Retrosynthesis graph models are genuinely useful for estimating synthetic accessibility. They have also created a class of compounds that pass computational screening with high synthesizability scores and then fail at the bench in predictable ways. Four structural patterns produce this failure reliably enough that they are worth knowing before you trust a score.
This is not an argument against synthesizability scoring. It is an argument for calibrating your expectations in specific structural contexts where the models systematically overestimate accessibility.
Case 1: Protecting group choreography in multi-functional substrates
Retrosynthesis models make disconnections. What they do not model explicitly is the sequence of protection and deprotection steps required to prevent unintended reactions at bystander functional groups during the key transformations.
A target molecule with a free primary amine, a phenol, and an aldehyde in the same structure may show a retrosynthesis path that disconnects the carbon skeleton cleanly at each step. The model finds valid bond formations. What it misses is that the aldehyde participates in every nucleophilic reaction you attempt elsewhere in the molecule unless you protect it, and protecting the aldehyde requires conditions that may affect the amine, requiring orthogonal protection throughout. A four-step synthetic route becomes an eight-step route when you account for the protection sequence.
The score reflects the disconnection graph quality, not the protection strategy overhead. Experienced synthetic chemists often recognize this immediately from the 2D structure. Computational scoring does not.
What to look for: structures with three or more polar functional groups capable of acting as nucleophiles or electrophiles under standard coupling conditions. Compare the number of formal disconnections in the retrosynthesis path to your own estimate of how many protection steps the route would require. When those numbers diverge significantly, treat the synthesizability score as an overestimate.
Case 2: Reactive functional group combinations that self-interfere
The second pattern involves functional group combinations that are individually unproblematic but create reactivity problems when present in the same molecule. The canonical example is a terminal alkyne adjacent to a thiol group: the alkyne provides a convenient handle for click chemistry or Sonogashira coupling, but the thiol oxidizes under the same conditions or coordinates metal catalysts and poisons the coupling. The model finds both functional groups as valid reactive sites without representing their interference.
Other common combinations: free amine plus boronic acid (boronate ester formation competes with the intended amide coupling), free hydroxyl plus vinyl sulfone (Michael addition proceeds on the hydroxyl oxygen rather than the intended nucleophile), and electron-rich aromatic systems plus strong electrophiles (electrophilic aromatic substitution competes with the directed metalation you planned).
Retrosynthesis models work on individual bond disconnections and do not typically encode these pairwise interaction effects. A compound with a thiol and a vinyl sulfone both present as unprotected functional groups might score 0.73 because the disconnection logic is sound, but every step of the synthesis that involves any reagent capable of thiol-Michael addition or radical reaction requires careful sequencing.
Case 3: Chiral center density with required absolute configuration
This case is distinct from the general chirality difficulty mentioned in discussions of synthesizability scores. The specific problem is when a target molecule requires a precise absolute configuration at multiple stereocenters, and those stereocenters cannot all be set in a single stereodetermining step.
A compound with four chiral centers can, in principle, be accessed by an asymmetric synthesis route that sets two centers in a Sharpless epoxidation and two more in a subsequent Evans oxazolidinone alkylation. The disconnection graph finds both steps as chemically precedented. What the model cannot assess is whether those stereocenters are close enough in space that the two-step approach actually delivers the required relative configuration, or whether the intermediate undergoes epimerization under the aqueous acidic conditions of the first step before the second can be applied.
Diastereomeric outcome in multi-step stereoselective synthesis is highly substrate-dependent. Literature precedent for each individual step does not guarantee that the steps will cooperate in your specific substrate. Programs running CNS candidates with two or three adjacent stereocenters frequently encounter this situation. The synthesizability score predicts the route exists; it does not predict whether the diastereomers will be separable by chiral HPLC or whether the epimerically sensitive intermediate will survive long enough to be functionalized.
Case 4: Ring-closure steps in macrocycle-adjacent geometries
The fourth failure mode is not limited to macrocycles, though macrocycles are the clearest example. Any synthesis requiring a ring-closure step where the two reactive termini are constrained by a molecular geometry that disfavors the productive conformation will score better than it performs.
The retrosynthesis model proposes a ring closure. The scoring reflects that the required bond formation is chemically precedented: a lactamization, a ring-closing metathesis, an intramolecular aldol. What it cannot assess is the effective molarity at the productive transition state for your specific substrate geometry.
Effective molarity is a function of conformational preferences of the acyclic precursor. A macrolide ring closure that works at 0.1 mM substrate concentration may require dilution to 0.01 mM for a substrate with a subtly different tether geometry, dropping isolated yield from 45% to 12%. The model has no representation of the pre-cyclization geometry; it only knows whether the ring-closing transformation is chemically competent in the general sense.
The practical signal: whenever a retrosynthesis route involves an intramolecular step where the ring size is 8 to 14 atoms, treat the synthesizability score as meaningfully optimistic and add a mental correction for effective molarity uncertainty. Smaller rings and larger rings (macrolides over 16 atoms) do not trigger this concern to the same degree.
Using these patterns in practice
None of these four patterns mean you should deprioritize compounds that fall into them. They mean you should flag them for structural review before finalizing synthesis prioritization.
Alkira's scoring output will show these compounds with standard synthesizability scores. The model does not yet have a specific correction factor for each failure mode, because training data on these specific failures is sparse. What you have, reading this, is a mental checklist: protecting group choreography, reactive functional group interference, multi-chiral-center absolute configuration requirements, and ring-closure geometry effects. Running that checklist against the top-ranked compounds in your synthesis queue before booking bench time takes 15 minutes and catches a predictable fraction of the false-confidence cases that the score alone will not surface.
The score is fastest when the compound is structurally clean and well-precedented. For structural classes that fall into any of these four patterns, add your own layer of manual review. That combination outperforms either approach used alone.