The Human Omnibus of Targetable Pockets

Kristy Carpenter & Russ Altman

This repository accompanies the manuscript "The Human Omnibus of Targetable Pockets." HOTPocket = "The Human Omnibus of Targetable Pockets", aka the dataset. hotpocketNN = the method.

Citation

K. A. Carpenter and R. B. Altman, “The Human Omnibus of Targetable Pockets,” J. Cheminformatics, vol. 17, no. 1, p. 180, Dec. 2025, doi: 10.1186/s13321-025-01125-x.

In addition to citing our manuscript, please cite all seven constituent methods directly:

B. Huang and M. Schroeder, “LIGSITEcsc: predicting ligand binding sites using the Connolly surface and degree of conservation,” BMC Struct. Biol., vol. 6, no. 19, 2006, doi: 10.1186/1472-6807-6-19.

R. Krivák and D. Hoksza, “P2Rank: machine learning based tool for rapid and accurate prediction of ligand binding sites from protein structure,” J. Cheminformatics, vol. 10, no. 1, p. 39, Dec. 2018, doi: 10.1186/s13321-018-0285-8.

V. Le Guilloux, P. Schmidtke, and P. Tuffery, “Fpocket: An open source platform for ligand pocket detection,” BMC Bioinformatics, vol. 10, no. 1, p. 168, Dec. 2009, doi: 10.1186/1471-2105-10-168.

A. Meller et al., “Predicting locations of cryptic pockets from single protein structures using the PocketMiner graph neural network,” Nat. Commun., vol. 14, no. 1, p. 1177, Mar. 2023, doi: 10.1038/s41467-023-36699-3.

P. A. Ravindranath and M. F. Sanner, “AutoSite: an automated approach for pseudo-ligands prediction—from ligand-binding sites identification to predicting key ligand atoms,” Bioinformatics, vol. 32, no. 20, pp. 3142–3149, Oct. 2016, doi: 10.1093/bioinformatics/btw367.

W. Tian, C. Chen, X. Lei, J. Zhao, and J. Liang, “CASTp 3.0: computed atlas of surface topography of proteins,” Nucleic Acids Res., vol. 46, no. W1, pp. W363–W367, Jul. 2018, doi: 10.1093/nar/gky473.

S. Wang, J. Xie, J. Pei, and L. Lai, “CavityPlus 2022 Update: An Integrated Platform for Comprehensive Protein Cavity Detection and Property Analyses with User-friendly Tools and Cavity Databases,” J. Mol. Biol., vol. 435, no. 14, p. 168141, Jul. 2023, doi: 10.1016/j.jmb.2023.168141.

Installation

1. Clone this repository: git clone https://github.com/Helix-Research-Lab/HOTPocket
2. Download the data from Zenodo and move all desired files into this directory (HOTPocket)
3. Install dependencies:
   1. Set up hotpocket environment (required): conda env create --name hotpocket --file=environment.yml
   2. Install dependencies for ESM2 (required to run hotpocketNN on new structures): follow instructions here
4. Run setup script: ./setup.sh

Usage

Browsing HOTPocket

We provide the 2.4 million predicted pockets across the whole human proteome as described in the accompanying manuscript in the proteome_hotpocket_embs_NN_score.csv file available on Zenodo. An alternate version computing using Feature Set C (see manuscript for feature set definitions) is available as proteome_hotpocket_comb_NN_score.csv, also on Zenodo. For more information on how to use these dataframes, see the data dictionary.

🚧 Coming soon: Website for easier browsing! 🚧

Running hotpocketNN on a novel structure(s)

If your structure of interest does not have precomputed hotpocketNN pockets, you can compute them yourself following these steps:

1. Get your structure file(s), sequence(s), and input dataframe

   You must have PDB structure files of any proteins you would like to run hotpocketNN on. Generate a fasta file containing the exact sequence represented in the PDB file using the pdbs_to_fasta function in utils.py.

   For the next step, you will need a dataframe of information about your structure. This should take a similar form as data/proteome_structure_files_canonical.csv; there needs to be uniprot id, structure type, structure id, and structure file columns. See the data dictionary for descriptions of these columns. The structure id does not need to match the filename of your PDB file. Note that the structure file column in the version of proteome_structure_files_canonical.csv from Zenodo has nan for all rows -- to use this dataframe, please fill in the path to your local copies of the structure files.

2. Run constituent pocket-finding methods

   assembly/make_db.py contains all functionality needed to run the 7 constituent pocket-finding methods on a novel structure. For AutoSite, CASTp, CavitySpace, Fpocket, LIGSITEcs, and P2Rank, you must use the hotpocket environment. For PocketMiner, you must be using the pocketminer environment. You may either run the script from the command line with the corresponding flag (e.g. for AutoSite, python make_db.py --autosite, for PocketMiner, python make_db.py --pocketminer). You can also run the functions individually from the console with your desired parameters (e.g. if you do not need to build the directory in ../data, you can call the function with makedir=False; we have also made some of these functions more easily parallelizable).

   You may also obtain BioLiP2 annotations for your structures of interest by running python make_db.py --knowndb or calling the make_known_df() function, passing along the location of your (perhaps more recent than our) BioLiP.txt file.

3. Get ESM2 embedding

   Use the esm-extract functionality as described in the ESM2 respository to generate ESM2 embeddings. You must have the esmfold environment activated. Note that you may need to increase the --truncation_seq_length parameter if your sequence is longer than the default.

   Run: esm-extract esm2_t36_3B_UR50D [FASTA] [OUTPUT DIRECTORY] --include per_tok --truncation_seq_length [MAXIMUM SEQUENCE LENGTH]

4. Run hotpocketNN

   hotpocketNN will take the candidate pockets predicted by the constituent pocket-finding methods and the ESM2 embeddings and generate a score, with which you can filter and/or rank the output pockets.

   Navigate into the assembly directory and run python run_hotpocketnn.py --in_strucid [INPUT DATAFRAME FROM STEP 1] --embdir [EMBEDDINGS DIRECTORY FROM STEP 3] --pdbdir [DIRECTORY CONTAINING STRUCTURES FROM STEP 1] --outdir [OUTPUT DIRECTORY] --errfile [FILENAME TO WRITE ERRORS]

5. Visualize

   Optionally, visualize the output pockets using the functions contained within evaluation/get_pymol_pockets.py.

Note that this whole process can be done with as few as one structure or as many as a whole other proteome. In the large-scale case, parallelization is imperative to get results in this lifetime. If you need help parallelizing these steps, please reach out and we can provide additional scripts.

Evaluating pocket predictions

The scripts in the evaluation directory contain all functions needed to repeat the analyses conducted in the accompanying manuscript. Specifically:

• dcccriterion.py is for calculating the DCCcriterion as shown in Table 4.
• get_pymol_pockets.py is for generating Pymol selection commands that can be used to create visualizations similar to Figure 2, Figure 3, Figure 5, Figure 9, and Figure 10.
• ligand_distance_histograms.py is for generating histograms of the distance between each pocket center of mass and the nearest atom of a relevant ligand as shown in Figure 6.
• naive_roc.py is for generating the ROC curve as shown in Figure 4.
• roc_prc.py is for generating the ROC and PRC curves as shown in Figure 7 and Figure 8,

Tuning the hotpocketNN model

If you would like to explore the hotpocketNN model beyond loading the state dictionary and running it in assembly/run_hotpocketnn.py, to try out additional hyperparameters, or to replicate the model tests and hyperparameter tuning described in the manuscript, all required functions are contained in the scripts within the model directory. cnn_sweep.py and nn_sweep.py contain everything needed to run the CNN and NN models, respectively, and are set up for tracking with W&B. logistic_regression.py contains everything needed to run all variants of the logistic regression model.

Questions?

Please raise an Issue in this repository if anything seems broken or if there is additional functionality you would like to see.