choose_highly_variable_genes.hpp

Go to the documentation of this file.

#ifndef SCRAN_VARIANCES_CHOOSE_HIGHLY_VARIABLE_GENES_HPP
#define SCRAN_VARIANCES_CHOOSE_HIGHLY_VARIABLE_GENES_HPP
 
#include <vector>
#include <algorithm>
#include <numeric>
#include <cstdint>
 
namespace scran_variances {
 
struct ChooseHighlyVariableGenesOptions {
    size_t top = 4000;
 
    bool larger = true;
 
    bool use_bound = false;
 
    double bound = 0;
 
    bool keep_ties = true;
};
 
namespace internal {
 
template<bool keep_index_, typename Index_, typename Stat_, class Output_, class Cmp_, class CmpEqual_>
void choose_highly_variable_genes(Index_ n, const Stat_* statistic, Output_& output, Cmp_ cmp, CmpEqual_ cmpeq, const ChooseHighlyVariableGenesOptions& options) {
    if (options.top == 0) {
        if constexpr(keep_index_) {
            ; // no-op, we assume it's already empty.
        } else {
            std::fill_n(output, n, false);
        }
        return;
    }
 
    Stat_ bound = options.bound;
    if (static_cast<size_t>(options.top) >= static_cast<size_t>(n)) {
        if (options.use_bound) {
            for (Index_ i = 0; i < n; ++i) {
                bool ok = cmp(statistic[i], bound);
                if constexpr(keep_index_) {
                    if (ok) {
                        output.push_back(i);
                    }
                } else {
                    output[i] = ok;
                }
            }
        } else {
            if constexpr(keep_index_) {
                output.resize(n);
                std::iota(output.begin(), output.end(), static_cast<Index_>(0));
            } else {
                std::fill_n(output, n, true);
            }
        }
        return;
    }
 
    std::vector<Index_> semi_sorted(n);
    std::iota(semi_sorted.begin(), semi_sorted.end(), static_cast<Index_>(0));
    auto cBegin = semi_sorted.begin(), cMid = cBegin + options.top - 1, cEnd = semi_sorted.end();
    std::nth_element(cBegin, cMid, cEnd, [&](Index_ l, Index_ r) -> bool { 
        auto L = statistic[l], R = statistic[r];
        if (L == R) {
            return l < r; // always favor the earlier index for a stable sort, even if options.larger = false.
        } else {
            return cmp(L, R);
        }
    });
 
    Stat_ threshold = statistic[semi_sorted[options.top - 1]];
 
    if (options.keep_ties) {
        if (options.use_bound && !cmp(threshold, bound)) {
            for (Index_ i = 0; i < n; ++i) {
                bool ok = cmp(statistic[i], bound);
                if constexpr(keep_index_) {
                    if (ok) {
                        output.push_back(i);
                    }
                } else {
                    output[i] = ok;
                }
            }
        } else {
            for (Index_ i = 0; i < n; ++i) {
                bool ok = cmpeq(statistic[i], threshold);
                if constexpr(keep_index_) {
                    if (ok) {
                        output.push_back(i);
                    }
                } else {
                    output[i] = ok;
                }
            }
        }
        return;
    }
 
    if constexpr(keep_index_) {
        output.reserve(options.top);
    } else {
        std::fill_n(output, n, false);
    }
 
    if (options.use_bound) {
        Index_ counter = options.top;
        while (counter > 0) {
            --counter;
            auto pos = semi_sorted[counter];
            if (cmp(statistic[pos], bound)) {
                if constexpr(keep_index_) {
                    output.push_back(pos);
                } else {
                    output[pos] = true;
                }
            }
        }
    } else {
        if constexpr(keep_index_) {
            output.insert(output.end(), semi_sorted.begin(), semi_sorted.begin() + options.top);
        } else {
            for (Index_ i = 0, end = options.top; i < end; ++i) {
                output[semi_sorted[i]] = true;
            }
        }
    }
 
    if constexpr(keep_index_) {
        std::sort(output.begin(), output.end());
    }
}
 
}
template<typename Stat_, typename Bool_>
void choose_highly_variable_genes(size_t n, const Stat_* statistic, Bool_* output, const ChooseHighlyVariableGenesOptions& options) {
    if (options.larger) {
        internal::choose_highly_variable_genes<false>(
            n, 
            statistic, 
            output, 
            [](Stat_ l, Stat_ r) -> bool { return l > r; },
            [](Stat_ l, Stat_ r) -> bool { return l >= r; },
            options
        );
    } else {
        internal::choose_highly_variable_genes<false>(
            n, 
            statistic, 
            output, 
            [](Stat_ l, Stat_ r) -> bool { return l < r; },
            [](Stat_ l, Stat_ r) -> bool { return l <= r; },
            options
        );
    }
}
 
template<typename Bool_ = uint8_t, typename Stat_>
std::vector<Bool_> choose_highly_variable_genes(size_t n, const Stat_* statistic, const ChooseHighlyVariableGenesOptions& options) {
    std::vector<Bool_> output(n
#ifdef SCRAN_VARIANCES_TEST_INIT
        , SCRAN_VARIANCES_TEST_INIT
#endif
    );
    choose_highly_variable_genes(n, statistic, output.data(), options);
    return output;
}
 
template<typename Index_, typename Stat_>
std::vector<Index_> choose_highly_variable_genes_index(Index_ n, const Stat_* statistic, const ChooseHighlyVariableGenesOptions& options) {
    std::vector<Index_> output;
    if (options.larger) {
        internal::choose_highly_variable_genes<true>(
            n, 
            statistic, 
            output,
            [](Stat_ l, Stat_ r) -> bool { return l > r; },
            [](Stat_ l, Stat_ r) -> bool { return l >= r; },
            options
        );
    } else {
        internal::choose_highly_variable_genes<true>(
            n, 
            statistic, 
            output,
            [](Stat_ l, Stat_ r) -> bool { return l < r; },
            [](Stat_ l, Stat_ r) -> bool { return l <= r; },
            options
        );
    }
    return output;
}
 
}
 
#endif