1. 1. Growth inhibition screens were conducted for each ESKAPE pathogen against an antibiotic-enriched chemical library containing 2,560 diverse molecules. Growth inhibition was quantified using optical density (OD).
2. 2. Each screen was binarized based on a cut-off of the mean minus one standard deviation to determine hits (antibacterial molecules) and non-hits (non-antibacterial molecules).
3. 3. Chemical structures and their binarized antibacterial activity values were used to train 3 model types for each ESKAPE pathogen: Chemprop, Chemprop-RDKit, and random forest. A total of 21 models are hosted on this website.
4. 4. Users input molecular structures through the web interface and receive antibacterial activity predictions from all 21 models.

Input options:

Enter SMILES in textbox or Upload a CSV file containing SMILES (one per row) with the column heading "smiles", similar to example below.

Output:

Prediction scores from each of the 21 models are computed for each molecule. A prediction score is a value between 0 and 1 that denotes how confident the model is that a molecule is antibacterial. Predicted antibacterial molecules will have prediction scores closer to 1, while predicted non-antibacterial molecules will have prediction scores closer to 0.

For any input compounds that were tested against the ESKAPE pathogens during training data acquisition, this tool will additionally output the experimental optical density (OD) values in the "validated" row. OD is a measure of bacterial cell growth, where a high OD means the bacteria grew in the presence of the compound, and a low OD means the compound was able to inhibit the growth of the bacteria. For reference, an OD less than 0.06 denotes full growth inhibition. All OD values were normalized by plate based on the interquartile mean.

Several metrics are also calculated for each compound:

Sum of PS: Sum of prediction scores from all pathogen models for one compound. This metric can be used to prioritize broad-spectrum antibacterial compounds.

PPF: The ratio of the highest prediction score for a compound (PS1) to the second highest (PS2). This metric can be used to prioritize pathogen-prioritized antibacterial compounds.

Molecular weight: Size of the molecule in g/mol

clogP: Calculated octanol-water partition coefficient, where high clogP values mean the compound is more lipophilic. clogP is an important metric for solubility and bioavailability.

TNN: The TNN similarity measures the structural similarity (value between 0-1) of an input molecule to the most similar molecule (nearest neighbour) from the training set. TNN similarity closer to 1 indicates the molecules are more similar (TNN similarity = 1 means the molecules are equal). Predictions on compounds that are more similar to the training set are likely to be more accurate. Nearest neighbor SMILES from the training set are included in the TSV.

View your results on the reports page or download them as a TSV file.

Interpretation:

While all models were trained on the same datasets using the same training scheme, the three model types differ in terms of architecture and molecular representation. Prediction scores for the same molecule and pathogen will therefore vary based on the model type. Note that prediction scores do not correlate directly with likelihood of activity or potency, but rather represent model confidence.

Runtime:

Note: Predictions on 1 molecule takes ~2 minutes. Predictions on 100 molecules takes ~3.5 minutes.

Please send feedback to Jon's email (stokesjm@mcmaster.ca) or the Stokes Lab website (https://www.thestokeslab.com).