README.md

Molecular Partitioning

Datasets quantifying the partitioning of small molecules in MED1, NPM1 and HP1α droplets were collected, consisting of 1143, 1055, and 963 molecules, respectively. To predict the partitioning ratio of molecules, we trained a random forest classifier and a directed message-passing neural network (MPNN) separately and aggregated their predictions. Given a molecule’s SMILES string, the models aimed to predict if the molecule’s partition ratio was above a preset threshold. A molecule’s partitioning ratio was predicted to be above a given threshold if both the random forest and MPNN models predicted a score greater than 0.5 and was predicted below the given threshold otherwise. In cases where the random forest yielded no positive molecules, only the MPNN predictions were used.

Partitioning ratios for molecules were given in 3 condensate scaffolding proteins: HP1a, MED1, and NPM1.

For each molecule-protein pair, we trained a model to predict whether the partitioning ratio ($\rho$) was above or below a threshold ($\tau$). For each condenste, 3 thresholds were used:

HP1a: [3.4, 2.7, 2.0]
MED1: [8.0, 5.4, 2.7]
NPM1: [4.5, 3.5, 2.7]

For a given threshold, the classification task is to predict $y$ from the molecule $x$:

[ y = P(\rho > \tau | x) ]

Dataset

The datasets are structured as CSV files with the following columns: ["SMILES", "PR", "THRESHOLD1", "THRESHOLD2", "THRESHOLD3"]. The SMILES column contains the SMILES string for the molecule. The PR column contains the continuous partitioning ratio for the molecule. The THRESHOLD{i} column contains the binary target label using the $i^{th}$ threshold.

For example, the a column in the HP1a dataset would look like:

SMILES	PR	THRESHOLD1	THRESHOLD2	THRESHOLD3
C1CC1	3.0	0	1	1

Deep Learning Models

Install chemprop from source as described in the chemprop docs.

1. Featurize molecule data

    python  scripts/save_features.py \ 
    --data_path /path/to/dataset.csv  
    --smiles_column SMILES \
    --save_path /path/to/rdkit_features.npz \
    --features_generator rdkit_2d_normalized
   

2. Training the models

    python train.py \
    --data_path /path/to/dataset.csv \
    --dataset_type classification \
    --target_columns Threshold{i} \       # target column, i = 1, 2, 3
    --batch_size 50 \
    --features_path /path/to/rdkit_features.npz \
    --split_type scaffold_balanced \
    --class_balance \
    --no_features_scaling \
    --split_sizes 0.8 0.1 0.1 \
    --smiles_column SMILES \
    --metric auc \
    --extra_metrics accuracy prc-auc \
    --epochs 50 \
    --ensemble 10 \
    --save_dir /path/to/models/condensate_name_tau_{i} # i = 1, 2, 3
   

3. Evaluating on held-out sets

    python predict.py \               
    --smiles_column SMILES  \
    --no_features_scaling \
    --features_path /path/to/eval_data_rdkit_features.npz \
    --test_path /path/to/eval_data.csv \
    --preds_path /path/to/eval_data_predictions/dataset.csv \
    --checkpoint_dir /path/to/models/
   

4. Extracting rationales

    python  interpret.py \
    --property_id 1 \
    --no_features_scaling  \
    --features_generator rdkit_2d_normalized  \
    --smiles_column SMILES \
    --data_path /path/to/eval_data_predictions/npm1_positive_smiles.csv  \
    --checkpoint_dir /path/to/models/ \
   

Similarly for the in vivo datasets.

Random Forest Model

Code for training and evaluating the random forest model is in the jupyter notebook.