Insights on AI-guided compound ranking, synthesizability scoring, and early-stage discovery from the Alkira Bio team.
A synthesizability score above 0.7 does not mean easy - it means possible. Here is what the number actually tells you and how to use it alongside your own chemical intuition.
Read article
AutoDock and graph neural network scoring are not interchangeable. One predicts binding geometry; the other predicts binding likelihood from structure. Knowing the difference changes how you combine them.
A failed synthesis run costs roughly 3-6 bench days. When 40% of your synthesis queue routinely fails early screens, the program cost compounds quietly. We ran the numbers.
We started with a known kinase inhibitor scaffold and used Alkira to score 118 bioisosteric replacements. The ranking changed which analogs we prioritized. Here is how it played out.
Retrosynthesis graphs struggle with protecting group choreography, reactive functional groups, and chiral center density. These four patterns produce high synth scores on compounds that experienced chemists would never queue.
Not all ADMET flags are equally fatal. For a CNS target, brain penetration and efflux are primary; for a gut-restricted target, oral bioavailability dominates. How to weight properties by target class.
The concern when filtering a compound series by score is that you discard chemically interesting outliers. Here is how to use confidence bands to retain the high-uncertainty compounds worth investigating.
The questions to ask any scoring vendor: what training data, what benchmark set, what does the confidence band actually represent, and can you run your own retrospective on known actives.
Sparse SAR data is not a blocker - it is a different problem. Scoring models trained on related scaffolds can still produce useful rank-ordering if you understand the transfer limitations.
A compound scoring 9.1 kcal/mol binding but flagging hERG inhibition is not a first-choice lead. The binding score is real. The tox flag is also real. Here is the decision logic for balancing them.
A compound with exceptional predicted binding but a synthesizability score under 0.4 is a research direction, not a synthesis candidate. How to use the conflict productively.
A walkthrough of submitting a 200-compound SMILES file, reading the ranked output, and identifying which compounds to prioritize for the first synthesis batch.