CleaveNet is an end-to-end AI pipeline for the design of protease substrates. Applied to matrix metalloproteinases, CleaveNet enhances the scale, tunability, and efficiency of substrate design. CleaveNet generates peptide substrates that exhibit sound biophysical properties and capture not only well-established but also novel cleavage motifs. To enable precise control over substrate design, CleaveNet incorporates a conditioning tag that enables generation of peptides guided by a target cleavage profile, enabling targeted design of efficient and selective substrates. CleaveNet-generated substrates were validated through a large-scale in vitro screen, demonstrating success even in the challenging task of achieving high selectivity for MMP13. We envision that CleaveNet will accelerate our ability to study and capitalize on protease activity, paving the way for new in silico design tools across enzyme classes.
Below we provide documentation on training, evaluating, and generating from CleaveNet.
CleaveNet is described in this preprint; if you use the code from this repository or the results, please cite the preprint.
- CleaveNet
- Table of contents
- Quick Setup
- Getting Started
- Training models
- Predicting cleavage
- Generating protease substrates
To get started quickly we have put together a jupyter notebook that simplifies using the code for substrate cleavage prediction, and unconditional substrate generation. We recommend starting here if you are less familiar with setting up and using a python package. It should take around 3-5 min to run the install line, and less than 20 minutes to execute the entire notebook using CPUs. A GPU will speed up inference for both the predictor and generator model.
To download our code, we recommend creating a clean environment with python v3.8, and installing tensorflow2.
conda env create --name cleavenet python=3.8
conda activate cleavenet
Install CleaveNet using:
pip install cleavenet
Or, for bleeding edge installation use:
pip install git+https://github.com/microsoft/cleavenet.git # bleeding edge, current repo main branch
├── README.md
├── cleavenet/ ~ package functions
├── data
├── kukreja.csv ~ dataset used for training/testing
├── cleavenet_design.txt ~ sequences and controls tested; sequences only
├── cleavenet_design_summary.xlsx ~ sequences and controls tested; MMP13 predicted Z-score, in vitro eff and selectivity
├── examples/ ~ example data needed to run test code in README.md
├── img/ ~ images in README.md
├── notebooks/
├── notebooks/
├── splits/ ~ processed data splits for model training (output of DataLoader)
├── bhatia/ ~ fluorescence test dataset
├── kukreja/ ~ mRNA display train/test splits
├── kukreja_rounded/ ~ rounded splits for conditional generation training
├── src ~ scripts for training and inference
├── generate.py
├── predict.py
├── train_generator.py
├── train_predictor.py
├── weights ~ saved weights
├── {training_scheme}_{model_arc} ~ generator models
├── {model_arc}_{ensemble} ~ predictor models
To train CleaveNet predictor:
python src/train_predictor.py
To train CleaveNet generator:
python src/train_generator.py
- The generator model is autoregressive trained over both conditional inputs (Z-scores included), and unconditional (sequence-only), next token prediction as described in the paper.
By default, these scripts will use the training splits in splits/ if present. If splits are not present it will create a new training split from data/kukreja.csv. This file was obtained from Kukreja et. al., 2015 and converted to a csv file. If you want to use your own data, we suggest either converting your data into csv file and passing it through the CleaveNet dataloader, or writing a compatible custom DataLoader. See cleavenet/data.py for an example.
You can use src/predict.py to predict cleavage of peptide sequences by our evaluated MMPs: MMP1, MMP2, MMP3, MMP7, MMP8, MMP9, MMP10,
MMP11, MMP12, MMP13, MMP14, MMP15, MMP16, MMP17, MMP19, MMP20, and MMP25.
Run src/predict.py by pointing to a csv file that contains peptide sequences. Output plots and predictions will be saved to --save-dir
For example, the following line will evaluate the Kukreja test set seqeunces.
python src/predict.py --path-to-sequence-csv splits/kukreja/X_test.csv --save-dir outputs/
The input csv should contain each sequence on a new line. Our model accepts sequences up to 10 residues in length. If sequences are shorter than 10 residues, pad them using a - token on both ends.
For example shorter sequences could look like:
---AHA----
---PRVA---
--LRVFL---
--PRVVFLR-
PRVFQLRVFL
Corresponding true Z-scores for the input sequences can be passed to the script in a CSV file using
--path-to-zscores scores.csv
These scores are not be used for prediction, but can be used to run additional evaluations between the predicted and true cleavage behavior. Output files and plots will be saved to --save-dir.
The following is an example of generating CleaveNet predictions, and comparing them to true Z-scores for the fluorescence test set.
python src/predict.py --path-to-sequence-csv splits/bhatia/X_all.csv \
--path-to-zscores splits/bhatia/y_all_labeled.csv \
--save-dir outputs/bhatia_test
Where y_all_labeled.csv contains the corresponding Z-scores for X_all.csv and the first line is a header line corresponding to the MMPs of each column. It should look like;
MMP1,MMP10,MMP12
-3.1,1.2,-2.1
4.5,2.6,-1.1
-1.5,2.1,-1.4
If you are interested in running predictions for the mRNA display test sequences evaluated in our paper you can use the following command
python src/predict.py --path-to-sequence-csv splits/kukreja/X_test.csv \
--path-to-zscores splits/kukreja/y_test.csv \
--no-csv-header
--save-dir outputs/kukreja_test
The flag --no-csv-header assumes the csv file contains no header information, and will assign each column to the MMPs used to create the splits, and used during training. This flag is not necessary for most use cases, and should be used with caution.
Notes:
- The model has not been robustly evaluated on sequences shorter or longer than 10 residues
- The default CleaveNet predictor model uses a transformer. Using the Transformer default should be sufficient for most use cases.
- In the paper we describe how in some cases an LSTM backbone can extrapolate better to unseen protease cleavage patterns. To instead use the baseline LSTM model for cleavage prediction add the flag
--model-architecture lstmtosrc/predict.py.
To generate protein sequences unconditionally using the CleaveNet generator run the following script:
python src/generate.py --num-seqs 20000 --output-dir generations
The sampling temperature is controlled by the flag --temperature, smaller numbers will result in less diverse samples and higher numbers will increase sample diversity.
A penalty can be applied to the frequency at which repeat amino acids occur by increasing the --repeat-penalty value. The default values for these are 1.0 and 1.2, respectively, as used in evaluations for the CleaveNet paper.
To conditionally generate sequences with a desired z-score profile, use the --z-scores flag. The following example will generate 400 sequences for each row provided in examples/efficient_conditional_seeds.csv, for a total of 20k samples.
python src/generate.py --num-seqs 400 --output-dir conditional_generations --z-scores examples/efficient_conditional_seeds.csv
An example z-score file is given in examples/efficient_conditional_seeds.csv. This file should contain 18 columns with the MMPs labeled and a z-score corresponding to each.
This project welcomes contributions and suggestions. Most contributions require you to agree to a Contributor License Agreement (CLA) declaring that you have the right to, and actually do, grant us the rights to use your contribution. For details, visit https://cla.opensource.microsoft.com.
When you submit a pull request, a CLA bot will automatically determine whether you need to provide a CLA and decorate the PR appropriately (e.g., status check, comment). Simply follow the instructions provided by the bot. You will only need to do this once across all repos using our CLA.
This project has adopted the Microsoft Open Source Code of Conduct. For more information see the Code of Conduct FAQ or contact opencode@microsoft.com with any additional questions or comments.
This project may contain trademarks or logos for projects, products, or services. Authorized use of Microsoft trademarks or logos is subject to and must follow Microsoft's Trademark & Brand Guidelines. Use of Microsoft trademarks or logos in modified versions of this project must not cause confusion or imply Microsoft sponsorship. Any use of third-party trademarks or logos are subject to those third-party's policies.