Add probability of overlap and weighted containment for Multisearch matches

Cargo.toml

@@ -27,6 +27,8 @@ camino = "1.1.9"
 glob = "0.3.1"
 rustworkx-core = "0.15.1"
 streaming-stats = "0.2.3"
+rust_decimal = { version = "1.36.0", features = ["maths"] }
+rust_decimal_macros = "1.36.0"
 
 [dev-dependencies]
 assert_cmd = "2.0.16"

src/lib.rs

@@ -22,6 +22,7 @@ mod manysearch_rocksdb;
 mod manysketch;
 mod multisearch;
 mod pairwise;
+mod search_significance;
 mod singlesketch;
 
 use camino::Utf8PathBuf as PathBuf;
@@ -231,7 +232,7 @@ fn do_check(index: String, quick: bool) -> anyhow::Result<u8> {
 }
 
 #[pyfunction]
-#[pyo3(signature = (querylist_path, siglist_path, threshold, ksize, scaled, moltype, estimate_ani, output_path=None))]
+#[pyo3(signature = (querylist_path, siglist_path, threshold, ksize, scaled, moltype, estimate_ani, estimate_prob_overlap, output_path=None))]
 #[allow(clippy::too_many_arguments)]
 fn do_multisearch(
     querylist_path: String,
@@ -241,6 +242,7 @@ fn do_multisearch(
     scaled: Option<u32>,
     moltype: String,
     estimate_ani: bool,
+    estimate_prob_overlap: bool,
     output_path: Option<String>,
 ) -> anyhow::Result<u8> {
     let _ = env_logger::try_init();
@@ -255,6 +257,7 @@ fn do_multisearch(
         selection,
         allow_failed_sigpaths,
         estimate_ani,
+        estimate_prob_overlap,
         output_path,
     ) {
         Ok(_) => Ok(0),

src/multisearch.rs

@@ -3,12 +3,134 @@ use anyhow::Result;
 use rayon::prelude::*;
 use sourmash::selection::Selection;
 use sourmash::signature::SigsTrait;
+use sourmash::sketch::minhash::KmerMinHash;
+use std::collections::HashMap;
 use std::sync::atomic;
 use std::sync::atomic::AtomicUsize;
 
+use crate::search_significance::{
+    compute_inverse_document_frequency, get_hash_frequencies, get_prob_overlap,
+    get_term_frequency_inverse_document_frequency, merge_all_minhashes, Normalization,
+};
+use crate::utils::multicollection::SmallSignature;
 use crate::utils::{csvwriter_thread, load_collection, MultiSearchResult, ReportType};
 use sourmash::ani_utils::ani_from_containment;
 
+#[derive(Default, Clone, Debug)]
+struct ProbOverlapStats {
+    prob_overlap: f64,
+    prob_overlap_adjusted: f64,
+    containment_adjusted: f64,
+    containment_adjusted_log10: f64,
+    tf_idf_score: f64,
+}
+
+/// Computes probability overlap statistics for a single pair of signatures
+fn compute_single_prob_overlap(
+    query: &SmallSignature,
+    against: &SmallSignature,
+    n_comparisons: f64,
+    query_merged_frequencies: &HashMap<u64, f64>,
+    against_merged_frequencies: &HashMap<u64, f64>,
+    query_term_frequencies: &HashMap<String, HashMap<u64, f64>>,
+    inverse_document_frequency: &HashMap<u64, f64>,
+    containment_query_in_target: f64,
+) -> ProbOverlapStats {
+    let overlapping_hashvals: Vec<u64> = query
+        .minhash
+        .intersection(&against.minhash)
+        .expect("Intersection of query and against minhashes")
+        .0;
+
+    let prob_overlap = get_prob_overlap(
+        &overlapping_hashvals,
+        query_merged_frequencies,
+        against_merged_frequencies,
+    );
+
+    let prob_overlap_adjusted = prob_overlap * n_comparisons;
+    let containment_adjusted = containment_query_in_target / prob_overlap_adjusted;
+
+    ProbOverlapStats {
+        prob_overlap,
+        prob_overlap_adjusted,
+        containment_adjusted,
+        containment_adjusted_log10: containment_adjusted.log10(),
+        tf_idf_score: get_term_frequency_inverse_document_frequency(
+            &overlapping_hashvals,
+            &query_term_frequencies[&query.md5sum],
+            inverse_document_frequency,
+        ),
+    }
+}
+
+/// Computes probability overlap statistics for queries and against signatures
+/// Estimate probability of overlap between query sig and against sig, using
+/// underlying distribution of hashvals for all queries and all againsts
+fn compute_prob_overlap_stats(
+    queries: &Vec<SmallSignature>,
+    againsts: &Vec<SmallSignature>,
+) -> (
+    f64,
+    HashMap<u64, f64>,
+    HashMap<u64, f64>,
+    HashMap<String, HashMap<u64, f64>>,
+    HashMap<u64, f64>,
+) {
+    let n_comparisons = againsts.len() as f64 * queries.len() as f64;
+
+    // Combine all the queries and against into a single signature each
+    eprintln!("Merging queries ...");
+    let queries_merged_mh: KmerMinHash =
+        merge_all_minhashes(queries).expect("Merging query minhashes");
+    eprintln!("\tDone.\n");
+
+    eprintln!("Merging against ...");
+    let against_merged_mh: KmerMinHash =
+        merge_all_minhashes(againsts).expect("Merging against minhashes");
+    eprintln!("\tDone.\n");
+
+    // Compute IDF
+    eprintln!("Computing Inverse Document Frequency (IDF) of hashes in all againsts ...");
+    let inverse_document_frequency =
+        compute_inverse_document_frequency(&against_merged_mh, againsts, Some(true));
+    eprintln!("\tDone.\n");
+
+    // Compute frequencies
+    eprintln!("Computing frequency of hashvals across all againsts (L1 Norm) ...");
+    let against_merged_frequencies =
+        get_hash_frequencies(&against_merged_mh, Some(Normalization::L1));
+    eprintln!("\tDone.\n");
+
+    eprintln!("Computing frequency of hashvals across all queries (L1 Norm) ...");
+    let query_merged_frequencies =
+        get_hash_frequencies(&queries_merged_mh, Some(Normalization::L1));
+    eprintln!("\tDone.\n");
+
+    // Compute term frequencies
+    eprintln!("Computing hashval term frequencies within each query (L2 Norm) ...");
+    let query_term_frequencies = HashMap::from(
+        queries
+            .par_iter()
+            .map(|query| {
+                (
+                    query.md5sum.clone(),
+                    get_hash_frequencies(&query.minhash, Some(Normalization::L2)),
+                )
+            })
+            .collect::<HashMap<String, HashMap<u64, f64>>>(),
+    );
+    eprintln!("\tDone.\n");
+
+    (
+        n_comparisons,
+        query_merged_frequencies,
+        against_merged_frequencies,
+        query_term_frequencies,
+        inverse_document_frequency,
+    )
+}
+
 /// Search many queries against a list of signatures.
 ///
 /// Note: this function loads all _queries_ into memory, and iterates over
@@ -21,6 +143,7 @@ pub fn multisearch(
     selection: Selection,
     allow_failed_sigpaths: bool,
     estimate_ani: bool,
+    estimate_prob_overlap: bool,
     output: Option<String>,
 ) -> Result<(), Box<dyn std::error::Error>> {
     // Load all queries into memory at once.
@@ -48,7 +171,7 @@ pub fn multisearch(
     let mut new_selection = selection;
     new_selection.set_scaled(expected_scaled);
 
-    let queries = query_collection.load_sketches(&new_selection)?;
+    let queries: Vec<SmallSignature> = query_collection.load_sketches(&new_selection)?;
 
     // Load all against sketches into memory at once.
     let against_collection = load_collection(
@@ -58,7 +181,25 @@ pub fn multisearch(
         allow_failed_sigpaths,
     )?;
 
-    let against = against_collection.load_sketches(&new_selection)?;
+    let againsts: Vec<SmallSignature> = against_collection.load_sketches(&new_selection)?;
+
+    let (
+        n_comparisons,
+        query_merged_frequencies,
+        against_merged_frequencies,
+        query_term_frequencies,
+        inverse_document_frequency,
+    ) = if estimate_prob_overlap {
+        compute_prob_overlap_stats(&queries, &againsts)
+    } else {
+        (
+            0.0,
+            Default::default(),
+            Default::default(),
+            Default::default(),
+            Default::default(),
+        )
+    };
 
     // set up a multi-producer, single-consumer channel.
     let (send, recv) =
@@ -75,7 +216,7 @@ pub fn multisearch(
 
     let processed_cmp = AtomicUsize::new(0);
 
-    let send = against
+    let send = againsts
         .par_iter()
         .filter_map(|against| {
             let mut results = vec![];
@@ -115,6 +256,30 @@ pub fn multisearch(
                     let mut match_containment_ani = None;
                     let mut average_containment_ani = None;
                     let mut max_containment_ani = None;
+                    let mut prob_overlap: Option<f64> = None;
+                    let mut prob_overlap_adjusted: Option<f64> = None;
+                    let mut containment_adjusted: Option<f64> = None;
+                    let mut containment_adjusted_log10: Option<f64> = None;
+                    let mut tf_idf_score: Option<f64> = None;
+
+                    // Compute probability overlap stats if requested
+                    if estimate_prob_overlap {
+                        let prob_stats = compute_single_prob_overlap(
+                            query,
+                            against,
+                            n_comparisons,
+                            &query_merged_frequencies,
+                            &against_merged_frequencies,
+                            &query_term_frequencies,
+                            &inverse_document_frequency,
+                            containment_query_in_target,
+                        );
+                        prob_overlap = Some(prob_stats.prob_overlap);
+                        prob_overlap_adjusted = Some(prob_stats.prob_overlap_adjusted);
+                        containment_adjusted = Some(prob_stats.containment_adjusted);
+                        containment_adjusted_log10 = Some(prob_stats.containment_adjusted_log10);
+                        tf_idf_score = Some(prob_stats.tf_idf_score);
+                    }
 
                     // estimate ANI values
                     if estimate_ani {
@@ -142,6 +307,11 @@ pub fn multisearch(
                         match_containment_ani,
                         average_containment_ani,
                         max_containment_ani,
+                        prob_overlap: prob_overlap,
+                        prob_overlap_adjusted: prob_overlap_adjusted,
+                        containment_adjusted: containment_adjusted,
+                        containment_adjusted_log10: containment_adjusted_log10,
+                        tf_idf_score: tf_idf_score,
                     })
                 }
             }

src/pairwise.rs

@@ -80,6 +80,12 @@ pub fn pairwise(
             let containment_q1_in_q2 = overlap / query1_size;
             let containment_q2_in_q1 = overlap / query2_size;
 
+            let prob_overlap = None;
+            let prob_overlap_adjusted = None;
+            let containment_adjusted = None;
+            let containment_adjusted_log10 = None;
+            let tf_idf_score = None;
+
             if containment_q1_in_q2 > threshold || containment_q2_in_q1 > threshold {
                 let max_containment = containment_q1_in_q2.max(containment_q2_in_q1);
                 let jaccard = overlap / (query1_size + query2_size - overlap);
@@ -113,6 +119,11 @@ pub fn pairwise(
                     match_containment_ani,
                     average_containment_ani,
                     max_containment_ani,
+                    prob_overlap,
+                    prob_overlap_adjusted,
+                    containment_adjusted,
+                    containment_adjusted_log10,
+                    tf_idf_score,
                 })
                 .unwrap();
             }
@@ -127,6 +138,11 @@ pub fn pairwise(
             let mut match_containment_ani = None;
             let mut average_containment_ani = None;
             let mut max_containment_ani = None;
+            let prob_overlap = None;
+            let prob_overlap_adjusted = None;
+            let containment_adjusted = None;
+            let containment_adjusted_log10 = None;
+            let tf_idf_score = None;
 
             if estimate_ani {
                 query_containment_ani = Some(1.0);
@@ -151,6 +167,11 @@ pub fn pairwise(
                 match_containment_ani,
                 average_containment_ani,
                 max_containment_ani,
+                prob_overlap,
+                prob_overlap_adjusted,
+                containment_adjusted,
+                containment_adjusted_log10,
+                tf_idf_score,
             })
             .unwrap();
         }

src/python/sourmash_plugin_branchwater/__init__.py

@@ -456,6 +456,12 @@ def __init__(self, p):
         p.add_argument(
             "-a", "--ani", action="store_true", help="estimate ANI from containment"
         )
+        p.add_argument(
+            "-p",
+            "--prob-significant-overlap",
+            action="store_true",
+            help="estimate probability of overlap for significance ranking of search results, of the specific query and match, given all queries and possible matches",
+        )
 
     def main(self, args):
         print_version()
@@ -468,6 +474,9 @@ def main(self, args):
         notify(
             f"searching all sketches in '{args.query_paths}' against '{args.against_paths}' using {num_threads} threads"
         )
+        notify(
+            f"estimate ani? {args.ani} / estimate probability of overlap? {args.prob_significant_overlap}"
+        )
 
         super().main(args)
         status = sourmash_plugin_branchwater.do_multisearch(
@@ -478,6 +487,7 @@ def main(self, args):
             args.scaled,
             args.moltype,
             args.ani,
+            args.prob_significant_overlap,
             args.output,
         )
         if status == 0: