4. Important microbial marker identification¶
In this section, we are going to calculate the global feature importance (GFI) to find out the key microbes that contributes to the CRC detection model. The simply-explainer is used to calculate the GFI, more about the AggMapNet model exaplaination can be found here. By calculating the importance score for each microbes, we can draw the saliency-map to find out the hot zone in the 2D-microbiomeprints.
Saliency-Map is an image in which the brightness of a pixel represents how salient the pixel is i.e brightness of a pixel is directly proportional to its saliency. It is generally a grayscale image. Saliency maps are also called as a heat map where hotness refers to those regions of the image which have a big impact on predicting the class which the object belongs to.
The purpose of the saliency-map is to find the regions which are prominent or noticeable at every location in the visual field and to guide the selection of attended locations, based on the spatial distribution of saliency.
from matplotlib.ticker import FormatStrFormatter
import matplotlib.pyplot as plt
import matplotlib.patches as patches
import seaborn as sns
import pandas as pd
import numpy as np
import os
sns.set(style='white', font='sans-serif', font_scale=2)
from aggmap import loadmap, AggMapNet
from aggmap.AggMapNet import load_model
4.1 Calculate the global feature importance¶
Let's calculate the GFI first. We need to reload the trained model that dumped in the disk. After that, we will use the simply_explainer to calculate GFI based on the training set data (i.e., the one country data used to train the model).
4.1.1 GFI for model trained on overall MEGMA Fmaps¶
The global feature importance is calculated based on the training set. Once we got the GFI values, The megma object can help us to reshape the GFI vector to the saliency map.
countries = ['AUS', 'CHN', 'DEU', 'FRA', 'USA']
model_dir = './megma_overall_model'
# load the pre-fitted megma_all object
megma = loadmap('./megma/megma.all')
gfis = []
for country in countries:
clf = load_model(os.path.join(model_dir, 'aggmapnet.%s' % country), gpuid=0)
sxp = AggMapNet.simply_explainer(clf, megma,
backgroud = 'global_min',
apply_smoothing = True)
gfi = sxp.global_explain()
gfis.append(gfi.simply_importance_class_0.to_frame(name = country))
dfimp1 = pd.concat(gfis, axis=1)
4.1.2 GFI for model trained on country specific MEGMA Fmaps¶
model_dir = './megma_country_model'
gfis2 = []
reshape_indexes = {}
for country in countries:
#load the country-specific megma
megma = loadmap('./megma/megma.%s' % country)
clf = load_model(os.path.join(model_dir, 'aggmapnet.%s' % country), gpuid=0)
sxp = AggMapNet.simply_explainer(clf, megma,
backgroud = 'global_min',
apply_smoothing = True)
gfi2 = sxp.global_explain(clf.X_, clf.y_)
gfis2.append(gfi2.simply_importance_class_0.to_frame(name = country))
## megma is different, therefore the reshape index and fmap_shape is also different
reshape_index = megma.feature_names_reshape
reshape_indexes.update({country: (megma.fmap_shape, reshape_index)})
dfimp2 = pd.concat(gfis2, axis=1)
4.2 Generate the explaination saliency map¶
4.2.1 Saliency map for overall MEGMA Fmaps¶
megma = loadmap('./megma/megma.all')
fig, axes = plt.subplots(1, 5, figsize=(24, 4))
for country, ax in zip(countries, axes):
IMPM = dfimp1[country].values.reshape(*megma.fmap_shape)
#print(IMPM.max().round(1))
sns.heatmap(IMPM,
cmap = 'rainbow', alpha = 0.8, ax =ax, xticklabels=0, yticklabels=0,
vmin = -0.1, vmax = 3.5,
cbar_kws = {'fraction':0.046, 'shrink':0.9, 'aspect': 40, 'pad':0.02, })
ax.set_title(country)
cbar = ax.collections[0].colorbar
cbar.ax.set_title('FI')
cbar.ax.yaxis.set_major_formatter(FormatStrFormatter("%.0f"))
plt.subplots_adjust(wspace = 0.18)
4.2.2 Saliency map country specific MEGMA Fmaps¶
fig, axes = plt.subplots(1, 5, figsize=(24, 4))
for country, ax in zip(countries, axes):
fmap_shape, reshape_idx = reshape_indexes[country]
IMPM = dfimp2.loc[reshape_idx][country].values.reshape(*fmap_shape)
#print(IMPM.max().round(1))
sns.heatmap(IMPM,
cmap = 'rainbow', alpha = 0.8, ax =ax, xticklabels=0, yticklabels=0,
vmin = -0.1, vmax = 3.5,
cbar_kws = {'fraction':0.046, 'shrink':0.9, 'aspect': 40, 'pad':0.02, })
bottom, top = ax.get_ylim()
#ax.set_ylim(bottom + 0.5, top - 0.5)
ax.set_title(country)
cbar = ax.collections[0].colorbar
cbar.ax.set_title('FI')
cbar.ax.yaxis.set_major_formatter(FormatStrFormatter("%.0f"))
plt.subplots_adjust(wspace = 0.18)
4.3 Global feature importance correlation¶
lets calculate the correlations between the global feature importance (GFI) scores in different countries.
fig, axes = plt.subplots(ncols = 2, figsize=(14,5))
ax1, ax2 = axes
sns.heatmap(dfimp1.corr(), annot=True, fmt='.2f', linewidths=.3, ax = ax1,
annot_kws={"size": 16}, square=True,robust=True, vmin = 0, vmax = 0.9,
cmap='Greens')
sns.heatmap(dfimp2.corr(), annot=True, fmt='.2f', linewidths=.3, ax = ax2,
annot_kws={"size": 16}, square=True,robust=True, vmin = 0, vmax = 0.9,
cmap='Greens')
ax1.set_title("Overall MEGMA")
ax2.set_title("Country MEGMA")
plt.tight_layout()
fig.savefig('./images/GFI_megma_all_vs_megma_country.png', bbox_inches='tight', dpi=400)
dfimp1.to_csv('./images/STST_GFI_megma_all.csv')
dfimp2.to_csv('./images/STST_GFI_megma_specific.csv')
Let's show the important microbes in the hot zones
dfimp1.mean(axis=1).sort_values(ascending=False).head(20)
S1 = dfimp2.AUS.sort_values(ascending=False).head(20).index.to_list()
S2 = dfimp2.CHN.sort_values(ascending=False).head(20).index.to_list()
S3 = dfimp2.DEU.sort_values(ascending=False).head(20).index.to_list()
S4 = dfimp2.FRA.sort_values(ascending=False).head(20).index.to_list()
S5 = dfimp2.USA.sort_values(ascending=False).head(20).index.to_list()
set(S2) & set(S3) & set(S4)
4.4 Discussions and conclusions on saliency map¶
We can see that there are some notable regions on the saliency map, i.e., the hot zones. The microbes in these zones are important microbes. For the 5 detection models built from megma_all transformed Fmaps, all of the important zones is in the left upper conner (as shown in Fig.1), and their GFI scores are highly-correlated to each other.
Fig. 1: Fmaps generated by overall megma and the saliency maps.
For the 5 detection models built from country-specific megma transformed Fmaps, the important zones located in different areas(as shown in Fig.2), but the GFI values are highly-correlated among the four countries of CHN, FRA, DEU, and USA.
Fig. 2: Fmaps generated by country-specific megma and the saliency maps.
By fitting megma with different data, the 2D arrangement of microorganisms is different. Therefore, the hot zones for the country-specific megma is in different regions. Although the hot regions are different, the microbes in these regions may be in high consistant (for examples, the species of Fusobacterium nucleatum s. vincentii [ref_mOTU_v2_0754], Parvimonas sp. [ref_mOTU_v2_4961], and unknown Dialister [meta_mOTU_v2_5867] are located in these zones.).
We can see that the performance of the models trained based on these Fmaps is also different. The performance of the model is increased if the key microbes are clustered together, so that the hotspots can also be seen in the saliency map. For example, the STST performance for the country-specific megma_AUS is worst among the 5 countries, and the hot zone and GFI calculated also shows not so consistant with the rest countries.
In conclusion, the overall megma (fitted by all metagenomic abundance data) shows better performance than country-specific megma (fitted by one country metagenomic abundance data), suggesting that unsupervised megma should also be fitted on larger samples. The MEGMA-AggMapNet-Explainer pipeline can be used to identify important microbes.