Stage 01

De novo molecule generation

The first stage has to deliver a broad, diverse set of drug-like molecules as input material for the filters that follow, rather than a narrow series of analogues of one chemotype.

Score distributions across library branches and recovery of known actives
Score distributions across generated library branches.

Reinforcement-learning setup#

Generation runs in REINVENT4 in reinforcement-learning mode: batches of 128 molecules per step, 10 to 100 steps, with early stopping at a mean reward of 0.60. The structural target is Escherichia coli dihydrofolate reductase (PDB 6XG5).

Reward function#

The declared reward is the geometric mean of two components:

In the campaign we ran, the docking component was not available in the data, so generation was effectively driven by the QED term alone. We state that limitation explicitly; it is the first entry in our register of negative results. The next campaign requires a structural reward term that returns a value for every molecule and has non-zero variance.

Diversity control#

To limit the risk of distribution collapse under aggressive reward optimisation, we used a generic ChEMBL prior, a ScaffoldSimilarity diversity filter and a low stopping threshold. The retained pre-selection pools cover 3,794,691 samples from 602 batches; these are not unique structures, nor a record of every epoch.

Quality of the generated library#

53,740compounds in the analysed sub-library
33,876unique Bemis-Murcko scaffolds, 85.4% singletons
0.869mean internal diversity
91.8%below Tanimoto 0.4 to known DHFR actives

The distribution shows no concentration on a single chemotype. The most frequent scaffold accounts for 0.96% of the library.