Running Search

Usage

syntheseus search \
    search_targets_file=[SMILES_FILE_WITH_SEARCH_TARGETS] \
    inventory_smiles_file=[SMILES_FILE_WITH_PURCHASABLE_MOLECULES] \
    model_class=[MODEL_CLASS] \
    model_dir=[MODEL_DIR] \
    time_limit_s=[NUMBER_OF_SECONDS_PER_TARGET]

Both the search targets and the purchasable molecules inventory are expected to be plain SMILES files, with one molecule per line.

The search command accepts further arguments to configure the search algorithm; see SearchConfig in cli/search.py for the complete list.

Concurrent targets and batched inference

Search remains serial by default (max_active_searches=1). To run independent target searches concurrently while batching their inference requests:

syntheseus search \
    search_targets_file=targets.smi \
    inventory_smiles_file=inventory.smi \
    model_class=RetroChimera \
    max_active_searches=32 \
    inference_batch_size=8 \
    inference_batch_wait_s=0.01  # 10 ms

max_active_searches limits the number of targets searched at once. inference_batch_size limits the number of molecules predicted together, while inference_batch_wait_s sets the maximum wait for other requests to fill a batch. Higher concurrency or larger batches may improve throughput, but can also increase memory use and per-target latency.

Per-target time limits include waiting for inference. Concurrent execution can produce different routes or solve different targets than serial execution, especially with time-limited searches or stochastic models. Output layout and resume behavior remain unchanged. If a target fails, remaining work is cancelled; running model calls must finish before the command exits.

Info

When using one of the natively supported single-step models you can omit model_dir, which will cause syntheseus to use a default checkpoint (see here for details).

Configuring the search algorithm

You can set the search algorithm explicitly using the search_algorithm argument to retro_star (default), mcts or pdvn. For all of those algorithms you can vary hyperparameters such as the policy/value functions or MCTS bound type/constant.

In practice however there may be no need to override any hyperparameters, especially if combining Retro* or MCTS with one of the natively supported models, as for those syntheseus will automatically choose sensible hyperparameter defaults (listed in cli/search_config.yml). In our experience both Retro* and MCTS show similar performance when tuned properly, but you may want to try both for your particular usecase and see which one works best empirically.