This repository contains Python source code required to train and evaluate chemical language models for generating natural compound structures.

The CLM code is forked from https://github.com/skinniderlab/CLM.

Complete documentation for the CLM is available at https://skinniderlab.github.io/CLM/

Overview

Harvest contains of an unconditional and a conditional CLM. The unconditional CLM is a model that generates natural product-like compounds similar to real natural products.

The conditional CLM is able to generate natural product(-like) compounds based on biosynthetic cues extracted from sequence data.

A conditional Harvest model is trained on synthetic natural products, retrobiosynthesized using RetroMol. At inference time, a conditional Harvest model can ingest antiSMASH generated output file to generate natural product(-like) compounds based on the biosynthetic cues contained in the GenBank file.

Installation

To install the CLM-Harvest package, clone this repository and install the package using Conda:

cond create -f environment.yml
conda activate clm-harvest

One installed, models can be trained and evaluated using the harvest command line interface (CLI).

Usage

See harvest --help for available commands and options.

Train a CLM

Train a CLM via the Snakemake workflow from any working directory:

harvest train \
  --configfile /absolute/path/to/config_harvest_cond.yaml \
  --jobs 10 \
  --default-resources slurm_partition=skinniderlab \
  --latency-wait 60 \
  --rerun-incomplete \
  --snakemake-args --slurm

Override config values from the command line (e.g., paths.output_dir and paths.dataset):

harvest train \
  --configfile /absolute/path/to/config_harvest_cond.yaml \
  --jobs 10 \
  --snakemake-args --config paths.output_dir=/absolute/path/to/output_dir paths.dataset=/absolute/path/to/dataset.smi

If for some reason Snakemake does not allow for dotted keys, pass a YAML mapping for paths instead:

harvest train \
  --configfile /absolute/path/to/config_harvest_cond.yaml \
  --jobs 10 \
  --snakemake-args --config 'paths={output_dir: /absolute/path/to/output_dir, dataset: /absolute/path/to/dataset.smi}'

Sample an unconditional CLM

Sample an unconditional model:

harvest sample-unconditional \
  --model-dir /absolute/path/to/trained_model_dir \
  --out-dir /absolute/path/to/output_dir \
  --device cpu \
  --num-samples 1000

See harvest sample-unconditional --help for more options when sampling the unconditional CLM.

Parse compounds with RetroMol

Parse compounds with RetroMol:

harvest run-retromol \
  --data-path /absolute/path/to/input_file.csv \
  --reaction-rules-path /absolute/path/to/reaction_rules.yaml \
  --matching-rules-path /absolute/path/to/matching_rules.yaml \
  --out-dir /absolute/path/to/output_dir \

You can adjust the number of workers with --num-workers to improve performance when parsing large datasets with RetroMol. See harvest run-retromol --help for more details.