Molecular Adaptive Search
describes an iterative search for molecules with desired properties, in which acquired results guide which individual measurement to obtain next within a limited evaluation budget. This is where MolScout comes in.
The screening problem
Finding useful molecules requires evidence about activity, selectivity and molecular properties. Evaluating every required property for every candidate can be costly. Sample-efficient virtual screening and active learning for molecular screening address this challenge by using available results to guide further evaluations.
How MolScout makes the evaluation budget count
MolScout coordinates where to search, which individual measurement to acquire and when to complete a candidate’s profile. New evidence changes both the molecular regions explored and the measurements prioritized. This connects exploration and follow-up evaluation to the same discovery objective and remaining budget, aiming to identify strong, fully evaluated candidates earlier. It supports single-target screening and budget-aware multi-target optimization.
For example, under a limited evaluation budget, MolScout may choose between evaluating a new molecule against JAK2 and evaluating LCK for a candidate with a promising measured JAK2 result. One option explores a new candidate; the other fills a gap in an existing profile and provides evidence about selectivity.
Learning from partial results and chemical spaces
Results guide the search before a molecule’s full profile is complete, while measured outcomes remain distinct from predictions. Structural, physicochemical, learned and response-based chemical spaces provide complementary views of molecular relationships, helping MolScout explore beyond close structural analogues.
Scalar and multi-objective optimization
MolScout supports both scalar objective functions, which express the discovery goal as a single score, and multi-objective optimization, which keeps competing criteria separate. A scalar score can combine information from multiple targets and molecular properties.
For multi-objective campaigns, the trade-offs found can be assessed through a Pareto front of non-dominated candidates, while hypervolume summarizes their quality and breadth against a fixed reference—helping track progress as new measurements arrive.
Where it helps—and its limits
The approach is useful when libraries are large, evaluations are costly and several requirements shape candidate selection. Results depend on the library, evaluation quality and budget. Promising candidates may remain undiscovered, and computational results still require experimental validation.