causal_cell_embedding

To use the spec file to create an identical conda environment on your machine: ''' conda create --name causalcell --file spec-file.txt '''

TODO

Add a diagram to the paper Read papers and add summaries to the git repo (as md files) in the folder Papers Investigate datasets Keep working on the ideas : Define different experiments we could make

Datasets

• Mambo - a framework to synthesize data from various data sources in order to construct and represent multimodal networks Link.

• Stanford Biomedical Network Dataset Collection Link

Explore the L1000 API

https://github.com/dhimmel/lincs/tree/abcb12f942f93e3ee839e5e3593f930df2c56845

Code a synthetic data generator

Remarks :

• We start with categorical latent variables (binary at first)
• Should we add noise in the generative process of latent variables + genes?

objects :

Gene_expresion_generator : initialized with number of latent variables and number of genes (among other things) class attributes :

• a function object F wich will be used in the generation process (take linear function at first)
• a directed graph objet representing latent DAG + parameters of F for each non-root node in the graph
• a binary matrix of size (number of latent variables, number of genes) which encodes which edges exist

methods for initialization :

build_latent _dag() input : number of latent variables, some parameters to define how we generate edges of the DAG (define some probability distribution over edges). The graph should not contain any cycle ! categorical case : each conditional probabilty P(l | PA(l)) will be generated from a family of functions that we should define (linear, NN etc) continuous case : we define the functions that will output the value of latent variables directly. Add noise as input output : a directed graph objet (networkx) whose nodes are the latent variables

build latent_gene_connexions() input : number of latent variables, number of genes, some distribution over edges Output : a binary matrix of size (number of latent variables, number of genes) which encodes which edges exist

build_env() TODO input : number of environments output : matrix

methods for sampling :

sample(env) TODO method which is used to get samples from the generator

Intersting links

• pgmpy : https://github.com/pgmpy
• python library for single cell data analysis https://scanpy.readthedocs.io/en/stable/
• task scheduler https://trello.com/en
• analysis toolkit for LINCS L1000 https://github.com/dhimmel/lincs/tree/abcb12f942f93e3ee839e5e3593f930df2c56845
• toolkit to analyse dynamics of scRNAseq https://buildmedia.readthedocs.org/media/pdf/scvelo/latest/scvelo.pdf
• collaboration ? https://mhi-omics.org/people/julie-hussin/
• Fabian Theis https://www.helmholtz-muenchen.de/icb/institute/staff/staff/ma/2494/Prof.%20Dr.%20Dr.-Theis/index.html
• Course ML for healthcare https://mlhc19mit.github.io/
• jupyter notebooks for RNAseq https://amp.pharm.mssm.edu/biojupies/
• pytorch metalearning library https://github.com/tristandeleu/pytorch-meta
• Depmap https://depmap.org/portal/achilles/
• Genome Research International journal that publishes studies on genomes.
• How to analyze any single-cell RNA-seq dataset without a computer Paper Demo

People to talk to for this project

• Fabian Theis and his student David Fischer
• David Gifford
• Alan Aspuru Guzik
• Dana Pe'er
• Christina Leslie