Skip to content

Latest commit

 

History

History
59 lines (50 loc) · 4.38 KB

document_all_options.md

File metadata and controls

59 lines (50 loc) · 4.38 KB
Raw

# Documentation

The aim of this file is to document all options available to run the MCTS and where to find them. More details are in the attached paper at https://doi.org/10.1101/800474 , especially in the Appendix.

### Global configuration options

  • where: in the config.py file

  • how: either by modying the config.py file by hand or by running the change_config.py with its argparser (recommanded)

  • DB_CACHE: uses the MongoDB cache when activated

  • DB_REPLACE: replaces data in the Mongo DB cache when activated

  • DB_time: time cut_off for loading in the DB: stored only if above the cut-off, otherwise the rule is applied by Python

  • use_cache: dictionnary for caching results within the script. Highly recommanded.

  • retrosynthesis: performs a retrosynthetic search; biosensor: performs a biosensor search Both cannot be activated at the same time. Main difference is how the state is considered sucessful: all compounds have to be found for retrosynthesis, and only 1 for biosensors

  • add_Hs: explicit hydrogens. Recommanded at false for faster calculations.

  • use_transpositions and transposition depth: not stable. Allow for sharing of information between nodes with the same chemical state but at different places in the tree, as done in doi:10.1007/BF03192151.

Tree search configuration:

  • stop_at_first_result: stops once a signle pathway is found.
  • c_name, c_smile,s c_inchi: information on the chemical compound of interest
  • fire_timeout, standardisation_timeout: time allowed for firing a rule/standardising a compound
  • organism_name: which model to use for production of compounds
  • complementary_sink: csv file containing compounds to add to the sink. If organism is None, is the full sink.
  • representation: how to print results in logs
  • itermax: maximum number of iterations allowed for running the Tree search
  • parallel: not possible to use at the moment due to workaround for RDKit rule application. Aimed at parallelising rollouts.
  • expansion_width: maximum number of children per node
  • time budget: time allowed for running the tree search. The search will stop at the end of the iteration that exceeds this allotted time
  • max_depth: maximum depth of the Tree (also the maximum number of pathway steps)
  • minimal_visit_counts: Minimal number of times a node has to be rolled out before his brothers can be expanded
  • UCT_policy: define the UCT policy to use, ie the way to rank the children of a node. Allows various bias, and scoring considerations.
  • UCTK: the constant defining the exploration/exploitation parameter in the UCT formula
  • bias_k: if progressive bias is used, define the weight of the progressive bias in the UCT formula
  • k_rave: if RAVE is used, how to weight the RAVE. Roughly for visits below this value RAVE values lead the UCT and above, rollout values lead.
  • use_RAVE: moves have scores each time they are used throughout the Tree, adapting RAVE (Rapid Action Value Estimation) principle to the whole tree and not just rollouts.
  • penalty: penalty when no compound of the state belongs to the organism
  • full_state_reward: reward when all compounds of the state belongs to the organism
  • pathway_scoring: how to score a pathway when it is found.
  • Rollout_policy: how to select moves for the Rollout: randomly, wieghting by which scores. Many options.
  • max_rollout: maximum length of the rollout (it also stops when max_depth is reached)
  • chemical_scoring: chose the way to chemically score reactions (considering only substrates or both substrates and products). Possibility to use ConstantChemicalScorer which always returns 1.
  • biological_score_cut_off: cuts off with biological score at the specified level
  • substrate_only_score_cut_off: cuts off with substrate similarity only score BEFORE applying the rule at the specified level
  • chemical_score_cut_off: cuts off with specified chemical score AFTER applying the rule at the specified level
  • virtual_visits: start nodes at virtual_visits values, to avoid stochasticity at initial simulations; used to avoid having to much variability at initial Monte Carlo simulations.
  • progressive_bias_strategy: policy for the progressive bias (untested)
  • progressive widening: add a child to nodes visited more than len(nodes)^2 (untested)
  • diameter: Speficy the diameters (as list) to use
  • EC_filter: allow only certain EC subclasses
  • small: development archive
  • seed: for reproducibility
  • tree_to_complete: if restarting the search from another tree.