Libinvent reaction filter, getting zero scores in reinvent 4.5.11 version

[nitishranalkar]

Hi,
We upgraded version from reinvent 4.2.6 to reinvent 4.5.11 version. We are using libinvent reaction filter scoring component. Libinvent without reaction filter is fine, but when using reaction filter its giving zero scores.
Can anybody please help, any possible fix?
Thanks

[halx]

Hi,

Many thanks for your interest in REINVENT and welcome to the community!

Can you please provide your input config file?

Cheers,
Hannes.

[nitishranalkar]

Below is the toml config file content.
run_type = "staged_learning"
device = "cpu" # set torch device e.g. "cpu"
tb_logdir = "tb_logs" # name of the TensorBoard logging directory
json_out_config = "_staged_learning.json" # write this TOML to JSON

[parameters]
summary_csv_prefix = "staged_learning" # prefix for the CSV file
use_checkpoint = false # if true read diversity filter from agent_file
purge_memories = false # if true purge all diversity filter memories after each stage

prior_file = "priors/libinvent.prior"
agent_file = "priors/libinvent.prior"
smiles_file = "scaffolds.smi" # 1 scaffold per line with attachment points

batch_size = 10 # network

unique_sequences = true # if true remove all duplicates raw sequences in each step
# only here for backward compatibility
randomize_smiles = true # if true shuffle atoms in SMILES randomly

[learning_strategy]

type = "dap" # dap: only one supported
sigma = 128 # sigma of the RL reward function
rate = 0.0001 # for torch.optim

[diversity_filter] # optional, comment section out or remove if unneeded

type = "IdenticalMurckoScaffold" # IdenticalTopologicalScaffold,
# ScaffoldSimilarity, PenalizeSameSmiles
bucket_size = 25 # memory size in number of compounds
minscore = 0.4 # only memorize if this threshold is exceeded
minsimilarity = 0.4 # minimum similarity for ScaffoldSimilarity
penalty_multiplier = 0.5 # penalty factor for PenalizeSameSmiles

[inception] # optional, comment sectionout or remove if unneeded

smiles_file = "sampled.smi" # "good" SMILES for guidance
memory_size = 100 # number of total SMILES held in memory
sample_size = 10 # number of SMILES randomly chosen each epoch

[[stage]]

chkpt_file = 'test1.chkpt' # name of the checkpoint file, can be reused as agent

termination = "simple" # termination criterion fot this stage
max_score = 0.98 # terminate if this total score is exceeded
min_steps = 2 # run for at least this number of steps
max_steps = 3 # terminate entire run when exceeded

[stage.scoring]
type = "geometric_mean" # aggregation function

[[stage.scoring.component]]
[stage.scoring.component.QED]
[[stage.scoring.component.QED.endpoint]]
weight = 1 # user chosen name for output
name = "QED Score" # weight to fine-tune the relevance of this component

[[stage.scoring.component]]
[stage.scoring.component.MolecularWeight]
[[stage.scoring.component.MolecularWeight.endpoint]]
name = "Molecular weight" # user chosen name for output
weight = 1 # weight to fine-tune the relevance of this component

transform.type = "double_sigmoid"
transform.high = 500.0
transform.low = 200.0
transform.coef_div = 500.0
transform.coef_si = 20.0
transform.coef_se = 20.0

[[stage.scoring.component]]
[stage.scoring.component.ReactionFilter]
[[stage.scoring.component.ReactionFilter.endpoint]]
name = "reaction_filter" # user chosen name for output
weight = 1 # weight to fine-tune the relevance of this component

params.reaction_smarts = [['C:2(=[O:3])[#6:4]>>[#7:1][].C,$(C=O):2([])[#6:4]']]
params.type = "selective"

[halx]

I had now some time to look into this. The reaction SMARTS does not make any sense. The atom mapping is inconsistent, there is a comma in the product which means OR (you can only have one defined product) and there are two strange empty brackets [] in the product. But RDKit doesn't even like the reagent side:

from rdkit.Chem import AllChem
smarts = 'C:2(=[O:3])[#6:4]>>[#7:1][].C,$(C=O):2([])[#6:4]'
AllChem.ReactionFromSmarts(smarts)

results in

ValueError: ChemicalReactionParserException: Problems constructing reactant from SMARTS: C:2(=[O:3])[#6:4]

Other issues in your TOML file:

• You set the Diversity Filter to IdenticalMurckoScaffold when you already predefine a scaffold. The only way to construct a new Bemis-Murcko scaffold is to add rings.
• The batch size is rather small. Typical values are somewhere between 50 and 150
• IIRC the LSTM Libinvent prior was not trained on randomized SMILES, so you may run into issues here.
• The Inception filter is redundant as only the classical Reinvent prior supports this.

[nitishranalkar]

Can you please provide sample data, file, which i can try?

[nitishranalkar]

Just attaching logs of run:-

You can see, all scoring components score is zero.

[halx]

The scores are all zero because the reaction filter, as the name implies, acts as a filter. Are you sure that your reaction SMARTS works properly?

[nitishranalkar]

Yeah means for previous reinvent it was working.

[nitishranalkar]

4.2.6 version

[nitishranalkar]

Can you please give some other sample data, which I can try

[nitishranalkar]

Hi @halx ,
In selective_filter.py under "reinvent\chemistry\library_design\reaction_filters\selective_filter.py"

Can we do this changes?

And in learning.py under "reinvent/runmodes/RL/learning.py"

Can we do this changes?

Will above things affect anythings in reinvent 4.5.11?

By doing this changes, my code is working fine.

Can you please suggest.. Any scientific issue could raise?
Thanks in advance.

[halx]

It is not particulary helpful to send screen shots as it makes it rather difficult to make changes. The conventional way is to issue a pull request for any code changes. A quick look suggests however that we would not go ahead with this as it changes the logic of the code which may be a backward compatibility problem and, in any case, would require careful validation.

I still have not seen your reaction SMILES which I suspect was garbled. You would need to use proper quoting using the "<>" icon from the icon list above.