score_markers_best.hpp

Go to the documentation of this file.

#ifndef SCRAN_MARKERS_SCORE_MARKERS_BEST_HPP
#define SCRAN_MARKERS_SCORE_MARKERS_BEST_HPP
 
#include <vector>
#include <cstddef>
#include <optional>
 
#include "scran_blocks/scran_blocks.hpp"
#include "tatami/tatami.hpp"
#include "sanisizer/sanisizer.hpp"
#include "topicks/topicks.hpp"
#include "quickstats/quickstats.hpp"
 
#include "scan_matrix.hpp"
#include "average_group_stats.hpp"
#include "block_averages.hpp"
#include "create_combinations.hpp"
#include "cohens_d.hpp"
#include "simple_diff.hpp"
#include "utils.hpp"
 
namespace scran_markers {
 
struct ScoreMarkersBestOptions {
    double threshold = 0;
 
    int num_threads = 1;
 
    bool compute_group_mean = true;
 
    bool compute_group_detected = true;
 
    bool compute_cohens_d = true;
 
    bool compute_auc = true;
 
    bool compute_delta_mean = true;
 
    bool compute_delta_detected = true;
 
    bool largest_cohens_d = true;
 
    bool largest_auc = true;
 
    bool largest_delta_mean = true;
 
    bool largest_delta_detected = true;
 
    std::optional<double> threshold_cohens_d = 0;
 
    std::optional<double> threshold_auc = 0.5;
 
    std::optional<double> threshold_delta_mean = 0;
 
    std::optional<double> threshold_delta_detected = 0;
 
    bool keep_ties = false;
 
    BlockAveragePolicy block_average_policy = BlockAveragePolicy::MEAN;
 
    scran_blocks::WeightPolicy block_weight_policy = scran_blocks::WeightPolicy::VARIABLE;
 
    scran_blocks::VariableWeightParameters variable_block_weight_parameters;
 
    double block_quantile = 0.5;
};
 
template<typename Stat_, typename Index_>
struct ScoreMarkersBestResults {
    std::vector<std::vector<Stat_> > mean;
 
    std::vector<std::vector<Stat_> > detected;
 
    std::vector<std::vector<topicks::TopQueue<Stat_, Index_> > > cohens_d;
 
    std::vector<std::vector<topicks::TopQueue<Stat_, Index_> > > auc;
 
    std::vector<std::vector<topicks::TopQueue<Stat_, Index_> > > delta_mean;
 
    std::vector<std::vector<topicks::TopQueue<Stat_, Index_> > > delta_detected;
};
 
template<typename Stat_, typename Index_>
std::vector<std::vector<std::vector<std::pair<Index_, Stat_> > > > queues_to_vectors(std::vector<std::vector<topicks::TopQueue<Stat_, Index_> > >& queued) {
    std::vector<std::vector<std::vector<std::pair<Index_, Stat_> > > > output;
    const auto num_groups = queued.size();
    sanisizer::resize(output, num_groups);
 
    for (I<decltype(num_groups)> g1 = 0; g1 < num_groups; ++g1) {
        sanisizer::resize(output[g1], num_groups);
        for (I<decltype(num_groups)> g2 = 0; g2 < num_groups; ++g2) {
            if (g1 == g2) {
                continue;
            }
 
            auto& current_in = queued[g1][g2];
            auto& current_out = output[g1][g2];
            current_out.reserve(current_in.size());
 
            while (!current_in.empty()) {
                const auto& best = current_in.top();
                current_out.emplace_back(best.second, best.first);
                current_in.pop();
            }
            std::reverse(current_out.begin(), current_out.end()); // earliest element should have the strongest effect sizes.
        }
    }
 
    return output;
}
 
namespace internal {
 
template<typename Stat_, typename Index_>
using PairwiseTopQueues = std::vector<std::vector<topicks::TopQueue<Stat_, Index_> > >;
 
template<typename Stat_, typename Index_>
void allocate_best_top_queues(
    PairwiseTopQueues<Stat_, Index_>& pqueues,
    const std::size_t num_groups,
    const Index_ top,
    const bool larger,
    const bool keep_ties,
    const std::optional<Stat_>& bound
) {
    topicks::TopQueueOptions<Stat_> opt;
    opt.check_nan = true;
    opt.keep_ties = keep_ties;
    if (bound.has_value()) {
        opt.bound = *bound;
    }
 
    sanisizer::resize(pqueues, num_groups);
    for (I<decltype(num_groups)> g1 = 0; g1 < num_groups; ++g1) {
        auto& x = pqueues[g1];
        x.reserve(num_groups);
        for (I<decltype(num_groups)> g2 = 0; g2 < num_groups; ++g2) {
            if (g1 == g2) {
                x.emplace_back(0, larger, opt); 
            } else {
                x.emplace_back(top, larger, opt);
            }
        }
    }
}
 
template<typename Stat_, typename Index_>
void add_best_top_queues(
    PairwiseTopQueues<Stat_, Index_>& pqueues,
    const Index_ gene,
    std::size_t num_groups,
    const std::vector<Stat_>& effects
) {
    for (I<decltype(num_groups)> g1 = 0; g1 < num_groups; ++g1) {
        for (I<decltype(num_groups)> g2 = 0; g2 < num_groups; ++g2) {
            const auto val = effects[sanisizer::nd_offset<std::size_t>(g2, num_groups, g1)];
            if (g1 != g2) {
                pqueues[g1][g2].emplace(val, gene);
            }
        }
    }
}
 
template<typename Stat_, typename Index_>
void report_best_top_queues(
    std::vector<std::optional<PairwiseTopQueues<Stat_, Index_> > >& pqueues,
    std::size_t num_groups,
    std::vector<std::vector<topicks::TopQueue<Stat_, Index_> > >& output
) {
    // We know it fits into an 'int' as this is what we got originally.
    const int num_available = pqueues.size();
 
    // If it's empty, we just create empty vectors and move on.
    if (num_available == 0) {
        sanisizer::resize(output, num_groups);
        topicks::TopQueue<Stat_, Index_> placeholder(0, false, {});
        for (I<decltype(num_groups)> g1 = 0; g1 < num_groups; ++g1) {
            sanisizer::resize(output[g1], num_groups, placeholder);
        }
        return;
    }
 
    // Consolidating all of the thread-specific queues into a single queue.
    auto& true_pqueue = *(pqueues.front());
    for (int t = 1; t < num_available; ++t) {
        auto& current_pqueue = *(pqueues[t]);
        for (I<decltype(num_groups)> g1 = 0; g1 < num_groups; ++g1) {
            for (I<decltype(num_groups)> g2 = 0; g2 < num_groups; ++g2) {
                auto& current_in = current_pqueue[g1][g2];
                auto& current_out = true_pqueue[g1][g2];
                while (!current_in.empty()) {
                    current_out.push(current_in.top());
                    current_in.pop();
                }
            }
        }
    }
 
    output = std::move(true_pqueue);
}
 
template<typename Index_, typename Stat_>
void find_best_simple_best_effects(
    const Index_ ngenes,
    const std::size_t num_groups,
    const std::size_t num_blocks,
    const std::size_t num_combos,
    const std::vector<Stat_>& combo_means,
    const std::vector<Stat_>& combo_vars,
    const std::vector<Stat_>& combo_detected,
    const BlockAverageInfo<Stat_>& average_info,
    const Index_ top,
    const ScoreMarkersBestOptions& options,
    ScoreMarkersBestResults<Stat_, Index_>& output
) {
    std::optional<std::vector<Stat_> > total_weights_per_group;
    const Stat_* total_weights_ptr = NULL;
    if (average_info.use_mean()) {
        if (options.compute_group_mean || options.compute_group_detected) {
            if (num_blocks > 1) {
                total_weights_per_group = compute_total_weight_per_group(num_groups, num_blocks, average_info.combo_weights().data());
                total_weights_ptr = total_weights_per_group->data();
            } else {
                total_weights_ptr = average_info.combo_weights().data();
            }
        }
    }
 
    std::vector<Stat_*> mptrs;
    if (options.compute_group_mean) {
        mptrs.reserve(num_groups);
        sanisizer::resize(output.mean, num_groups);
        for (auto& x : output.mean) {
            sanisizer::resize(x, ngenes);
            mptrs.push_back(x.data());
        }
    }
 
    std::vector<Stat_*> dptrs;
    if (options.compute_group_detected) {
        dptrs.reserve(num_groups);
        sanisizer::resize(output.detected, num_groups);
        for (auto& x : output.detected) {
            sanisizer::resize(x, ngenes);
            dptrs.push_back(x.data());
        }
    }
 
    std::optional<PrecomputedPairwiseWeights<Stat_> > preweights;
    if (average_info.use_mean()) {
        if (options.compute_cohens_d || options.compute_delta_mean || options.compute_delta_detected) {
            preweights.emplace(num_groups, num_blocks, average_info.combo_weights().data());
        }
    }
 
    // Setting up the output queues.
    std::optional<std::vector<std::optional<PairwiseTopQueues<Stat_, Index_> > > > threaded_cohens_d_queues, threaded_delta_detected_queues, threaded_delta_mean_queues;
    if (options.compute_cohens_d) {
        threaded_cohens_d_queues.emplace(sanisizer::cast<I<decltype(threaded_cohens_d_queues->size())> >(options.num_threads));
    }
    if (options.compute_delta_mean) {
        threaded_delta_mean_queues.emplace(sanisizer::cast<I<decltype(threaded_delta_mean_queues->size())> >(options.num_threads));
    }
    if (options.compute_delta_detected) {
        threaded_delta_detected_queues.emplace(sanisizer::cast<I<decltype(threaded_delta_detected_queues->size())> >(options.num_threads));
    }
 
    const auto num_groups2 = sanisizer::product<typename std::vector<Stat_>::size_type>(num_groups, num_groups);
 
    int num_used = tatami::parallelize([&](const int t, const Index_ start, const Index_ length) -> void {
        std::optional<PairwiseTopQueues<Stat_, Index_> > local_cohens_d_queue, local_delta_mean_queue, local_delta_detected_queue;
        if (options.compute_cohens_d) {
            local_cohens_d_queue.emplace();
            allocate_best_top_queues(*local_cohens_d_queue, num_groups, top, options.largest_cohens_d, options.keep_ties, options.threshold_cohens_d);
        }
        if (options.compute_delta_mean) {
            local_delta_mean_queue.emplace();
            allocate_best_top_queues(*local_delta_mean_queue, num_groups, top, options.largest_delta_mean, options.keep_ties, options.threshold_delta_mean);
        }
        if (options.compute_delta_detected) {
            local_delta_detected_queue.emplace();
            allocate_best_top_queues(*local_delta_detected_queue, num_groups, top, options.largest_delta_detected, options.keep_ties, options.threshold_delta_detected);
        }
 
        std::vector<Stat_> buffer;
        if (options.compute_cohens_d || options.compute_delta_mean || options.compute_delta_detected) {
            buffer.resize(num_groups2);
        }
 
        std::optional<std::vector<Stat_> > qbuffer, qrevbuffer;
        std::optional<quickstats::SingleQuantileVariableNumber<Stat_> > qcalc;
        if (!average_info.use_mean()) {
            qbuffer.emplace();
            qrevbuffer.emplace();
            qcalc.emplace(num_blocks, average_info.quantile());
        }
 
        for (Index_ gene = start, end = start + length; gene < end; ++gene) {
            auto in_offset = sanisizer::product_unsafe<std::size_t>(gene, num_combos);
 
            if (options.compute_group_mean) {
                const auto tmp_means = combo_means.data() + in_offset;
                if (average_info.use_mean()) {
                    average_group_stats_blockmean(gene, num_groups, num_blocks, tmp_means, average_info.combo_weights().data(), total_weights_ptr, mptrs);
                } else {
                    average_group_stats_blockquantile(gene, num_groups, num_blocks, tmp_means, *qbuffer, *qcalc, mptrs);
                }
            }
 
            if (options.compute_group_detected) {
                const auto tmp_detected = combo_detected.data() + in_offset;
                if (average_info.use_mean()) {
                    average_group_stats_blockmean(gene, num_groups, num_blocks, tmp_detected, average_info.combo_weights().data(), total_weights_ptr, dptrs);
                } else {
                    average_group_stats_blockquantile(gene, num_groups, num_blocks, tmp_detected, *qbuffer, *qcalc, dptrs);
                }
            }
 
            // Computing the effect sizes.
            if (options.compute_cohens_d) {
                const auto tmp_means = combo_means.data() + in_offset;
                const auto tmp_variances = combo_vars.data() + in_offset;
                if (average_info.use_mean()) {
                    compute_pairwise_cohens_d_blockmean(tmp_means, tmp_variances, num_groups, num_blocks, options.threshold, *preweights, buffer.data());
                } else {
                    compute_pairwise_cohens_d_blockquantile(tmp_means, tmp_variances, num_groups, num_blocks, options.threshold, *qbuffer, *qrevbuffer, *qcalc, buffer.data());
                }
                add_best_top_queues(*local_cohens_d_queue, gene, num_groups, buffer);
            }
 
            if (options.compute_delta_mean) {
                const auto tmp_means = combo_means.data() + in_offset;
                if (average_info.use_mean()) {
                    compute_pairwise_simple_diff_blockmean(tmp_means, num_groups, num_blocks, *preweights, buffer.data());
                } else {
                    compute_pairwise_simple_diff_blockquantile(tmp_means, num_groups, num_blocks, *qbuffer, *qcalc, buffer.data());
                }
                add_best_top_queues(*local_delta_mean_queue, gene, num_groups, buffer);
            }
 
            if (options.compute_delta_detected) {
                const auto tmp_detected = combo_detected.data() + in_offset;
                if (average_info.use_mean()) {
                    compute_pairwise_simple_diff_blockmean(tmp_detected, num_groups, num_blocks, *preweights, buffer.data());
                } else {
                    compute_pairwise_simple_diff_blockquantile(tmp_detected, num_groups, num_blocks, *qbuffer, *qcalc, buffer.data());
                }
                add_best_top_queues(*local_delta_detected_queue, gene, num_groups, buffer);
            }
        }
 
        // Only moving it into the shared buffer at the end to minimize false sharing.
        if (options.compute_cohens_d) {
            (*threaded_cohens_d_queues)[t] = std::move(local_cohens_d_queue);
        }
        if (options.compute_delta_mean) {
            (*threaded_delta_mean_queues)[t] = std::move(local_delta_mean_queue);
        }
        if (options.compute_delta_detected) {
            (*threaded_delta_detected_queues)[t] = std::move(local_delta_detected_queue);
        }
    }, ngenes, options.num_threads);
 
    // Now figuring out which of these are the top dogs.
    if (options.compute_cohens_d) {
        threaded_cohens_d_queues->resize(num_used);
        report_best_top_queues(*threaded_cohens_d_queues, num_groups, output.cohens_d);
    }
    if (options.compute_delta_mean) {
        threaded_delta_mean_queues->resize(num_used);
        report_best_top_queues(*threaded_delta_mean_queues, num_groups, output.delta_mean);
    }
    if (options.compute_delta_detected) {
        threaded_delta_detected_queues->resize(num_used);
        report_best_top_queues(*threaded_delta_detected_queues, num_groups, output.delta_detected);
    }
}
 
template<
    bool single_block_,
    typename Stat_,
    typename Value_,
    typename Index_,
    typename Group_,
    typename Block_
>
ScoreMarkersBestResults<Stat_, Index_> score_markers_best(
    const tatami::Matrix<Value_, Index_>& matrix, 
    const Group_* const group, 
    const std::size_t num_groups,
    const Block_* const block,
    const std::size_t num_blocks,
    const std::size_t* const combo,
    const std::size_t num_combos,
    const std::vector<Index_>& combo_sizes,
    const Index_ top,
    const ScoreMarkersBestOptions& options
) {
    const auto ngenes = matrix.nrow();
    const auto payload_size = sanisizer::product<typename std::vector<Stat_>::size_type>(ngenes, num_combos);
    std::vector<Stat_> combo_means, combo_vars, combo_detected;
    if (options.compute_group_mean || options.compute_cohens_d || options.compute_delta_mean) {
        combo_means.resize(payload_size);
    }
    if (options.compute_cohens_d) {
        combo_vars.resize(payload_size);
    }
    if (options.compute_group_detected || options.compute_delta_detected) {
        combo_detected.resize(payload_size);
    }
 
    // For a single block, this usually doesn't really matter, but we do it for consistency with the multi-block case,
    // and to account for variable weighting where non-zero block sizes get zero weight.
    BlockAverageInfo<Stat_> average_info;
    if (options.block_average_policy == BlockAveragePolicy::MEAN) {
        average_info = BlockAverageInfo<Stat_>(
            scran_blocks::compute_weights<Stat_>(
                combo_sizes,
                options.block_weight_policy,
                options.variable_block_weight_parameters
            )
        );
    } else {
        average_info = BlockAverageInfo<Stat_>(options.block_quantile);
    }
 
    ScoreMarkersBestResults<Stat_, Index_> output;
 
    if (options.compute_auc) {
        auto auc_queues = sanisizer::create<std::vector<std::optional<PairwiseTopQueues<Stat_, Index_> > > >(options.num_threads);
 
        struct AucResultWorkspace {
            AucResultWorkspace(const std::size_t num_groups) : pairwise_buffer(sanisizer::product<typename std::vector<Stat_>::size_type>(num_groups, num_groups)) {};
            std::vector<Stat_> pairwise_buffer;
            PairwiseTopQueues<Stat_, Index_> queue;
        };
 
        const auto num_used = scan_matrix_by_row_custom_auc<single_block_>(
            matrix, 
            group,
            num_groups,
            block,
            num_blocks,
            combo,
            num_combos,
            combo_sizes,
            average_info,
            combo_means,
            combo_vars,
            combo_detected,
            /* do_auc = */ true,
            /* auc_result_initialize = */ [&](const int) -> AucResultWorkspace {
                AucResultWorkspace res_work(num_groups);
                allocate_best_top_queues(res_work.queue, num_groups, top, options.largest_auc, options.keep_ties, options.threshold_auc);
                return res_work;
            },
            /* auc_result_process = */ [&](const Index_ gene, AucScanWorkspace<Value_, Group_, Stat_, Index_>& auc_work, AucResultWorkspace& res_work) -> void {
                process_auc_for_rows(auc_work, num_groups, num_blocks, options.threshold, res_work.pairwise_buffer.data());
                add_best_top_queues(res_work.queue, gene, num_groups, res_work.pairwise_buffer);
            },
            /* auc_result_finalize = */ [&](const int t, AucResultWorkspace& res_work) -> void {
                auc_queues[t] = std::move(res_work.queue);
            },
            options.num_threads
        );
 
        auc_queues.resize(num_used);
        report_best_top_queues(auc_queues, num_groups, output.auc);
 
    } else if (matrix.prefer_rows()) {
        scan_matrix_by_row_full_auc<single_block_>(
            matrix, 
            group,
            num_groups,
            block,
            num_blocks,
            combo,
            num_combos,
            combo_sizes,
            average_info,
            combo_means,
            combo_vars,
            combo_detected,
            static_cast<Stat_*>(NULL),
            options.threshold,
            options.num_threads
        );
 
    } else {
        scan_matrix_by_column(
            matrix,
            [&]{
                if constexpr(single_block_) {
                    return group;
                } else {
                    return combo;
                }
            }(),
            [&]{
                if constexpr(single_block_) {
                    return num_groups;
                } else {
                    return num_combos;
                }
            }(),
            combo_sizes,
            combo_means,
            combo_vars,
            combo_detected,
            options.num_threads
        );
    }
 
    find_best_simple_best_effects(
        matrix.nrow(),
        num_groups,
        num_blocks,
        num_combos,
        combo_means,
        combo_vars,
        combo_detected,
        average_info,
        top,
        options,
        output
    );
 
    return output;
}
 
}
template<typename Stat_, typename Value_, typename Index_, typename Group_>
ScoreMarkersBestResults<Stat_, Index_> score_markers_best(
    const tatami::Matrix<Value_, Index_>& matrix, 
    const Group_* const group, 
    const std::size_t num_groups,
    const Index_ top,
    const ScoreMarkersBestOptions& options
) {
    const auto group_sizes = tabulate_groups(matrix.ncol(), group, num_groups); 
    return internal::score_markers_best<true, Stat_>(
        matrix,
        group,
        num_groups,
        static_cast<int*>(NULL),
        1,
        static_cast<std::size_t*>(NULL),
        num_groups,
        group_sizes,
        top,
        options
    );
}
 
template<typename Stat_, typename Value_, typename Index_, typename Group_, typename Block_>
ScoreMarkersBestResults<Stat_, Index_> score_markers_best_blocked(
    const tatami::Matrix<Value_, Index_>& matrix, 
    const Group_* const group, 
    const std::size_t num_groups,
    const Block_* const block,
    const std::size_t num_blocks,
    const Index_ top,
    const ScoreMarkersBestOptions& options
) {
    const auto combo_out = create_combinations(matrix.ncol(), group, num_groups, block, num_blocks);
    return internal::score_markers_best<false, Stat_>(
        matrix,
        group,
        num_groups,
        block,
        num_blocks,
        combo_out.combinations.data(),
        combo_out.num_combinations,
        combo_out.frequencies,
        top,
        options
    );
}
 
}
 
#endif