/*
* Parameters:
* SNPs [nIndividuals by nSNPs]:
* Matrix stored in column-major order.
* This will hold the result.
* NaNs will be set to 0.0 in the result.
*/
void SUFFIX(ImputeAndZeroMeanSNPs)(
REAL *SNPs,
const size_t nIndividuals,
const size_t nSNPs,
const bool betaNotUnitVariance,
const REAL betaA,
const REAL betaB
)
{
bool seenSNC = false; //Keep track of this so that only one warning message is reported
#ifdef ORDERF
for ( size_t iSnp = 0; iSnp < nSNPs; ++iSnp )
{
REAL n_observed = 0.0;
REAL sum_s = 0.0; //the sum of a SNP over all observed individuals
REAL sum2_s = 0.0; //the sum of the squares of the SNP over all observed individuals
size_t end = nIndividuals;
size_t delta = 1;
for( size_t ind = 0; ind < end; ind+=delta )
{
if (SNPs[ind] == SNPs[ind])
{
//check for not NaN
sum_s += SNPs[ ind ];
sum2_s+= SNPs[ ind ] * SNPs[ ind ];
++n_observed;
}
}
if ( n_observed < 1.0 )
{
printf( "No individual observed for the SNP.\n");
}
REAL mean_s = sum_s / n_observed; //compute the mean over observed individuals for the current SNP
REAL mean2_s = sum2_s / n_observed; //compute the mean of the squared SNP
//When beta standardization is being done, check that data is 0,1,2
if (betaNotUnitVariance && sum_s <= (REAL)0.0)
{
REAL freqT = sum_s/n_observed;
fprintf(stderr, "Observed SNP freq is %.2f. for a SNPs[:][%i]\n", freqT, iSnp );
exit(1);
}
//The SNP frequency
REAL freq = (sum_s) / (n_observed * (REAL)2.0); //compute snp-freq as in the Visscher Height paper (Nat Gen, Yang et al 2010).
if ((freq != freq) || betaNotUnitVariance && ((freq >= (REAL)1.0) || (freq <= (REAL)0.0)))
{
if (!seenSNC)
{
seenSNC = true;
fprintf(stderr, "Illegal SNP frequency: %.2f for SNPs[:][%i]\n", freq, iSnp);
}
}
REAL variance = mean2_s-mean_s * mean_s; //By the Cauchy Shwartz inequality this should always be positive
REAL std = sqrt( variance ); //The SNP frequency
bool isSNC = false;
if ( (std != std) || (std <= (REAL)0.0) )
{
// a std == 0.0 means all SNPs have the same value (no variation or Single Nucleotide Constant (SNC))
// however ALL SNCs should have been removed in previous filtering steps
// This test now prevents a divide by zero error below
std = 1.0;
isSNC = true;
if (!seenSNC)
{
seenSNC = true;
fprintf(stderr, "std=.%2f has illegal value for SNPs[:][%i]\n", std, iSnp );
}
}
if (betaNotUnitVariance && freq > .5)
{
freq = 1.0 - freq;
}
for( size_t ind = 0; ind < end; ind+=delta )
{
//check for NaN
if ( (SNPs[ ind ]!=SNPs[ ind ]) || isSNC)
{
SNPs[ ind ] = 0.0;
}
else
{
SNPs[ ind ] -= mean_s; //subtract the mean from the data
if (betaNotUnitVariance )
{
REAL rT = SUFFIX(BetaPdf)( freq, betaA, betaB );
//fprintf(stderr, "BetaPdf(%f,%f,%f)=%f\n", freq, betaA, betaB, rT);
SNPs[ ind ] *= rT;
}
else
{
SNPs[ ind ] /= std; //unit variance as well
}
}
}
SNPs += nIndividuals;
}
#else //Order C
// Make one pass through the data (by individual, because that is how it is laid out), collecting statistics
std::vector<REAL> n_observed(nSNPs); // C++ inits to 0's
std::vector<REAL> sum_s(nSNPs); //the sum of a SNP over all observed individuals. C++ inits to 0's
std::vector<REAL> sum2_s(nSNPs); //the sum of the squares of the SNP over all observed individuals. C++ inits to 0's
for( size_t ind = 0; ind < nIndividuals; ++ind)
{
size_t rowStart = ind * nSNPs;
for ( size_t iSnp = 0; iSnp < nSNPs; ++iSnp )
{
REAL value = SNPs[rowStart+iSnp];
if ( value == value )
{
sum_s[iSnp] += value;
sum2_s[iSnp] += value * value;
++n_observed[iSnp];
}
}
}
std::vector<REAL> mean_s(nSNPs); //compute the mean over observed individuals for the current SNP
std::vector<REAL> mean2_s(nSNPs); //compute the mean of the squared SNP
std::vector<REAL> std(nSNPs); //the standard deviation
std::vector<REAL> freq(nSNPs); //The SNP frequency
std::vector<bool> isSNC(nSNPs); // Is this a SNC (C++ inits to false)
for ( size_t iSnp = 0; iSnp < nSNPs; ++iSnp )
{
if ( n_observed[iSnp] < 1.0 )
{
printf( "No individual observed for the SNP.\n");
}
mean_s[iSnp] = sum_s[iSnp] / n_observed[iSnp]; //compute the mean over observed individuals for the current SNP
mean2_s[iSnp] = sum2_s[iSnp] / n_observed[iSnp]; //compute the mean of the squared SNP
//When beta standardization is being done, check that data is 0,1,2
if (betaNotUnitVariance && sum_s[iSnp] <= (REAL)0.0)
{
REAL freqT = sum_s[iSnp]/n_observed[iSnp];
fprintf(stderr, "Observed SNP freq is %.2f. for a SNPs[:][%i]\n", freqT, iSnp );
exit(1);
}
freq[iSnp] = (sum_s[iSnp]) / (n_observed[iSnp] * (REAL)2.0); //compute snp-freq[iSnp] as in the Visscher Height paper (Nat Gen, Yang et al 2010).
if ((freq[iSnp] != freq[iSnp]) || betaNotUnitVariance && ((freq[iSnp] >= (REAL)1.0) || (freq[iSnp] <= (REAL)0.0)))
{
if (!seenSNC)
{
seenSNC = true;
fprintf(stderr, "Illegal SNP frequency: %.2f for SNPs[:][%i]\n", freq[iSnp], iSnp);
}
}
REAL variance = mean2_s[iSnp]-mean_s[iSnp] * mean_s[iSnp]; //By the Cauchy Shwartz inequality this should always be positive
std[iSnp] = sqrt( variance );
if ( (std[iSnp] != std[iSnp]) || (std[iSnp] <= (REAL)0.0) )
{
// a std == 0.0 means all SNPs have the same value (no variation or Single Nucleotide Constant (SNC))
// however ALL SNCs should have been removed in previous filtering steps
// This test now prevents a divide by zero error below
std[iSnp] = 1.0;
isSNC[iSnp] = true;
if (!seenSNC)
{
seenSNC = true;
fprintf(stderr, "std=.%2f has illegal value for SNPs[:][%i]\n", std[iSnp], iSnp );
}
}
if (betaNotUnitVariance && freq[iSnp] > .5)
{
freq[iSnp] = 1.0 - freq[iSnp];
}
}
for( size_t ind = 0; ind < nIndividuals; ++ind)
{
size_t rowStart = ind * nSNPs;
for ( size_t iSnp = 0; iSnp < nSNPs; ++iSnp )
{
REAL value = SNPs[rowStart+iSnp];
//check for NaN
if ( (value != value) || isSNC[iSnp])
{
value = 0.0;
}
else
{
value -= mean_s[iSnp]; //subtract the mean from the data
if (betaNotUnitVariance )
{
REAL rT = SUFFIX(BetaPdf)( freq[iSnp], betaA, betaB );
//fprintf(stderr, "BetaPdf(%f,%f,%f)=%f\n", freq, betaA, betaB, rT);
value *= rT;
}
else
{
value /= std[iSnp]; //unit variance as well
}
}
SNPs[rowStart+iSnp] = value;
}
}
#endif
}
/*
* Parameters:
* SNPs [nIndividuals by nSNPs]:
* Matrix stored in column-major order.
* This will hold the result.
* NaNs will be set to 0.0 in the result.
*/
void SUFFIX(ImputeAndZeroMeanSNPs)(
REAL *SNPs,
const size_t nIndividuals,
const size_t nSNPs,
const bool betaNotUnitVariance,
const REAL betaA,
const REAL betaB,
const bool apply_in_place,
const bool use_stats,
REAL *stats
)
{
bool seenSNC = false; //Keep track of this so that only one warning message is reported
#ifdef ORDERF
for ( size_t iSnp = 0; iSnp < nSNPs; ++iSnp )
{
REAL mean_s;
REAL std;
REAL freq = 0;
size_t end = nIndividuals;
size_t delta = 1;
bool isSNC;
if (use_stats)
{
mean_s = stats[iSnp];
std = stats[iSnp + nSNPs];
isSNC = isinf(std);
}
else
{
isSNC = false;
REAL n_observed = 0.0;
REAL sum_s = 0.0; //the sum of a SNP over all observed individuals
REAL sum2_s = 0.0; //the sum of the squares of the SNP over all observed individuals
for (size_t ind = 0; ind < end; ind += delta)
{
if (SNPs[ind] == SNPs[ind])
{
//check for not NaN
sum_s += SNPs[ind];
sum2_s += SNPs[ind] * SNPs[ind];
++n_observed;
}
}
if (n_observed < 1.0)
{
printf("No individual observed for the SNP.\n");
//LATER make it work (in some form) for n of 0
}
mean_s = sum_s / n_observed; //compute the mean over observed individuals for the current SNP
REAL mean2_s = sum2_s / n_observed; //compute the mean of the squared SNP
if ((mean_s != mean_s) || betaNotUnitVariance && ((mean_s > (REAL)2.0) || (mean_s < (REAL)0.0)))
{
if (!seenSNC)
{
seenSNC = true;
fprintf(stderr, "Illegal SNP mean: %.2f for SNPs[:][%i]\n", mean_s, iSnp);
}
}
REAL variance = mean2_s - mean_s * mean_s; //By the Cauchy Shwartz inequality this should always be positive
std = sqrt(variance);
if ((std != std) || (std <= (REAL)0.0))
{
// a std == 0.0 means all SNPs have the same value (no variation or Single Nucleotide Constant (SNC))
// however ALL SNCs should have been removed in previous filtering steps
// This test now prevents a divide by zero error below
isSNC = true;
if (!seenSNC)
{
seenSNC = true;
fprintf(stderr, "std=.%2f has illegal value for SNPs[:][%i]\n", std, iSnp);
}
std = std::numeric_limits<REAL>::infinity();
}
stats[iSnp] = mean_s;
stats[iSnp + nSNPs] = std;
}
if (apply_in_place)
{
for (size_t ind = 0; ind < end; ind += delta)
{
//check for NaN
if ((SNPs[ind] != SNPs[ind]) || isSNC)
{
SNPs[ind] = 0.0;
}
else
{
SNPs[ind] -= mean_s; //subtract the mean from the data
if (betaNotUnitVariance) //compute snp-freq as in the Visscher Height paper (Nat Gen, Yang et al 2010).
{
REAL freq = mean_s / 2.0;
if (freq > .5)
{
freq = 1.0 - freq;
}
REAL rT = SUFFIX(BetaPdf)(freq, betaA, betaB);
//fprintf(stderr, "BetaPdf(%f,%f,%f)=%f\n", freq, betaA, betaB, rT);
SNPs[ind] *= rT;
}
else
{
SNPs[ind] /= std; //unit variance as well
}
}
}
}
SNPs += nIndividuals;
}
#else //Order C
std::vector<REAL> mean_s(nSNPs); //compute the mean over observed individuals for the current SNP
std::vector<REAL> std(nSNPs); //the standard deviation
std::vector<bool> isSNC(nSNPs); // Is this a SNC (C++ inits to false)
if (use_stats)
{
for (size_t iSnp = 0; iSnp < nSNPs; ++iSnp)
{
mean_s[iSnp] = stats[iSnp*2];
std[iSnp] = stats[iSnp * 2+1];
isSNC[iSnp] = isinf(std[iSnp]);
}
}
else
{
// Make one pass through the data (by individual, because that is how it is laid out), collecting statistics
std::vector<REAL> n_observed(nSNPs); // C++ inits to 0's
std::vector<REAL> sum_s(nSNPs); //the sum of a SNP over all observed individuals. C++ inits to 0's
std::vector<REAL> sum2_s(nSNPs); //the sum of the squares of the SNP over all observed individuals. C++ inits to 0's
for( size_t ind = 0; ind < nIndividuals; ++ind)
{
size_t rowStart = ind * nSNPs;
for ( size_t iSnp = 0; iSnp < nSNPs; ++iSnp )
{
REAL value = SNPs[rowStart+iSnp];
if ( value == value )
{
sum_s[iSnp] += value;
sum2_s[iSnp] += value * value;
++n_observed[iSnp];
}
}
}
std::vector<REAL> mean2_s(nSNPs); //compute the mean of the squared SNP
for (size_t iSnp = 0; iSnp < nSNPs; ++iSnp)
{
if (n_observed[iSnp] < 1.0)
{
printf("No individual observed for the SNP.\n");
}
mean_s[iSnp] = sum_s[iSnp] / n_observed[iSnp]; //compute the mean over observed individuals for the current SNP
mean2_s[iSnp] = sum2_s[iSnp] / n_observed[iSnp]; //compute the mean of the squared SNP
if ((mean_s[iSnp] != mean_s[iSnp]) || betaNotUnitVariance && ((mean_s[iSnp] > (REAL)2.0) || (mean_s[iSnp] < (REAL)0.0)))
{
if (!seenSNC)
{
seenSNC = true;
fprintf(stderr, "Illegal SNP mean: %.2f for SNPs[:][%i]\n", mean_s[iSnp], iSnp);
}
}
REAL variance = mean2_s[iSnp] - mean_s[iSnp] * mean_s[iSnp]; //By the Cauchy Shwartz inequality this should always be positive
std[iSnp] = sqrt(variance);
if ((std[iSnp] != std[iSnp]) || (std[iSnp] <= (REAL)0.0))
{
// a std == 0.0 means all SNPs have the same value (no variation or Single Nucleotide Constant (SNC))
// however ALL SNCs should have been removed in previous filtering steps
// This test now prevents a divide by zero error below
std[iSnp] = std::numeric_limits<REAL>::infinity();
isSNC[iSnp] = true;
if (!seenSNC)
{
seenSNC = true;
fprintf(stderr, "std=.%2f has illegal value for SNPs[:][%i]\n", std[iSnp], iSnp);
}
}
stats[iSnp*2] = mean_s[iSnp];
stats[iSnp*2+1] = std[iSnp];
}
}
if (apply_in_place)
{
for (size_t ind = 0; ind < nIndividuals; ++ind)
{
size_t rowStart = ind * nSNPs;
for (size_t iSnp = 0; iSnp < nSNPs; ++iSnp)
{
REAL value = SNPs[rowStart + iSnp];
//check for NaN
if ((value != value) || isSNC[iSnp])
{
value = 0.0;
}
else
{
value -= mean_s[iSnp]; //subtract the mean from the data
if (betaNotUnitVariance)
{
//compute snp-freq as in the Visscher Height paper (Nat Gen, Yang et al 2010).
REAL freq = mean_s[iSnp] / 2.0;
if (freq > .5)
{
freq = 1.0 - freq;
}
REAL rT = SUFFIX(BetaPdf)(freq, betaA, betaB);
//fprintf(stderr, "BetaPdf(%f,%f,%f)=%f\n", freq, betaA, betaB, rT);
value *= rT;
}
else
{
value /= std[iSnp]; //unit variance as well
}
}
SNPs[rowStart + iSnp] = value;
}
}
}
#endif
}