Abstract
Motivation: Retrobiosynthesis tools use different reaction rules, search strategies and biochemical priors, so each can recover transformations missed by others. Integrating their outputs could broaden pathway coverage, but ensembling can also worsen ranking. Results: SynBioCrow runs independent retrosynthesis engines, normalizes predictions into a provenance-preserving reaction graph, and separates route generation from biochemical evidence, sequence design and bounded learning. On the frozen 65-path Galaxy-SynBioCAD literature benchmark used for SynBioCrow 2.3, connectivity-level partial recovery was 19/65 for the ensemble, 13/65 for RetroBioCat2, 9/65 for RetroPath and 0/65 for DORAnet; the ensemble preserved two exact RetroPath routes but worsened their ranks. SynBioCrow 2.4 subsequently repaired structured-reaction parsing and target-connected Candidate generation and prospectively locked two new external targets, 1,3-propanediol and 1,4-butanediol, before pathway truth was examined. RetroBioCat2 produced 3 and 2 backend-internal SOLVED search outcomes, respectively, whereas DORAnet produced none; however, neither backend recovered any predeclared explicit literature intermediate by exact structure. The negative prospective result was frozen without post-reveal tuning. Production ranking in 2.4 is therefore NONE and the release remains Candidate-stage; validated complete-pathway recovery is not claimed. Availability: SynBioCrow 2.4.0 is archived at DOI 10.5281/zenodo.23173173 with source code and frozen supporting evidence.