This notebook is used to compare the number of non-zero coefficients in the elastic net models trained by EnPACT for different modalities and conditions. This gives some idea of how “complex” each modality is with respect to the Enformer predictions.
en_net_path <- "/beagle3/haky/users/saideep/projects/Con_EnPACT/models/Flu_ATAC_ws8/intermediates/train_enpact/trained_enpact_eln_Flu.linear.rds"
loaded_en_net <- readRDS(en_net_path)
coefs = predict(loaded_en_net, type="coef")
nonzero_coefs = coefs != 0
sum(nonzero_coefs)[1] 2749proj_path = "/beagle3/haky/users/saideep/projects/Con_EnPACT/models"
optimal_window_sizes = list(
"H3K27ac"=8,
"H3K27me3"=64,
"H3K4me1"=32,
"H3K4me3"=8,
"ATAC"=8,
"RNAseq"=4
)
modalities = c()
conditions = c()
window_sizes = c()
nonzero_count = c()
optws_modalities = c()
for (x in 1:length(optimal_window_sizes)) {
for (condition in c("Flu","NI")) {
for (ws in c(2,4,8,16,32,64,128)) {
modality = names(optimal_window_sizes)[x]
if (modality == "RNAseq") {
path_to_en_net = glue("{proj_path}/{condition}_ws{ws}_ns_3_7_lnormFALSE_beforeFALSE_scale1_sourcekallisto/intermediates/train_enpact/trained_enpact_eln_{condition}.linear.rds")
} else {
path_to_en_net = glue("{proj_path}/{condition}_{modality}_ws{ws}/intermediates/train_enpact/trained_enpact_eln_{condition}.linear.rds")
}
cur_en_net <- readRDS(path_to_en_net)
coefs = predict(cur_en_net, type="coef")
nonzero_coefs = coefs != 0
modalities = c(modalities, modality)
window_sizes = c(window_sizes, as.character(ws))
conditions = c(conditions, condition)
nonzero_count = c(nonzero_count, sum(nonzero_coefs))
optws_modalities = c(optws_modalities, glue(modality,"_", as.character(optimal_window_sizes[modality])))
}
}
}
en_summary_df <- data.frame(modalities, conditions, nonzero_count, window_sizes, optws_modalities)
en_summary_df$window_sizes = factor(en_summary_df$window_sizes, levels=c("2","4","8","16","32","64","128"))en_summary_df modalities conditions nonzero_count window_sizes optws_modalities
1 H3K27ac Flu 2669 2 H3K27ac_8
2 H3K27ac Flu 2401 4 H3K27ac_8
3 H3K27ac Flu 2230 8 H3K27ac_8
4 H3K27ac Flu 2211 16 H3K27ac_8
5 H3K27ac Flu 2213 32 H3K27ac_8
6 H3K27ac Flu 1312 64 H3K27ac_8
7 H3K27ac Flu 336 128 H3K27ac_8
8 H3K27ac NI 1974 2 H3K27ac_8
9 H3K27ac NI 2047 4 H3K27ac_8
10 H3K27ac NI 2032 8 H3K27ac_8
11 H3K27ac NI 1939 16 H3K27ac_8
12 H3K27ac NI 2056 32 H3K27ac_8
13 H3K27ac NI 1849 64 H3K27ac_8
14 H3K27ac NI 467 128 H3K27ac_8
15 H3K27me3 Flu 1522 2 H3K27me3_64
16 H3K27me3 Flu 1276 4 H3K27me3_64
17 H3K27me3 Flu 1414 8 H3K27me3_64
18 H3K27me3 Flu 1386 16 H3K27me3_64
19 H3K27me3 Flu 1477 32 H3K27me3_64
20 H3K27me3 Flu 1698 64 H3K27me3_64
21 H3K27me3 Flu 1384 128 H3K27me3_64
22 H3K27me3 NI 1464 2 H3K27me3_64
23 H3K27me3 NI 1261 4 H3K27me3_64
24 H3K27me3 NI 1383 8 H3K27me3_64
25 H3K27me3 NI 1333 16 H3K27me3_64
26 H3K27me3 NI 1421 32 H3K27me3_64
27 H3K27me3 NI 1679 64 H3K27me3_64
28 H3K27me3 NI 1397 128 H3K27me3_64
29 H3K4me1 Flu 3621 2 H3K4me1_32
30 H3K4me1 Flu 3214 4 H3K4me1_32
31 H3K4me1 Flu 2917 8 H3K4me1_32
32 H3K4me1 Flu 2715 16 H3K4me1_32
33 H3K4me1 Flu 2507 32 H3K4me1_32
34 H3K4me1 Flu 2223 64 H3K4me1_32
35 H3K4me1 Flu 1480 128 H3K4me1_32
36 H3K4me1 NI 3750 2 H3K4me1_32
37 H3K4me1 NI 3205 4 H3K4me1_32
38 H3K4me1 NI 2931 8 H3K4me1_32
39 H3K4me1 NI 2720 16 H3K4me1_32
40 H3K4me1 NI 2630 32 H3K4me1_32
41 H3K4me1 NI 2406 64 H3K4me1_32
42 H3K4me1 NI 1576 128 H3K4me1_32
43 H3K4me3 Flu 1422 2 H3K4me3_8
44 H3K4me3 Flu 1187 4 H3K4me3_8
45 H3K4me3 Flu 1232 8 H3K4me3_8
46 H3K4me3 Flu 1040 16 H3K4me3_8
47 H3K4me3 Flu 769 32 H3K4me3_8
48 H3K4me3 Flu 562 64 H3K4me3_8
49 H3K4me3 Flu 667 128 H3K4me3_8
50 H3K4me3 NI 1223 2 H3K4me3_8
51 H3K4me3 NI 1255 4 H3K4me3_8
52 H3K4me3 NI 1395 8 H3K4me3_8
53 H3K4me3 NI 1137 16 H3K4me3_8
54 H3K4me3 NI 1107 32 H3K4me3_8
55 H3K4me3 NI 1646 64 H3K4me3_8
56 H3K4me3 NI 656 128 H3K4me3_8
57 ATAC Flu 2674 2 ATAC_8
58 ATAC Flu 2575 4 ATAC_8
59 ATAC Flu 2749 8 ATAC_8
60 ATAC Flu 2168 16 ATAC_8
61 ATAC Flu 2054 32 ATAC_8
62 ATAC Flu 1234 64 ATAC_8
63 ATAC Flu 1330 128 ATAC_8
64 ATAC NI 3254 2 ATAC_8
65 ATAC NI 3374 4 ATAC_8
66 ATAC NI 2887 8 ATAC_8
67 ATAC NI 2666 16 ATAC_8
68 ATAC NI 2332 32 ATAC_8
69 ATAC NI 1714 64 ATAC_8
70 ATAC NI 1556 128 ATAC_8
71 RNAseq Flu 327 2 RNAseq_4
72 RNAseq Flu 311 4 RNAseq_4
73 RNAseq Flu 343 8 RNAseq_4
74 RNAseq Flu 319 16 RNAseq_4
75 RNAseq Flu 466 32 RNAseq_4
76 RNAseq Flu 419 64 RNAseq_4
77 RNAseq Flu 232 128 RNAseq_4
78 RNAseq NI 548 2 RNAseq_4
79 RNAseq NI 532 4 RNAseq_4
80 RNAseq NI 585 8 RNAseq_4
81 RNAseq NI 621 16 RNAseq_4
82 RNAseq NI 615 32 RNAseq_4
83 RNAseq NI 547 64 RNAseq_4
84 RNAseq NI 460 128 RNAseq_4ggplot(en_summary_df) + geom_bar(aes(x=modalities, y=nonzero_count, fill=conditions), stat="identity", position="dodge") + theme_minimal() + theme(axis.text.x = element_text(angle = 45, hjust = 1))
Interestingly, RNAseq seems the “least” complex by this metric.
ggplot(en_summary_df[en_summary_df$condition == "Flu",]) + geom_bar(aes(x=optws_modalities, y=nonzero_count, fill=window_sizes), stat="identity", position="dodge") + theme_minimal() + theme(axis.text.x = element_text(angle = 45, hjust = 1))