <- quos(...)
	callgr$truth <- mcols(callgr)[[paste0("Id",truthid)]]
	# TODO: filter duplicate calls
	if (ignore.duplicates) {
		# ignore calls that have a -2 in the truth column (don't touch the truth set though!)
		callgr <- callgr[callgr$Id != truthid & callgr$truth != -2]
	}
	eventCount <- sum(callgr$Id == truthid)
	as.data.frame(callgr) %>%
		dplyr::select(Id, CallSet, !!!groupingCols, QUAL, truth) %>%
		group_by(Id, CallSet, !!!groupingCols, QUAL) %>%
		summarise(
			fp=sum(truth < 0),
			tp=sum(truth >= 0)) %>%
		group_by(Id, CallSet, !!!groupingCols) %>%
		arrange(desc(QUAL)) %>%
		mutate(
			fp=cumsum(fp),
			tp=cumsum(tp)) %>%
		# each QUAL score is a point on the ROC plott
		group_by(Id, CallSet, !!!groupingCols, QUAL) %>%
		summarise(tp=max(tp), fp=max(fp)) %>%
		# QUAL scores with the same number of tp calls can be merged on the ROC plot
		group_by(Id, CallSet, !!!groupingCols, tp) %>%
		summarise(fp=max(fp), QUAL=min(QUAL)) %>%
		# subsample along tp and tp+fp axis
		group_by(Id, CallSet, !!!groupingCols) %>%
		dplyr::slice(unique(c(
			1,
			findInterval(seq(0, max(tp), max(tp)/rocSlicePoints), tp),
			findInterval(seq(0, max(tp + fp), max(tp + fp)/rocSlicePoints), tp + fp),
			n()
		))) %>%
		ungroup() %>%
		mutate(
			fn=eventCount-tp,
			precision=tp / (tp + fp),
			fdr=1-precision,
			sens=tp/eventCount)
}

```

By Shared hits & event type

```{r}

rocbySharedHits <- rocby(callgr, simpleEvent, caller_hits_ex_truth)

ggplot(rocbySharedHits %>% filter(Id != "00000000000000000000000000000001")) +
	aes(y=precision, x=tp, colour=Id, linetype=CallSet) +
	geom_line() +
	facet_grid(simpleEvent ~ caller_hits_ex_truth)
	labs(title="Precision-Recall by number of callers sharing call")

```

By event type

```{r}

rocbyeventType <- rocby(callgr, simpleEvent)

ggplot(rocbySharedHits %>% filter(Id != "00000000000000000000000000000001")) +
	aes(y=precision, x=tp, colour=Id, linetype=CallSet) +
	geom_line() +
	facet_wrap(~ simpleEvent)
	labs(title="Precision-Recall by event type")

```

By shared hits

```{r}
ggplot(rocby(callgr, caller_hits_ex_truth, truthid="00000000000000000000000000000003") %>% filter(!(Id %in% c("00000000000000000000000000000000", "00000000000000000000000000000001", "00000000000000000000000000000002", "00000000000000000000000000000003")))) +
	aes(y=precision, x=tp, colour=Id, linetype=CallSet) +
	geom_line() +
	facet_wrap( ~ caller_hits_ex_truth, scale="free") + 
	scale_colour_brewer(palette="Set1") + 
	labs(title="Precision-Recall by number of callers sharing call") +
	coord_cartesian(ylim=c(0,1))

```

Overall ROCs by truth set

```{r}
rocfull <- rocby(callgr, truthid="00000000000000000000000000000000") %>% mutate(truth="00000000000000000000000000000000") %>% rbind(
	rocby(callgr, truthid="00000000000000000000000000000001") %>% mutate(truth="00000000000000000000000000000001")) %>% rbind(
	rocby(callgr, truthid="00000000000000000000000000000002") %>% mutate(truth="00000000000000000000000000000002")) %>% rbind(
	rocby(callgr, truthid="00000000000000000000000000000003") %>% mutate(truth="00000000000000000000000000000003"))

ggplot(rocfull %>% filter(!(Id %in% c("00000000000000000000000000000000", "00000000000000000000000000000001", "00000000000000000000000000000002", "00000000000000000000000000000003")))) +
	aes(y=precision, x=tp, colour=Id, linetype=CallSet) +
	geom_line() +
	facet_wrap( ~ truth, scale="free") + 
	scale_colour_brewer(palette="Set1")
	labs(title="Precision-Recall by number of callers sharing call")

```

By SNP/indel context

```{r}
callgr$snp50bpbin <- cut(callgr$snp50bp, breaks=c(0, 1, 2, 3, 4, 5, 1000), right=FALSE)

```

Missed by GRIDSS
```{r}


callgr[
	callgr$Idb1112f1c3cbd28c464f58fc5c5c02f9b == -1 &
	callgr$caller_hits_ex_truth > 0 &
	callgr$Id =="00000000000000000000000000000001"]

```

Unique by annotation

```{r}
truthvcf <- readVcf("../data.chm/00000000000000000000000000000001.vcf", "")
repeattruth <- info(truthvcf)$REPEAT_TYPE
names(repeattruth) <- names(rowRanges(truthvcf))
truthgr <- breakpointRanges(truthvcf 
                            # , info_columns="REPEAT_TYPE" # chm-specific
                            )
truthgr$repeatAnn <- ifelse(is.na(truthgr$REPEAT_TYPE), "NotMasked", truthgr$REPEAT_TYPE)
callgr$repeatAnn <- ""
callgr$repeatAnn[callgr$Id =="00000000000000000000000000000001"] <- repeattruth[callgr$vcfId[callgr$Id =="00000000000000000000000000000001"]]
callgr$repeatAnn[callgr$Id =="00000000000000000000000000000001" & is.na(callgr$repeatAnn)] <- truthgr$repeatAnn[findOverlaps(
	callgr[callgr$Id =="00000000000000000000000000000001" & is.na(callgr$repeatAnn)],
	truthgr,
	type="equal", select="first")]
```

Common calls by repeat annotation

```{r}

ggplot(as.data.frame(callgr[callgr$Id =="00000000000000000000000000000001"])) +
	aes(fill=factor(caller_hits_ex_truth), x=repeatAnn) +
	geom_bar() + 
	theme(axis.text.x=element_text(angle=90,hjust=1)) +
	facet_wrap(simpleEvent ~ cut(abs(svLen), breaks=c(0, 50, 100, 1000, 100000), right=FALSE), scales = "free")

```

# pull in TRF annotation
```{r}

callgr$trf <- overlapsAny(callgr, grtrf[[1]], type="any")
ggplot(rocby(callgr, trf, truth_id=truth_id) %>% filter(Id != truth_id) %>% left_join(metadata)) +
		aes(y=precision, x=tp / 2, colour=StripCallerVersion(CX_CALLER), linetype=CallSet) +
		geom_line() +
		caller_colour_scheme +
		facet_wrap( ~ trf, scales="free") +
		scale_y_continuous(limits=c(0,1)) +
		labs(title=paste("Precision-Recall by presence of tandem repeat at breakpoint\n", sample_name, truth_name), colour="Caller")
```

## MDS plot

```{r}
call_df <-
    mcols(callgr) %>% 
    as.data.frame() %>% 
    as.tbl()

call_df_idcols <-
    call_df %>%
    select(starts_with("Id"), starts_with("fId"))
```

### First, with truth as query

```{r}
call_df_truth_query <-
    call_df_idcols %>%
    filter(Id == "00000000000000000000000000000001")

truth_query_hit_matrix <-
    call_df_truth_query %>%
    select(-Id) %>%
    as.matrix %>%
    (function(x) {x != -1}) %>%
    t()
```

#### Using the LogisticPCA package

```{r}
require(logisticPCA)
```

```{r}
logsvd_model <- 
    logisticSVD(
        truth_query_hit_matrix, 
        k = 2)
```

```{r}
raw_caller_ids <- 
    rownames(truth_query_hit_matrix) 

caller_ids <-
    raw_caller_ids %>%
    str_replace(".?Id", "")

caller_name_lookup <- 
    ifelse(is.na(metadata$CX_CALLER),
           str_trunc(metadata$Id, width = 5, side = "left", ellipsis = ""),
           metadata$CX_CALLER)

names(caller_name_lookup) <- metadata$Id

caller_names <-
    caller_name_lookup[caller_ids]

caller_names[is.na(caller_names)] <-
    names(caller_names)[is.na(caller_names)]
```


```{r}
require(cowplot)
require(ggrepel)

exp_fam_pca_plot <-
    plot(logsvd_model, type = "scores") + 
    geom_point(aes(color = str_detect(raw_caller_ids, "fId"))) +
    ggtitle("Exponential Family PCA of call") + 
    ggrepel::geom_label_repel(aes(label = caller_names)) +
    cowplot::background_grid(minor = "none") +
    theme_cowplot()

exp_fam_pca_plot
```


* * * * * *

```{r}
caller_similarity_matrix <-
    call_df_idcols %>%
    select(-Id) %>%
    as.matrix() %>%
    (function(x) {x != -1})
```

## Plot 3

```{r}
# Depends on caller_hits_ex_truth computed above
truth_hits_df <-
    callgr[callgr$Id == "00000000000000000000000000000001"] %>%
    as.data.frame() %>% as.tbl() %>%
    # Remove duplicates (filtered vs. unfiltered meaningless for truth)
    # filter(CallSet == ALL_CALLS) %>%
    # Remove non-filtered subject columns
    select(vcfId,
           matches("fId.+"), # "f" is for "full"
           -ends_with("0000000"), # Ignore other truth sets
           -ends_with("0000002"),
           -ends_with("0000003"),
           -ends_with("0000004")
           ) 
```

```{r}
# truth_hits_df$caller_hits_ex_truth %>% table()
```

```{r}
truth_hits_df_long <-
    truth_hits_df %>%
    gather(
        key = "CallSet_Id_subject", 
        value = "qual", 
        starts_with("fId")) # "f" is for "full"
```

Double-checking `caller_hits_ex_truth`:

```{r}
truth_hit_count_df <-
    truth_hits_df_long %>%
    group_by(vcfId) %>%
    summarise(caller_hits_ex_truth_2 = sum(qual != -1) / 2)
```

```{r}
# test_df <- 
#    truth_hits_df %>% 
#    left_join(truth_hit_count_df) %>% 
#    select(caller_hits_ex_truth, caller_hits_ex_truth_2)

# all(test_df$caller_hits_ex_truth == test_df$caller_hits_ex_truth_2)
```

```{r}
require(cowplot)
require(colorspace)

shared_calls_plot_df <-
    truth_hits_df_long %>% 
    # Note: -2 is placeholder for duplicate hit
    filter(qual != -1) %>%
    left_join(truth_hit_count_df) %>%
    mutate(
        caller_name = caller_name_lookup[
                str_replace(CallSet_Id_subject, "fId", "")] # "f" is for "full"
                %>% StripCallerVersion(),
        is_truth = caller_name == "00001")

n_callers_plus_truth <-
    length(table(shared_calls_plot_df$caller_name))

n_callers_palette <-
    c("black", 
      sequential_hcl(n_callers_plus_truth - 2, h = 235, c. = c(30, 20), l = c(20, 100)),
      "white") %>%
    rev()

shared_calls_plot_base <-
    shared_calls_plot_df %>%
    ggplot(aes(
        x = caller_name,
        fill = factor(caller_hits_ex_truth_2 / 2 - 1))) +
    facet_grid(factor(1 - is_truth) ~ ., scales = "free", space = "free") +
    scale_y_continuous(expand = c(0,0)) +
    geom_bar(color = "black") +
    scale_fill_manual(values = n_callers_palette, name = "# callers\nsharing") +
    xlab("") +
    coord_flip() +
    ggtitle("Sharing and distribution of true positive calls") +
    theme_cowplot()

shared_calls_plot <-
    shared_calls_plot_base +
    # Blanks out "is_truth" panels
    theme(strip.text = element_blank(), strip.background = element_blank())

# Add truth and subtract 1
# Add correct caller names/pretty labels
# Add plot titles
# Try "caller-as-query" version -- facet -- Yes!

shared_calls_plot
```

Faceted version:

```{r}
shared_calls_plot_faceted <-
    shared_calls_plot +
    facet_grid(
        factor(1 - is_truth) ~ factor(-1 * (caller_hits_ex_truth_2 / 2 - 1)),
        scales = "free", space = "free") +
    theme(
        panel.spacing.x = unit(0, "cm"),
        axis.line.x = element_blank(),
        axis.ticks.x = element_blank(),
        axis.text.x = element_blank(),
        axis.line.y = element_blank(),
        axis.ticks.y = element_blank()) +
    geom_bar(size = .65, color = "black")

shared_calls_plot_faceted
```

Using explicit boxes (DOESN'T (YET) WORK):

```{r}
box_start_lookup <-
    shared_calls_plot_df %>%
    filter(caller_name == "00001") %>%
    group_by(caller_hits_ex_truth_2) %>%
    summarise(n=n()) %>%
    arrange(desc(caller_hits_ex_truth_2)) %>%
    mutate(offset=cumsum(n)-n, tpcount=n) %>%
    select(caller_hits_ex_truth_2, offset, tpcount)

shared_calls_box_df <- shared_calls_plot_df %>% 
    group_by(CallSet_Id_subject, caller_name, caller_hits_ex_truth_2) %>% 
    summarise(n = n()) %>%
    left_join(box_start_lookup)
shared_calls_box_df <- shared_calls_box_df %>%
    mutate(tp=1) %>%
    rbind(shared_calls_box_df %>%
        mutate(tp=0, n=n-tpcount))

shared_calls_box_plot <- shared_calls_box_df %>%
    ggplot() +
    aes(x = caller_name, 
        y=n,
        alpha=as.numeric(!tp),
        #ymin = offset, 
        #ymax = offset + n,
        fill = as.factor(caller_hits_ex_truth_2),
        group=interaction(as.numeric(!tp), as.factor(caller_hits_ex_truth_2))) +
    geom_col()

shared_calls_box_plot
```

```{r}
precision_recall_by_shared_calls_plot <-
    rocby(
        callgr, caller_hits_ex_truth, 
        truthid="00000000000000000000000000000003") %>% 
    filter(!(Id %in% c(
        "00000000000000000000000000000000", 
        "00000000000000000000000000000001", 
        "00000000000000000000000000000002", 
        "00000000000000000000000000000003"))) %>%
    ggplot() +
	aes(y = precision, 
	    x = tp, 
	    colour = StripCallerVersion(caller_name_lookup[Id]), 
	    linetype = CallSet) +
	geom_line() +
	facet_wrap(
	    ~ factor(caller_hits_ex_truth, levels = max(caller_hits_ex_truth):1),
	    scale = "free",
	    ncol = 2) + 
	scale_colour_brewer(palette = "Set1") + 
	labs(
	    title = "Precision-recall by number of callers sharing call",
	    color = "caller",
	    linetype = "call set",
	    x = "# true positives",
	    y = "precision") +
	coord_cartesian(ylim=c(0,1)) +
    theme_cowplot()

precision_recall_by_shared_calls_plot
```

### Sharing of *false* positives

```{r}
# Depends on caller_hits_ex_truth computed above
false_positive_plot_df <-
    callgr[callgr$Id != "00000000000000000000000000000001"] %>%
    as.data.frame() %>% as.tbl() %>%
    # Exclude High confidence only hitting a truth call, inc. duplicate true positives
    filter(
        fId00000000000000000000000000000001 < 0) %>%
    # Remove duplicates (filtered vs. unfiltered meaningless for truth)
    # filter(CallSet == ALL_CALLS) %>%
    # Remove non-filtered subject columns
    select(Id, CallSet, caller_hits_ex_truth, matches("fId")) %>%
    gather(key=subject_Id_CallSet, value=QUAL, matches("fId")) %>%
    filter(paste0("fId",Id)==subject_Id_CallSet & QUAL != -2) %>%
    left_join(metadata) %>%
    mutate(caller_name = StripCallerVersion(CX_CALLER),
           is_truth = is.na(CX_CALLER))
```

```{r}
shared_false_positive_plot_base <-
    false_positive_plot_df %>%
    ggplot(aes(
        x = caller_name,
        fill = factor(caller_hits_ex_truth))) +
    facet_grid(factor(1 - is_truth) ~ ., scales = "free", space = "free") +
    scale_y_continuous(expand = c(0,0)) +
    geom_bar(color = "black") +
    scale_fill_manual(values = n_callers_palette, name = "# callers\nsharing") +
    xlab("") +
    # Chop off top of plot
    coord_flip(
        # ylim = c(0, 40000)
        ) +
    ggtitle("Sharing and distribution of false positive calls") +
    theme_cowplot()
    
shared_false_positive_plot_base
```

### Duplicate call counts

```{r}
dup_plot_df <- callgr %>%
    as.data.frame() %>% as.tbl() %>%
        # dirty hack to get truth not gathered
    select(Id, CallSet, Id00000000000000000000000000000001, matches("fId")) %>%
    gather(key=subject_Id_CallSet, value=QUAL, matches("fId")) %>%
    filter(paste0("fId",Id)==subject_Id_CallSet) %>%
    mutate(
        isTp=Id00000000000000000000000000000001 != -1,
        isFp=!isTp,
        isDup=QUAL == -2) %>%
    group_by(Id, CallSet) %>%
    summarise(prop_isDup = sum(isDup) / n()) %>%
    ungroup() %>%
    left_join(metadata) %>%
    mutate(caller_name = StripCallerVersion(CX_CALLER))
```

```{r}
dup_plot <- 
    ggplot(dup_plot_df) +
    aes(x = caller_name, y = prop_isDup, color = CallSet) +
    geom_errorbar(ymin = 0, ymax = .5, color = "grey50", width = .1) +
    geom_errorbar(ymin = .5, ymax = 1, color = "grey50", width = .1) +
    geom_point(alpha = 0.7, size = 5) +
    scale_color_brewer(palette = "Dark2", name = "") +
    # Actually: percentage of calls overlapping one of higher quality
    labs(x = "", y = "duplicate calls") +
    scale_y_continuous(labels = scales::percent, limits = c(0, 1), expand = c(.05,.05)) +
    theme(axis.line = element_blank(), axis.ticks.y = element_blank(), 
          axis.text.y = element_text(margin = margin(t = 0, r = -10, b = 0, l = 0))) +
    coord_flip()

dup_plot
```



#### Arrange these into a grid

```{r}
precision_recall_by_shared_calls_grob <-
    ggplotGrob(precision_recall_by_shared_calls_plot)

shared_calls_grob <-
    ggplotGrob(shared_calls_plot)

layout_matrix <-
    cbind(c(1,3), c(2,2))

fig_3_grob <-
    grid.arrange(
    grobs = list(
        shared_calls_plot, 
        precision_recall_by_shared_calls_grob,
        grid.rect(gp=gpar(fill=NA))),
    layout_matrix = layout_matrix)

grid.newpage()
grid.draw(fig_3_grob)
```


### TO DO

-   Common false positive calls by repeat masker annotation,
    event size, ENCODE blacklist
-   Stacked bar chart -- all vs. truth, but *also*
    each-as-query
-   Why do the 7-of-7 precision-recall curves reach different
    x-positions? Is it due to de-duplication (-2s), 
    duplication, ... ?