Team4_AI_Drug_Targets_Rare_Diseases: RARExDrug

Docker Image to run the RARExDrug pipeline

This Docker image provides an environment for running protein-ligand interaction prediction using pretrained models such as ProtBERT and ChemBERTa with PyTorch.

🐳 Image Info

• Base Image: python:3.10-slim
• Includes:
  • PyTorch
  • HuggingFace Transformers
  • RDKit
  • pandas, scikit-learn, tqdm, and other scientific packages
• Supports: CPU only

📦 Pull the Image

docker pull fedecargo/protchem:cpu

🏗️ Run RARExDrug pipeline through Snakemake

Requires Singularity installed and configured

snakemake --use-singularity --cores 4

👥 Team members

David Liu, CareDx, Brisbane California, USA

Federico Rossi, SOPHiA GENETICS, Rolle, Switzerland

John Adedeji, Osun State University, Nigeria

Mukul Sherekar, John Hopkins University, Baltimore, MD, USA

🧪 Methods

🧬 Find a relevant rare-disease for which genetic information are available regarding the existing variants

We will focus on cystic fibrosis that is common among the world populations but still considered rare. Drug treatment of cystic fibrosis mainly targets the CFTR gene (Cystic Fibrosis Transmembrane Conductance Regulator). Drugs aim to correct the expression expression or function of the defective CFTR protein. The relevance of CFTR for th treatment of cystic fibrosis is also remarked by the number of pathogenic or likely-pathogenic variants hosted on Clinvar (https://www.ncbi.nlm.nih.gov/clinvar) that are linked to this gene. Based on Clinvar, CFTR presents 1102 Pathogenic/Likely-pathogenic variants. These 1102 variants include the following main variant types:

Mutation type	Count
Deletion	404
Duplication	127
Indel	28
Insertion	10
Microsatellite	20
Copy number loss	1
single nucleotide variant	504

Drugs currently available to treat cystic fibrosis are the following:

Drug Class	Target	Example	SMILES Structure
Potentiator	CFTR gating	Ivacaftor	CC(C)(C)C1=CC(=C(C=C1NC(=O)C2=CNC3=CC=CC=C3C2=O)O)C(C)(C)C
Corrector	CFTR folding/trafficking	Lumacaftor	CC1=C(N=C(C=C1)NC(=O)C2(CC2)C3=CC4=C(C=C3)OC(O4)(F)F)C5=CC(=CC=C5)C(=O)O
	CFTR folding/trafficking	Tezacaftor	CC(C)(CO)C1=CC2=CC(=C(C=C2N1C[C@H](CO)O)F)NC(=O)C3(CC3)C4=CC5=C(C=C4)OC(O5)(F)F
Amplifier	CFTR expression	Nesolicaftor (experimental)	C[C@H](C1=NN=C(O1)C2CC(C2)NC(=O)C3=CC(=NO3)C4=CC=CC=C4)O
Readthrough agent	Premature stop codons	Ataluren	C1=CC=C(C(=C1)C2=NC(=NO2)C3=CC(=CC=C3)C(=O)O)F
Gene therapy	CFTR gene	CRISPR, mRNA delivery	N/A
ENaC inhibitors	Sodium channel	BI 1265162 (investigational)	CCN1C2=C(C=CC(=C2)OCC(=O)NCCOCP(=O)(C)C)[N+](=C1CNC(=O)C3=NC(=CN=C3N)Cl)CC.OP(=O)(O)[O-]
Anti-inflammatory	Neutrophilic inflammation	Lenabasum	CCCCCCC(C)(C)C1=CC(=C2[C@@H]3CC(=CC[C@H]3C(OC2=C1)(C)C)C(=O)O)O
	Neutrophilic inflammation	Ibuprofen	CC(C)CC1=CC=C(C=C1)C(C)C(=O)O
Mucolytics	Mucus degradation/hydration	Dornase alfa	N/A*
	Mucus degradation/hydration	Hypertonic saline	N/A

N.B: N/A- Not applicable, N/A*- Not available

We obtained the chemical structure of Ivacaftor and we plan to obtain the same information for others drugs that targets cystic fibrosis with the aim of comparing the structure of the drugs currently used to treat cystic fibrosis with the ligand that will be identified through our machine learning model.

💾 Select a database of protein-ligand interactions to be used as input for the training of the ML model

We selected BindingDB as the database of protein-ligand interactions to be used as input for the training of the ML model (https://www.bindingdb.org/rwd/bind/chemsearch/marvin/Download.jsp). BindingDB was selected because:

1. it is provided in several iteration (data extracted from ChEMBL, Patents, Publications) thus allowing to test the quality of distinct source of information or to combine them to obtain better accuracy in the prediction of the ML model;
2. it is freely available for academic use;
3. it contains a number of protein-ligand interaction that is sufficiently high for the training of an ML model (e.g. >500,000 entries available in ChEMBL BindingDB database);
4. it provides ligand in SMILE format and sequences in AA format, which can fit the type of input files needed for the training of the ML model.

BindingDB contains examples in which a ligand binds a protein. However, the training of an ML model also requires "negative" examples of interactions where a ligand is not expected to bind the protein. To generate pairs of proteins and ligands expected to not interact, we replaced the 30% of the aminoacid of the protein sequence with some other aminoacid thus leading to a protein that is likely to lack the structural properties that would allow the establishment of a protein-ligand interaction.

🔗 Find a tool to generate embeddings from the protein sequences

SMILE representation of ligands and protein sequences cannot be used as input for the training of the ML model as they are, but instead need to be converted into embeddings which are "low-dimensional, learned vector representations of high-dimensional or complex data, where similar inputs are mapped to nearby points in the vector space". To convert SMILE and aminoacid sequences into embeddings, we first considered the tool ProteinBERT (https://github.com/nadavbra/protein_bert). ProteinBERT is able to convert protein SMILE representations into amino acid sequences and we further integrated them with ChemBERTa to effectively encode chemical structures from SMILES representations. This allows the model to be applicable across diverse datasets by relying on amino acid sequences rather than PDB structures.

🤖 Build a first-prototype of the deep learning model using a basic architecture

We built a first prototype of a deep-learning (DL) model that is expected to take the BindingDB dataset as an input and to be used to predict the interaction between known biomarkers of cystic fibrosis and ligands (drugs) that are available in BindingDB. We first considered a simple DL architecture that includes and input layer, two hiddens layers and an output layer. The model was then compiled setting the loss to 'binary_crossentropy', the optimizer to 'adam' and requiring the 'accuracy' metrics to be computed. The performance of this basic DL architecture will be compared with the performance of models that are already available.

📊 Results

🧾 Model performance

RareXDrug is trained on 7,000 samples extracted from the BindingDB ChEMBL database including 4683 positive binding events and 2317 negative binding events. The evaluation of the performance of the model reported the following metrics:

Metric	Value
True Positive	1919
True Negative	9
False Positive	1061
False Negative	11
AUC-ROC score	0.49
Accuracy	0.64
Binary cross entropy loss	0.66

The poor performance of the model was likely affected by the low number of samples (7,000) and training epochs (11). Such a low number of samples, 7,000 out of the 750,000 available in the initial dataset from BindingDB, was due to the limited computational capacity available to perform the training of the deep learning model. We plan to expand the dataset used for the training of the deep learning model and to increase the epochs included in the training to improve the performance of the model.

⚙️ Making RareXDrug available to external users

Users can access the RareXDrug model for the prediction of binding between a custom set of proteins and ligands. Moreover, the training and evaluation scripts of RareXDrug are provided as a Snakemake pipeline linked to a docker container to allow the re-training of deep learning models with custom datasets.

🎯 Conclusions

Deep learning models are a promising tool for the prediction of protein-ligand (drug) interactions and can thus be used to identify the ligand that can bind proteins implied in disease mechanism and modulate their activity. Here, we built a pipeline for the generation of protein and ligand embeddings then used for the training of a deep learning models aiming to predict the likelihood of interaction between proteins implied in cystic fibrosis and ligand candidates. We successfully generated the protein and ligand embeddings from a dataset of 7,000 samples extracted from the BindingDB ChEMBL database and we used those embeddings for the training of a deep learning model. We obtained a poor model performance in terms of accuracy, AUC-ROC score and binary cross entropy loss, probably due to the limited size of the training dataset and the low number of training epochs. Size of training dataset and number of training epochs were limited by the computational resources available at the time of training. We plan to expand the computational capacity of the training process by accessing GPU resources.

📝 Future work

1. Evaluate the model's performance on more advanced computational architectures, including graph neural networks (GNNs)
2. Evalute the model's performance on an increased number of samples
3. Extend the training across a longer number of epochs
4. Integrating multi-omics data to provide information on broader biological contexts
5. Partnering with clinical researchers will ensure real-world clinical validation of potential drug candidates.
6. Predict the ligands able to bind to proteins classified as biomarkers for Cystic Fibrosis. Protein sequences will be edited to take into account the mutations that most frequently occur in Cystic Fibrosis.
7. Deploy an interactive web-based interface to offer non-technical researchers and clinicians a rare disease drug repurposing platform.

📚 References

ClinVar

@article{landrum2016clinvar,
  title={ClinVar: public archive of interpretations of clinically relevant variants},
  author={Landrum, Melissa J and Lee, Jennifer M and Benson, Mark and Brown, Garth R and Chao, Christine and Chitipiralla, Shanmuga and Gu, Baoshan and Hart, Jennifer and Hoffman, David and Hoover, Jeff and Jang, Won and Katz, Kathi and Ovetsky, Michael and Riley, Greg and Sethi, Amar and Tully, Rachel E and Villamarin-Salomon, Ricardo and Rubinstein, Wendy S and Maglott, Donna R},
  journal={Nucleic Acids Research},
  volume={44},
  number={D1},
  pages={D862--D868},
  year={2016},
  publisher={Oxford University Press},
  doi={10.1093/nar/gkv1222}
}

Docker

@article{merkel2014docker,
  title={Docker: lightweight linux containers for consistent development and deployment},
  author={Merkel, Dirk},
  journal={Linux Journal},
  volume={2014},
  number={239},
  pages={2},
  year={2014}
}

ProtBert

@article{elnaggar2021prottrans,
  title={ProtTrans: Towards Cracking the Language of Life's Code Through Self-Supervised Deep Learning and High Performance Computing},
  author={Elnaggar, Ahmed and Heinzinger, Michael and Dallago, Christian and Rehawi, Ghalia and Wang, Yu and Jones, Llion and Gibbs, Tom and Feher, Tamas and Angerer, Christoph and Steinegger, Martin and Bhowmik, Debsindhu and Rost, Burkhard},
  journal={IEEE Transactions on Pattern Analysis and Machine Intelligence},
  volume={44},
  number={10},
  pages={7112--7127},
  year={2022},
  publisher={IEEE}
}

ChemBert

@article{chithrananda2020chemberta,
  title={ChemBERTa: Large-Scale Self-Supervised Pretraining for Molecular Property Prediction},
  author={Chithrananda, Soham and Grand, Gabriel and Ramsundar, Bharath},
  journal={arXiv preprint arXiv:2010.09885},
  year={2020}
}

Pytorch

@article{paszke2019pytorch,
  title={PyTorch: An Imperative Style, High-Performance Deep Learning Library},
  author={Paszke, Adam and Gross, Sam and Massa, Francisco and Lerer, Adam and Bradbury, James and Chanan, Gregory and Killeen, Trevor and Lin, Zeming and Gimelshein, Natalia and Antiga, Luca and Desmaison, Alban and Kopf, Andreas and Yang, Edward and DeVito, Zachary and Raison, Martin and Tejani, Alykhan and Chilamkurthy, Sasank and Steiner, Benoit and Fang, Lu and Bai, Junjie and Chintala, Soumith},
  journal={arXiv preprint arXiv:1912.01703},
  year={2019}
}

Flowchart of ML pipeline for Cystic Fibrosis Drug Prediction