Maitra, C., Seal, D.B., Das, V. and De, R.K., 2022. UMINT: unsupervised neural network for single cell multi-omics integration. BioRxiv, pp.2022-04. https://www.biorxiv.org/content/10.1101/2022.04.21.489041v1

Now published at Frontiers in Molecular Biosciences

Maitra, C., Seal, D.B., Das, V. and De, R.K., Unsupervised neural network for single cell Multi-omics INTegration (UMINT): An application to health and disease. Frontiers in Molecular Biosciences, 10, p.335. https://www.frontiersin.org/articles/10.3389/fmolb.2023.1184748/full

To run UMINT, import umint.py from the Proposed directory and run the function CombinedEncoder. All the parameters are mentioned below for better understanding. One can also follow a .ipynb file from the Proposed directory.

To run umint, one needs to install tensorflow, sklearn, scipy and pandas packages. Installation codes are as follows:

• pip install tensorflow
• pip install scikit-learn
• pip install scipy
• pip install pandas

All input parameters are as follows: layer_neuron, mid_neuron, seed, lambda_act, lambda_weight, epoch, bs

• data: List of input data matrices for training. [The input data matrices should be in the form of cells x features.]
• val: List of data matrices for validation. [The validation data matrices also should be in the same format as input data matrices i.e., cells x features.]
• layer_neuron: List of neurons in the modality encoding layer [modality wise].
• mid_neuron: Dimension onto which the data is being projected.
• seed: To reproduce the results set a seed value.
• lambda_act: Activity regularizer parameter.
• lambda_weight: Kernel regularizer parameter.
• epoch: Total number of iteration for training.
• bs: Training batch size.

To run UMINT, one needs to import the script umint (within the Proposed directory) first. An example is provided below. Let x1_train [cells x features] and x2_train [cells x features] be two training datasets, coming from two different omics modalities, and x1_test [cells x features], x2_test [cells x features] be their respective counterparts for validation.

import umint
MyEncoder, MyAE = umint.CombinedEncoder(data=[x1_train, x2_train], val=[x1_test, x2_test],
                                        layer_neuron=[128, 10], mid_neuron=64, seed=98,
                                        lambda_act=0.0001, lambda_weight=0.001, epoch=25, bs=16)

Once UMINT is trained, to find the latent lower dimensional embedding produced by UMINT, run the code below.

low = MyEncoder.predict([x1, x2]) 

To integrate multiple modalities please change the input accordingly. The sizes of data, val and layer_neuron must match in order to run umint.py successfully.

Chayan Maitra

Machine Intelligence Unit, Indian Statistical Institute, 203 Barrackpore Trunk Road, Kolkata 700108, India

E-mail: [email protected]

Dibyendu B. Seal

Tatras Data Services Pvt. Ltd., E64, Vasant Marg, Vasant Vihar, New Delhi 110057, India.

E-mail: [email protected]

Vivek Das

Novo Nordisk A/S, Novo Nordisk Park 1, 2760 M ̊aløv, Denmark

E-mail: [email protected]

Rajat K. De

Machine Intelligence Unit, Indian Statistical Institute, 203 Barrackpore Trunk Road, Kolkata 700108, India

E-mail: [email protected]

https://doi.org/10.5281/zenodo.7723340