Custom Multi-Target Models and Patent Screening for a Competitive Assessment
Case Study 02
Purpose-built Assay Central models across three targets — including a cross-species reconstruction — applied to a client molecule, its competitor, and thousands of patent derived analogues.
Challenge
A pharmaceutical company had a small molecule and a competitor molecule that they wanted to assess against two human targets and one non-human target. They also asked for predictions on thousands of compounds similar to their small molecule that had appeared in the patent literature — a search space far too large to evaluate by hand.
What We Did
-
1
Curated data and built models for two human targets
Two unrelated proteins — one agonist model and one inhibitor model. We curated the data and built machine learning models with our Assay Central software, selecting the best-performing model for each target.
-
2
Reconstructed a modelable third (non-human) target
Receptor-sequence recovery, reconstructed protein sequences, transcriptomic support, cross-species sequence comparison, and patent-reported organismal-activity analysis — extensive sequence work to identify the closest species with data usable for modeling in Assay Central.
-
3
Mined the patent literature
A comprehensive patent search surfaced several thousand similar molecules, which were then scored against the three Assay Centraltarget models.
-
4
Benchmarked client vs. competitor molecule
A detailed cross-target comparison of the client's small molecule against the competitor molecule, supported by the underlying predictions.
Models were validated with 5-fold cross-validation, giving the team a clear read on performance before acting on any prediction.