Technical Summary: Margin-Calibrated Classifier Guidance for Property-driven Synthesis Planning

Problem Statement

Synthesis planning aims to generate efficient sequences of chemical reactions to produce a target molecule. While data-driven single-step (autoregressive) retrosynthesis models have advanced significantly, they typically operate without regard for specific chemist preferences, such as favoring robust reaction types (e.g., C–C cross-couplings) or avoiding hazardous intermediates. Existing approaches to incorporate these constraints are limited: retraining generators per property is computationally expensive and brittle to constraint changes, while post-hoc filtering restricts the search to sequences already generated by the base model, leaving valid but unexplored routes unreachable.

Classifier guidance offers a lightweight alternative by steering the output of a pretrained generator at inference time using an auxiliary signal. However, the authors identify a fundamental theoretical and empirical failure mode: standard classifiers trained with cross-entropy (CE) loss on retrosynthesis data cannot provide the necessary token-level discriminability. In sparse single-disconnection reaction datasets, the "rare cell" of training examples (where a specific prefix leads to a target property but a generator-preferred token) is structurally underrepresented. Consequently, CE-trained classifiers fail to assign meaningfully different probabilities to token continuations that lead toward versus away from the target property, rendering them ineffective for overriding the generator's unconditional distributions during beam search.

Methodology: Sequence Completion Ranking (SCR)

To overcome the limitations of cross-entropy training, the authors propose Sequence Completion Ranking (SCR), a novel training procedure that calibrates token-level classifiers to ensure they can discriminate between continuations during decoding.

1. Theoretical Analysis of CE Failure

The authors formally analyze the sample complexity required for a CE-trained classifier to achieve the necessary discriminability gap ($\Delta_k > G_k$). They demonstrate that on retrosynthesis data, the number of "rare-cell" examples (prefixes where the target property requires a token the generator disfavors) is insufficient to overcome sampling variance. Even with large datasets, the CE estimator collapses toward the generator's preference, failing to provide the signal needed to steer the beam search.

2. SCR Training Pipeline

SCR addresses this through two complementary modifications:

• Contrastive Augmentation: For every partial sequence, the method constructs a negative example by replacing the final token with a uniformly random vocabulary token. This token is assigned a "catch-all" label distinct from all reaction classes. This synthetic augmentation populates the "rare cell" at off-policy positions, providing the cross-class boundary signal that standard annotation lacks.
• Margin-Based Ranking Loss: Instead of relying solely on posterior probability estimation, SCR enforces a direct margin constraint on the guided log-scores. For a ground-truth token $r^*_k$ and the generator's top alternative $\hat{r}_k$, the loss enforces:
  $$\tilde{s}_\alpha(r^*_k) - \tilde{s}_\alpha(\hat{r}_k) \geq \gamma$$
  where $\tilde{s}_\alpha$ is the sum of the generator's log-probability and the classifier's log-probability. This loss directly installs the discriminability gap required for guidance, decoupling it from the per-prefix sample count.

3. Theoretical Guarantee

The authors prove (Proposition 1) that classifiers satisfying these per-step requirements (specifically, assigning probabilities $c_1$ to property-satisfying continuations and $c_2$ to non-satisfying ones at divergent prefixes) provably expand the set of property-satisfying sequences reachable under guided beam search. This establishes that SCR can recover targets that unguided decoding would discard.

Key Contributions

1. Theoretical Limitation of CE: The paper establishes that cross-entropy training cannot deliver the token-level discriminability required for classifier guidance on retrosynthesis data due to the structural sparsity of rare token-class combinations in single-disconnection corpora.
2. SCR Algorithm: The proposal of Sequence Completion Ranking, which combines contrastive augmentation with a margin-based ranking loss to calibrate classifiers by construction, ensuring the discriminability gap is met.
3. Reachability Proof: A formal proof demonstrating that SCR-trained classifiers expand the chemical space reachable by guided beam search, theoretically guaranteeing the recovery of property-satisfying sequences previously unreachable.
4. Empirical Validation: Extensive evaluation showing that SCR effectively steers generation toward specific reaction types and starting materials, significantly outperforming both template-free and template-based baselines.

Experimental Results

The method was evaluated on the USPTO-190 and USPTO-50k datasets using reaction type and starting material similarity (Tanimoto) as guidance targets.

• Single-Step Retrosynthesis: On USPTO-190-Steps (640 products), SCR achieved the highest steering breadth among all baselines. It recovered valid steered precursors for 231 of the 287 products where the unguided generator failed entirely. SCR achieved a steering breadth of 0.64, surpassing the best template-based method (RetroKNN at 0.51) and closing the diversity gap between template-free and template-based methods.
• Multi-Step Synthesis: Using Retro* search on 190 targets:
  • Reaction-Type Guidance: Lifted the solve rate from 16.8% (unguided) to 78.4%, a 4.7x improvement, exceeding the best template-free baseline by 12.6 percentage points.
  • Tanimoto Guidance: Achieved a 95.3% solve rate with the highest route diversity.
  • Novelty: Across both signals, guidance unlocked valid routes for 33 targets (17.4%) that no baseline could solve.
• Diversity: Guided and unguided samples exhibited only 39% Jaccard overlap, confirming that SCR explores genuinely new chemical space rather than simply reranking existing candidates.
• Efficiency: SCR-guided search required significantly fewer search nodes and model calls compared to template-based methods while achieving higher solve rates.

Significance and Claims

The paper claims that SCR provides a theoretically grounded and empirically validated solution to the problem of steerable synthesis planning. By moving beyond the limitations of cross-entropy training, SCR enables autoregressive models to incorporate chemist-specified constraints (reaction types, starting materials) at inference time without retraining.

The authors emphasize that their approach expands the reachable chemical space beyond what is possible with post-hoc filtering or search-level interventions alone. The method effectively bridges the long-standing diversity gap between template-free and template-based approaches, offering a flexible framework for property-driven synthesis planning that respects both chemical validity and specific user preferences. The work suggests that token-level guidance is a superior granularity for enforcing constraints compared to hard filtering, which often prunes the search frontier too aggressively.