Troc (2,2,2-trichloroethoxycarbonyl) deprotection using zinc/acid conditions is a key step in Takeda's synthesis route. SDLabs optimized this reaction to improve yield from 50% to over 90%, navigating a deceptive optimization landscape.

The optimization landscape is deceptive: the intuitive strategy of more zinc + higher temperature leads to a local optimum (~75% yield) that masks the true global optimum (~93%). Zinc has diminishing returns above 5 eq, temperature decomposes the product above 45 °C, and acid must be carefully dosed. Aprotic solvents like THF outperform protic alternatives by 30–35%. See the full landscape description below.

The Troc deprotection has approximately 600 possible parameter combinations. Yield started at 50% and needed to exceed 90%. The landscape is deceptive: the intuitive approach of "more reagent + higher temperature" leads to a local optimum (~75%) but not the global optimum (~93%).

Single-objective Bayesian optimization navigates the deceptive landscape. Zinc has diminishing returns above 5 eq. Temperature has a sweet spot at 25–35 °C, with decomposition occurring above 45 °C. THF is the best solvent for Zn dissolution. The optimizer avoids the "high Zn + high T" trap that leads to the local optimum.

In 25 experiments (5 iterations of 5), the optimizer improves yield from an initial best of 73% to 89% — escaping the deceptive local optimum that traps conventional approaches. The first iteration finds the "high zinc + high temperature" region (~73%) which looks promising but is a dead end. Over iterations 2–4, the Gaussian Process builds a model of the temperature decomposition cliff and zinc diminishing returns. The breakthrough comes at iteration 5 when the model predicts that moderate conditions (Zn ~5.5 eq, 30 °C, THF) outperform aggressive ones — a counterintuitive result that the AI discovers from data. The progress chart shows the characteristic BO pattern: initial plateau while exploring, then a sharp improvement once the model confidence is high enough to exploit the true optimum region. Further iterations would push yield toward the theoretical maximum of ~93%.

Yield convergence — The optimizer escapes the local optimum and finds the global optimum region by iteration 5:

Temperature vs Yield — Showing the sweet spot at 25-35 °C and the decomposition cliff above 45 °C:

Solvent comparison — THF and DMF emerge as the best solvents, while protic solvents (MeOH) underperform:

This landscape is deceptive. The intuitive strategy — more zinc, higher temperature, stronger acid — leads to a local optimum around 75% yield that looks reasonable but is far from the best possible result.

Zinc has diminishing returns. Yield climbs sharply from 1 to 5 equivalents as more zinc surface area drives the reduction. But above ~6 eq, excess zinc promotes side reactions and the yield starts to drop. An experimenter adding more zinc "just to be safe" would never find the true optimum.

Temperature has a narrow sweet spot. The reaction needs enough heat to proceed (below 10 °C it barely runs), but above 45 °C the deprotected product decomposes. The optimal window is 25–35 °C — room temperature conditions that are easy to miss if you assume higher temperature means faster reaction. Worse, combining high temperature with high acid accelerates degradation even further.

Acid must be carefully dosed. Too little acid (below 1 eq) fails to activate the zinc surface. Too much (above 4 eq) decomposes the product. The optimum sits around 2–3 equivalents.

Solvent choice matters. Aprotic solvents like THF and DMF dissolve the substrate while keeping zinc reactive. Protic solvents like MeOH react with zinc directly, reducing its effectiveness and lowering yield by 30–35% relative to THF.

Reaction time has an optimal window. The reaction builds toward completion over 4–8 hours, but extended times (>12 h) lead to slow product degradation. The optimizer learns to balance completeness against stability.