The general setup for each run was as follows:
model=MPN--conf-method=mve--metric=ucb--top-k=0.001--pool=lazy--minimize--objective=lookup
When pruning was performed, the following parameters were used:
--prune--prune-min-hit-prob=0.025
See the following sections for specific details on each experiment
download the full dataset TSV file from here. You can either use as-is, using \t as your delimiter for a LookupObjective or convert it to a CSV and remove the inchikey column. Using this file with a LookupObjective entails creating an objective config file with the --score-col value set corresponding to the column of your selected target. Yes, this means that to run MolPAL across all 58 DOCKSTRING targets, you will need to write 58 separate objective config files. This same file was used as the --library. Each experiment was repeated 5 times
The output directories should be organized in the following hierarchy:
DOCKSTRING_RUNS_ROOT
├── 0.001
│ ├── prune
│ │ ├── abl1
│ │ │ ├── rep-0 <-- output directory of a MolPAL run
│ │ │ ...
│ │ │ └── rep-N
│ │ ...
│ │ └── thrb
│ └── full
│ ├── abl1
│ ...
│ └── thrb
├── 0.002
| ├── prune
│ └── full
└── 0.004
download the dataset from here and delete any placeholder scores. When we received the dataset, there were a few molecules with docking scores around +10000 kcal/mol, so we removed any ridiculous outliers like those. Run MolPAL using that dataset as both your LookupObjective and MoleculePool. Each experiment was repeated 3 times.
The output directories should be organized in the following hierarchy:
AMPC_RUNS_ROOT
├── 0.001
│ ├── prune
│ │ ├── rep-0 <-- output directory of a MolPAL run
│ │ ...
│ │ └── rep-N
│ └── full
│ ├── rep-0
│ ...
│ └── rep-N
├── 0.002
| ├── prune
│ └── full
└── 0.004
greedy metric!
Run MolPAL using this CSV as your LookupObjective and the Enamine HTS library as your MoleculePool. The active learning experiments were repeated 5 times, and the single-batch experiments were repeated 3 times.
The output directories of the active learning runs should organized like so
HTS_SB_ROOT
└── 0.004
├── rep-0 <-- output directory of a MolPAL run
...
└── rep-N
The output directories of the single-batch runs should be organized like so:
HTS_AL_ROOT
├── 0.004
| ├── rep-0 <-- output directory of a MolPAL run
| ...
│ └── rep-N
└── 0.020
├── rep-0 <-- output directory of a MolPAL run
...
└── rep-N
First, run the process.py script for each group of experiments (i.e., the group of repetitions) as the arguments to --expts. Organize the resulting .npy files in a similar hierarchy as above, naming each file simply TARGET.npy. We analyzed the top 250 scores for each task.
NOTE: you must have the same number of repetitions for each group of experiments before proceeding.
Next, run the collate.py script, supplying the proper root directory of the top of the hierarchy and the batch sizes for which you've run the experiments.
See the corresponding notebooks for the figures you'd like to make.
Note that the code to generate figures from the original MolPAL publication has been refactored and largely removed from the repo, but the general scaffold to recreate most of these figures with the new code can be found in hts-figures.ipynb. Note: this notebook can be used to recreate figures for non-HTS datasets too! To recreate the UMAP figure, see umap_fig.py. To recreate the score histograms, see dockstring-figures.ipynb.
The general directory structure for the raw data was organized like so:
RUNS_HOME # some arbitrary directory name
├── SIZE # the init/batch-size: 0.01, 0.004, 0.002, or 0.001
│ ├── MODEL # the model used: MPN, NN, or RF
│ │ ├── METRIC # the acquisition metric used: greedy, UCB, PI, EI, TS
│ │ | ├── rep_0 # the output directory of a MolPAL run
| | | ...
│ │ | └── rep_N
│ │ ...
│ │ └── METRIC
| ...
│ └── MODEL
...
└── SIZE