AI_AFM

Repository for implementing AI in Atomic Force Microscopy (AFM)

Sample plots

The figures above shows the sample plots for (a) no rupture, (b) single rupture, (c) double rupture and (d) multiple rupture.

Classification Labels

Labels	3-class	2-class (single vs rest)	2-class (no vs rest)
0	No rupture and multiple rupture	Rest	No rupture
1	Single rupture	Single rupture	Rest
2	Double rupture rupture	-	-

Algorithm Overview

The figure above shows our few-shot framework. Training, shown in the top block, is done by inputting three samples from the train set into a triplet loss architecture. Testing, shown in the bottom block, is done by inputting train-test sample pairs into the model to compare embedding distances.

Results

Accuracy, class accuracies, precision, recall, and F-1 score for our few-shot approach, SVM, RF, and KNN are shown above for the 3-class case.

Accuracy and class accuracies are shown above for our few-shot approach on 2-class scenarios.

Loss Functions

Assuming Y = binary class label where similar class label = 1, dissimilar class label = 0, m = margin (default = 1.0) and let distance $D_{1,2} = (embedding_1 - embedding_2)^2$,

Simplified Contrastive Loss (ContrastiveLoss_ori):

$$L = \left( YD \right) + \left( 1-Y \right) \left[ max \left( 0, m-D \right)\right]$$

Modified Contrastive Loss (ContrastiveLoss):

$$L = 0.5 \left[ \left( YD \right) + \left( 1-Y \right) max \left(0, m-\sqrt{D} \right)^2 \right]$$

Triplet Loss (TripletLoss)

$$L = max \left( 0, D_{(+ve, anchor)} - D_{(-ve, anchor)} + m \right)$$

Install

Clone repo and install requirements.txt in a Python>=3.9.12 enviornment.

git clone https://github.com/JoshuaRWaite/AI_AFM  # clone
cd AI_AFM
pip install -r requirements.txt  # install

Running few-shot experiments

Example run for 3 class (No/Multipe Rupture, Single Rupture, Double Rupture) case:

python main.py -d 3class_matching -m convo1D2 -mod 10 -ep 300 -bs 32 -opt sgd --lr 0.01 -sche_step 1 -sche_gamma 1 -mt triplet -s 0 -exp 4

Example run for 2 class (Single Rupture vs Rest) case:

python main.py -d 2class_s_vs_r -m convo1DS2 -mod 10 -ep 300 -bs 32 -opt sgd --lr 0.01 -sche_step 1 -sche_gamma 1 -mt triplet -g 0 -s 0 -up 0 -exp 5

Example run for 2 class (No Rupture vs Rest) case:

python main.py -d 2class_n_vs_r -m convo1DDrp2 -mod 10 -ep 300 -bs 16 -opt sgd --lr 0.01 -sche_step 1 -sche_gamma 1 -mt triplet -g 0 -s 0 -up 0 -exp 6

Running few-shot evaluation with pretrained weights

Example eval for 3 class (No/Multipe Rupture, Single Rupture, Double Rupture) case:

python eval.py -d 3class_matching -m convo1D2 -mod 10 -ep 300 -bs 32 -opt sgd --lr 0.01 -sche_step 1 -sche_gamma 1 -mt triplet -s 0 -exp 1 -tar 28-06-2022_220804_best

Example eval for 2 class (Single Rupture vs Rest) case:

python eval.py -d 2class_s_vs_r -m convo1DS2 -mod 10 -ep 300 -bs 32 -opt sgd --lr 0.01 -sche_step 1 -sche_gamma 1 -mt triplet -g 0 -s 0 -up 0 -exp 2 -tar 28-06-2022_163140_best

Example eval for 2 class (No Rupture vs Rest) case:

python eval.py -d 2class_n_vs_r -m convo1DDrp2 -mod 10 -ep 300 -bs 16 -opt sgd --lr 0.01 -sche_step 1 -sche_gamma 1 -mt triplet -g 0 -s 0 -up 0 -exp 3 -tar 28-06-2022_163633_best

List of Optimizer (-opt argument)

SGD (sgd)
ADAM (adam)

List of Loss Functions (-mt argument)

Contrastive (siam)
Triplet (triplet)

List of Models (-m argument)

Linear (toy, toyL, toyS, toyS2, toyS3, toyXS, cerb, cerbL, cerbXL)
Convolutional (convo1D, convo1DS, convo1DDrp, convo1DDrp2, convo1DS2, convo1D2, convo1D3)

List of Datasets (-d argument)

No/multiple rupture, single rupture, and double rupture (3class_matching)
Single rupture vs rest (2class_s_vs_r)
No rupture vs rest (2class_n_vs_r)

Running shallow experiments

Example run for shallow methods with SVM and varying percentage of noisy data for the 3 class case:

python re_shallow.py -mode noise -algo SVM -d 3class_matching

-mode has noise and trainsz
-algo has KNN, RF, and SVM

Acknowledgements

We would like to acknowledge Prof. Peter Hoffmann (Department of Physics, Wayne State University) and Prof. Rafael Fridman (Department of Pathology, Wayne State University) for their help in culturing cells with overexpressed receptor and performing AFM force-distance measurements. This work was supported by Iowa State College of Engineering Exploratory Research Program (A.S., S.S.).

Citation

Please cite our paper in your publications if it helps your research:

@article{waite_tan_saha_sarkar_sarkar_2023,
  title={Few-shot deep learning for AFM Force curve characterization of single-molecule interactions},
  author={Waite, Joshua R. and Tan, Sin Yong and Saha, Homagni and Sarkar, Soumik and Sarkar, Anwesha},
  journal={Patterns},
  year={2023}
}

Paper Links

Few-shot deep learning for AFM force curve characterization of single-molecule interactions

ZhiyuanChen/AI_AFM