of 11
Cell Systems, Volume
15
Supplemental information
Accurate single-molecule spot detection
for image-based spatial transcriptomics
with weakly supervised deep learning
Emily Laubscher, Xuefei Wang (Julie), Nitzan Razin, Tom Dougherty, Rosalind J.
Xu, Lincoln Ombelets, Edward Pao, William Graf, Jeffrey R. Mof
fi
tt, Yisong
Yue, and David Van Valen
1 Supplementary Information
1.1 Supplementary Note 1: Evaluation of generative model performance for the creation of
training data
Supplementary Figure 1:
Benchmarking consensus annotation output of the generative model.
(a) Error
distribution for EM estimates of TPR and FPR values for 100 trials with three simulated classical methods. (b)
Fraction of simulated detections correctly classified with increasing dataset size (number of spots in the dataset).
(c) Fraction of simulated detections correctly classified as a true or false detection by EM for an increasing number
of classical spot detection methods used in the EM method.
1
1.2 Supplementary Note 2: Generalization of Polaris’ spot detection model to a variety of
spot images
Supplementary Figure 2:
Polaris’ spot detection model generalizes to spot images generated with a
variety of single-molecule assays.
The spot probability prediction images encode the pixel-wise spot proba-
bility. The regression image is the sum of the square of the subpixel distances to the nearest spot in the x- and
y-dimensions. Pixels beyond a threshold value are set to zero. These outputs are used together to generate a set
of predicted spot locations with subpixel resolution, plotted over the raw image.
2
1.3 Supplementary Note 3: Mutual nearest-neighbor method for matching sets of spots
Supplementary Figure 3:
Example cases handled by a mutual nearest-neighbor matching algorithm.
(a) Example with spots inside and outside the threshold distance to a ground-truth spot. Ground truth spots
and their threshold distance are shown in grey. True positive detections are shown in green and false positive
detections are shown in orange. (b) Example with two spots inside the threshold distance to a ground-truth spot.
(c) Example with two spots within the threshold distance of two ground-truth spots.
3
1.4 Supplementary Note 4: Inter-algorithm agreement of spot detection results
Supplementary Figure 4:
Quantification of agreement between Polaris’ deep learning model and dif-
ferent classical spot detection methods.
The benchmarked methods include maximum-intensity filtering
(PLM), the Crocker-Grier centroid-finding algorithm (Trackpy), Laplacian of Gaussian (LoG), difference of Gaus-
sians (DoG), Airlocalize, and Polaris.
4
1.5 Supplementary Note 5: Benchmarking the receptive field of Polaris’ spot detection
model
Supplementary Figure 5:
Benchmarking the receptive field parameter of Polaris’ spot detection model.
(a-d) Violin plot quantifying the performance metrics ((a) precision, (b) recall, (c) F1, (d) best validation loss
during training) for models trained with different values for receptive field of Polaris’ spot detection model. n=24
trained models per receptive field condition. Plots illustrate the minimum, mean, and maximum values for each
experiment.
5
1.6 Supplementary Note 6: Benchmarking Polaris’ spot detection model on simulated im-
ages with ranging spot characteristics
Supplementary Figure 6:
Benchmarking model performance on simulated spot images with a range
of spot intensities and densities.
(a) Performance quantification for models with Polaris’ deep learning
architecture trained with various datasets predicting on images with a range of spot density. (b) Performance
quantification for models with Polaris’ deep learning architecture trained with various datasets predicting on
images with a range of levels of simulated noise. The low noise condition corresponds to a signal-to-noise ratio
of greater than approximately 16. The medium noise condition corresponds to a SNR of approximately 8-15.
The high noise condition corresponds to a SNR of approximately 3-7. (c) Performance quantification for various
spot detection methods on images with a range of spot densities. (d) Performance quantification for various spot
detection methods on images with a range of levels of simulated noise. The SNR ranges for each noise condition
are the same as those in (b).
6