To help users assess and filter generated molecules, Makya returns the following proprietary generation scores:
Iktos Ranking
The Iktos Ranking is the actual generation scoring function, the agglomerated reward that is used to guide generators towards good molecules through a process known as Reinforcement Learning.
It is the geometric mean of several scores (which are first Gaussian transformed): the reward scores (Similarity, predictions of the selected QSAR models, 3D scores, API scores), the Confidence (if QSAR models are selected), and in the case of the Fine Tuning, the QED.
Post processed scores are not included in the Iktos ranking score.
Targets in Blueprint
This is the number of rewards scores (QSAR scores, 3D - structure or ligand based scores - and API objectives) matched by the molecule. For example, if a generator runs with 2 QSAR and 3DLB rewards (which adds 2 scores to the Iktos Ranking - the Pharmacophore Score and the Shape Score), there is a total of 4 scores that are used as rewards. Molecules would then have a "Targets in Blueprint" score between 0 and 4, 0 indicating that none of the objectives is met, and 4 indicating they are all met.
Confidence
Only with generators that use QSAR models
The Confidence score is an in-house metric that controls QSAR models are within their applicability domain. It is based on the presence and absence of known structural features in the full initial training set. For instance, if no alcohol is present in the training set of the QSAR model, generating such functional group during a generation would trigger a low confidence.
We recommend Confidence ≥ 0.7.
Quality
Only with the Fine Tuning
Quality is an in-house metric for molecular complexity. It helps avoid "ugly" generated compounds. It makes sure the generator does not deviate too much from drug-likeness in terms of important molecular descriptors (for example, the number of rotatable bonds).
We recommend Quality ≥ 0.7.