pinellolab/dna-diffusion

🧬 Generative modeling of regulatory DNA sequences with diffusion probabilistic models 💨

dna-diffusion's Issues

Implement train.py/sampler.py CLI argument parser

label and close inactive issues and PRs

See actions/stale.

merge "codebase" work into default branch

Please see the project information pane for information regarding the migration to distributed trunk-based collaboration.

Implement W&B logger

Create function motif_count Input

def motif_count Input(sequences, return_probability=False)

'sequences': list of sequences to scan
return_probability: if True scale the occurrences based on the total number of sequences in input
Return: pandas dataframe with index 'motif_id_name' value: 'occurences' (or 'probability') of motif in the sequences

Write unit tests

Implement score-interpolation

GOAL : Implements score-interpolation described in Continuous diffusion for categorical data

Please see Figure 3 below for overall implementation picture requirements.

Up-to-date notebook progress could be found at https://github.com/pinellolab/DNA-Diffusion/blob/score_interpolation_latest/notebooks/experiments/conditional_diffusion/dna_diff_baseline_conditional_UNET.ipynb , click on history button to check actual code commits modifications.

DONE :

1. Added DPM-Solver++ ODE solver block
2. Added Interpolate Embeddings block
3. Added reparameterization trick for p_losses()
4. Fed the output of Time Embedding stage into LayerNorm inside Unet model

TODO :

• Enable the self_conditioning logic properly without runtime dimension issue

• Verify the Interpolate Embeddings block

• Verify that the LearnedSinusoidalPosEmb class is actually following the theoretical concept behind Random Fourier Embedding block

• Enable the initial denoising inside p_losses() properly considering the actual mean and std at denoising timestep t=0

• Implement all aspects of Input embedding stage properly for the Interpolate scores block

• Enhance the denoising process with noise = sigma * epsilon where sigma is a learned NN parameter, such that the noise will be a deterministic function of the input and the parameter rather than being independent and randomly generated as suggested by chatGPT

• Verify the integration of Time Embedding stage into Unet model

• Enhance the training pipeline using modified version of tweedie formula in soft-diffusion paper

Create a Data Loader Class with Pytorch Lightning

For the first version make sure the latest version provided by @meuleman here:

https://www.meuleman.org/research/synthseqs/#material

Specifically, we have the following datasets available:

training set: 160k sequences, 10k per NMF component (chr3-chrY, .csv.gz)
validation set: 16k sequences, 1k per NMF component (chr2 only, .csv.gz)
test set: 16k sequences, 1k per NMF component (chr1 only) .csv.gz)

Each of these contains the genomic locations (human genome assembly hg38, first 3 columns) of accessible genome elements, their majority NMF component (column: component, see legend figure above) as well as their 200bp nucleotide sequence (column: raw_sequence).

Column Example Description
seqname chr16 Chromosome
start 68843660 Start position
end 68843880 End position
DHS_width 220 Width of original DHS
summit 68843790 Position of center of mass of DHS
total_signal 122.770678 Sum of DNase-seq signal across biosamples
numsamples 61 Number of biosamples with the DHS
raw_sequence GAGGCATTG… 200bp extracted nucleotide sequence
component 1 Dominant DHS Vocabulary component
proportion 0.767371514 Proportion of NMF loadings in this component

remove merged/unused branches

We would like to clean up branches that have been merged into the default branch while making sure we do not delete existing branches where work is ongoing. Please comment here with the name of the branch you are currently working on and a projected timeframe to merge.

update metadata files

Fix script pathing issues when training on cluster

• Once moved to distributed training on cluster, code from dna-diffusion branch outputs various relative path errors

Prototype HyperNetwork for conditioning on Luca's notebook

Prototype HyperNetwork for weights generation for our main denoising network.

Decide and implement evaluation metrics

Create a function called KL_divergence_motifs

def KL_divergence_motifs(original, generated)

'original': list or pandas Series with original sequences used to train the model
'generated':list or pandas Series with generated sequences used to train the model

return a 'score' (float) calculating the Kullback-Leibler divergence

add pre-commit.ci configuration

add pre-commit.ci configuration

Benchmarking against other tools (GAN paper, Transformer implementation)

migrate to src-full layout

See, for example,

• https://packaging.python.org/en/latest/discussions/src-layout-vs-flat-layout/
• https://hatch.pypa.io/latest/config/project-templates/#src-layout
• https://blog.ionelmc.ro/2014/05/25/python-packaging/#the-structure
• https://blog.ionelmc.ro/2014/05/25/python-packaging/#alternatives

What is a contribution? contribution file linked to read me

Discussed in #22

^{Originally posted by IhabBendidi October 17, 2022}
An idea is to make a contribution file, to keep account of contributors.
But what is a contribution?

A code push? that would not count the think tanks that theorized the project, so it is unfair. At the same time, what counts as contibuting discussion, and what isn't? Lines can be blurry in some edge cases.

What are your thoughts?

Contribution list is gonna become huge, and we need to keep proper track of it, so a file would be better, that would be linked to the readme

add code of conduct

See the contributor covenant code of conduct.

Implement gaussian log likelihood metric as in OpenAI's initial diffusion repository

Would be awesome to have the gaussian log likelihood metric to combine it with KL divergence as they do in https://github.com/openai/guided-diffusion.

Implement train.py and configs

Create training script and configs using pytorch lightning and hydra

pre-commit cannot parse config.py containing yaml

see pre-commit.ci run referring to src/config.py, which appears to contain a mixture of python and yaml

Refactor the DDPM and UNet code for conditional diffusion

Implement a training loop

Use reformers to greatly speed up training time and reduce memory usage

Transformers (like the ones used in your model) have a memory and time complexity of O(n^2). Reformers (explained here) have a memory and time complexity of O(n log n) and, as such, require far less ram and compute to train.

I am strongly suggesting you use reformers since they can replace transformers with almost no change and with a massive benefit. For reference, replacing transformers with reformers on one of my projects took the memory usage down from >10GB to <1GB and was able to train at 45 mins per epoch on an 8-core CPU (I was testing it out before migrating to colab)

Integrate Maximal Update Parametrization (MuP) for hyperparameter tuning

Implement a sequence quality metric based on k-mer composition

Roughly this is the idea: using a set of sequences (e.g. training set or generated set) we write first them in a temporary fasta file, using the obtained file (e.g. sequences.fa) call the kat hist function to get an histogram of kmers count. Here we can use k=7. Then we rescale the number count obtained of each k-mer based on the total sum to get probabilities of k-mers. Now assuming you have two probabilities vectors for k-mer corresponding to train sequences and generated sequences you can calculate the KL-diverge using the pytorch implementation that Cesar is also using. See: #14

Implement a sequence quality metric based on existing neural networks that can predict enhancer activity (Bert based), expression (Enformer) or chromatin accessibility (BPnet)

@sg134 Can you help to explore sequences -> enhancer models?
We can use these enhancer classifiers as orthogonal metrics to evaluate our synthetic sequences.
Probably we can find tons of these classifiers, but here is someplace to start your search:

bert enhancer
DeepEnhancer

What should be the focus:

• Sequence length?
• Do we need to retrain it? Can we retrain it on our 16 cell types?
• Create a function to receive a sequence and report the probability of it being classified as an enhancer.

Integrate PyTorch Ligthning for distributed training

Develop a standardized PR ticket template

My interaction over at #50 made me think:

We may or may not have previously talked about setting a template for opening new pull requests. The former interaction has a call for:

1. A link to the ticket it's addressing
2. A small explanation so as to why (based on what though?).

Maybe these two could be good starting points for a default ticket template? A la, copy and paste then fill in with your contribution. Additionally maybe, add if your contribution is going into the codebase/experiments/other folder (or where in the codebase, once it grows), although this is probably filled by 1 already. Anything else you guys consider relevant or good/standard practice?

Add p2 weighting to the conditioned DDPM

Consider DNA embedding models to transition to use stable diffusion

Create a CLIP-like embedding+tokenizer for our usecase that can be hooked up to stable diffusion pipeline

use hydra-zen for configuration

See, for example, the hydra-zen lightning tutorial.

Implement DDIM sampler based on Lucas' conditioned DDPM

Replicate the DDPM logic but with the mathematical tricks from the DDIM paper: https://arxiv.org/abs/2010.02502.

Design config file

Design a template config file which we can use to decide the what parameters the train and sampler parsers have to receive and pass to pl.Trainer and what the Diffusion and UNet classes final parameters are.

Prototype HyperNetwork for conditioning on Luca's notebook

Training a VQ-VAE for DNA-sequences for stable diffusion

Current notebook here: https://github.com/pinellolab/DNA-Diffusion/blob/latent-space-representation/vq_vae_diffusion.ipynb

Implement the guided diffusion based on new model

Please use @LucasSilvaFerreira this notebook as starting point: https://github.com/pinellolab/DNA-Diffusion/blob/dna-diffusion/models/vanilla_diffusion/Code_to_refactor_UNET_ANNOTATED%20v2.ipynb

Implement unit tests

Implement a scalable data loading pipeline

Separate sampling metrics from validation metrics

• Validation metrics currently exist in both validation_metrics.py and sampling_metrics.py

add workflow to label issues

Create tests and github actions

• Add tests for all code functions
• Add github actions, for continuous testing of the code

Refactor dataloader code to include conditioning

What is your timeline?

3b base pairs, etc

add build system

Implement DDIM sampler based on conditioned DDPM

Implement a sequence quality metric based on k-mer composition

document the procedure to regenerate `model4cells_train_split_3_50_dims.pkl`

In #101 a model file was introduced.
We should document how to regenerate it and remove it from the repository.

Split single script into subscripts / begin populating dna-diffusion branch

Refactoring and splitting single script obtained from dna-diffusion/notebooks/experiments