proseg 3.1.1

Probabilistic cell segmentation for in situ spatial transcriptomics
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245

use num::{Signed, Zero, cast::AsPrimitive};
use rayon::prelude::*;
use std::f32;

use super::csrmat::CSRMat;

// Bookkeeping for P² algorithm
#[derive(Clone, Copy)]
struct P2Values {
    h: [f32; 5], // estimates
    n: [f32; 5], // positions
}

impl P2Values {
    fn new() -> P2Values {
        P2Values {
            h: [f32::INFINITY; 5],
            n: [1.0, 2.0, 3.0, 4.0, 5.0],
        }
    }

    fn update(&mut self, t: usize, dn: &[f32; 5], obs: f32) {
        // if we are still on the first 5 observations, just initialize
        if t < 5 {
            self.h[t] = obs;

            // slide the observation left until we find its place
            let mut j = t;
            while j > 0 {
                if self.h[j] < self.h[j - 1] {
                    self.h.swap(j, j - 1);
                    j -= 1;
                } else {
                    break;
                }
            }

            return;
        }

        let k = if obs < self.h[0] {
            self.h[0] = obs;
            1
        } else if obs < self.h[1] {
            1
        } else if obs < self.h[2] {
            2
        } else if obs < self.h[3] {
            3
        } else if obs <= self.h[4] {
            4
        } else {
            self.h[4] = obs;
            4
        };

        // increment position markers
        for n_i in &mut self.n[k..5] {
            *n_i += 1.0;
        }

        // adjust marker heights
        #[allow(clippy::needless_range_loop)]
        for i in 1..4 {
            let delta = dn[i] - self.n[i];
            if (delta >= 1.0 && self.n[i + 1] - self.n[i] > 1.0)
                || (delta <= -1.0 && self.n[i - 1] - self.n[i] < -1.0)
            {
                let d = delta.signum();

                // parabolic predictor
                let h_adj = self.h[i]
                    + (d / (self.n[i + 1] - self.n[i - 1]))
                        * ((self.n[i] - self.n[i - 1] + d) * (self.h[i + 1] - self.h[i])
                            / (self.n[i + 1] - self.n[i])
                            + (self.n[i + 1] - self.n[i] - d) * (self.h[i] - self.h[i - 1])
                                / (self.n[i] - self.n[i - 1]));

                if self.h[i - 1] < h_adj && h_adj < self.h[i + 1] {
                    self.h[i] = h_adj;
                } else {
                    // linear prediction
                    if delta.is_positive() {
                        self.h[i] += (self.h[i + 1] - self.h[i]) / (self.n[i + 1] - self.n[i]);
                    } else {
                        self.h[i] -= (self.h[i] - self.h[i - 1]) / (self.n[i] - self.n[i - 1]);
                    }
                }

                if delta.is_positive() {
                    self.n[i] += 1.0;
                } else {
                    self.n[i] -= 1.0;
                }
            }
        }
    }

    fn estimate(&self) -> f32 {
        self.h[2]
    }
}

pub struct ScalarQuantileEstimator {
    estimates: P2Values,
    dn: [f32; 5],
    quantile: f32,
    t: usize,
}

impl ScalarQuantileEstimator {
    pub fn new(quantile: f32) -> Self {
        ScalarQuantileEstimator {
            estimates: P2Values::new(),
            dn: [
                1.0,
                1.0 + 2.0 * quantile,
                1.0 + 4.0 * quantile,
                3.0 + 2.0 * quantile,
                5.0,
            ],
            quantile,
            t: 0,
        }
    }

    pub fn update(&mut self, value: f32) {
        // increment desired positions
        self.dn[1] += self.quantile / 2.0;
        self.dn[2] += self.quantile;
        self.dn[3] += (1.0 + self.quantile) / 2.0;
        self.dn[4] += 1.0;
        self.estimates.update(self.t, &self.dn, value);
        self.t += 1;
    }

    pub fn estimate(&self) -> f32 {
        self.estimates.estimate()
    }
}

/*
// Estimating transcript count quantiles using the P² algorithm
pub struct CountQuantileEstimator {
    estimates: SparseMat<P2Values, u32>,
    dn: [f32; 5], // "desired" positions
    quantile: f32,
    t: usize, // iteration number
}

impl CountQuantileEstimator {
    pub fn new(m: usize, n: usize, quantile: f32, shardsize: usize) -> Self {
        CountQuantileEstimator {
            estimates: SparseMat::zeros(m, n as u32 - 1, shardsize),
            dn: [
                1.0,
                1.0 + 2.0 * quantile,
                1.0 + 4.0 * quantile,
                3.0 + 2.0 * quantile,
                5.0,
            ],
            quantile,
            t: 0,
        }
    }

    pub fn update<T>(&mut self, counts: &SparseMat<T, u32>)
    where
        T: AsPrimitive<f32> + Sync + Send + Zero,
    {
        assert_eq!(self.estimates.m, counts.m);
        assert_eq!(self.estimates.n, counts.n);

        // increment desired positions
        self.dn[1] += self.quantile / 2.0;
        self.dn[2] += self.quantile;
        self.dn[3] += (1.0 + self.quantile) / 2.0;
        self.dn[4] += 1.0;

        self.estimates
            .par_rows()
            .zip(counts.par_rows())
            .for_each(|(estimates_c, counts_c)| {
                // update estimates with estimates entries with latest counts
                let counts_c_lock = counts_c.read();
                let mut estimates_c_lock = estimates_c.write();
                for (gene, estimates_cg) in estimates_c_lock.iter_nonzeros_mut() {
                    let count_cg = counts_c_lock.get(gene);
                    estimates_cg.update(self.t, &self.dn, count_cg.as_());
                }

                // update estimates where there is no entry
                for (gene, count_cg) in counts_c_lock.iter_nonzeros() {
                    estimates_c_lock.update(gene, P2Values::new, |est| {
                        est.update(self.t, &self.dn, count_cg.as_())
                    });
                }
            });

        self.t += 1;
    }
}
*/

pub struct CountMeanEstimator {
    pub estimates: CSRMat<u32, f32>,
    t: usize, // iteration number
}

impl CountMeanEstimator {
    pub fn new(m: usize, n: usize, _shardsize: usize) -> Self {
        CountMeanEstimator {
            estimates: CSRMat::zeros(m, n as u32 - 1),
            t: 0,
        }
    }

    pub fn update<T>(&mut self, counts: &CSRMat<u32, T>)
    where
        T: AsPrimitive<f32> + Sync + Send + Zero + PartialEq,
    {
        self.t += 1;
        let t = self.t as f32;
        let scale_factor = (t - 1.0) / t;

        self.estimates
            .par_rows()
            .zip(counts.par_rows())
            .for_each(|(estimates_c, counts_c)| {
                let counts_c_lock = counts_c.read();
                let mut estimates_c_lock = estimates_c.write();

                // E_t = E_{t-1} * (t-1)/t + X_t / t
                estimates_c_lock.guard.scale_all(|est| *est *= scale_factor);

                for (gene, count_cg) in counts_c_lock.iter_nonzeros() {
                    estimates_c_lock.update(
                        gene,
                        || 0.0,
                        |est| *est += count_cg.as_() / t,
                    );
                }
            });
    }
}