Research map / Label-efficient and robust learning / OOD and anomaly detection

RetroOOD: Understanding Out-of-Distribution Generalization in Retrosynthesis Prediction

AAAI 2024 · Label-efficient and robust learning › OOD and anomaly detection · approach: Out-of-distribution generalization theory · 4 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Defines label shift (retro-strategy) and covariate shift (target molecule) for single-step retrosynthesis, builds USPTO-50K OOD splits, benchmarks five models, and proposes IRM-based invariant learning and EBM-based concept enhancement as model-agnostic fixes.

Problem

Single-step retrosynthesis models degrade substantially under out-of-distribution molecules or reactions, and prior work lacked rigorous disentangling of label shift (retro-strategy distribution) from covariate shift (target molecule distribution) with benchmarks.

Approach

Formally separate label shift P(T) from covariate shift P(M) in retrosynthesis, build OOD splits per shift (template-disjoint splits by minimal/retro-template granularity; size/scaffold splits within each template), and apply two model-agnostic remedies: IRM regularization on the disconnection-site loss for covariate shift, and an energy-based model over a target-template bipartite graph to denois

Key results

• On covariate-shift splits, IRM variants improve OOD top-1 for most baselines (e.g., R_Composer OOD top-1 41.2% to 41.6% and GraphRetro 38.5% to 39.6% on mol-size), with gains described as marginal overall.
• On retro-template label-shift OOD split, concept enhancement raises MT OOD top-10 from 43.9% to 55.9%, and GLN OOD top-10 from 46.6% to 52.6%.
• Under minimal-template label shift, all baselines reach near-zero OOD top-1 (0-2.8%), showing template-free and template-based models both fail to generalize to unseen minimal templates.

Measured effects reported in the paper

Surprising findings

• Template-free models do not show a clear generalization advantage over template-based models under retro-template label shift, and GLN and MHN generalize to some unseen retro-templates.
• Minimal-template label shift causes near-total OOD failure (near 0% top-1) for nearly all baselines, contrary to the assumption that template-free models generalize well to novel templates.

Limitations

• Benchmarks are built only on the open-source USPTO-50K dataset due to licensing of larger reaction databases.
• IRM gains are marginal overall, attributed to concept drift and selection bias in USPTO-50K that violates the invariance assumption.
• Concept enhancement effect is minimal on the minimal-template OOD set because the enhancement uses only retro-templates.

Experimental setup

Datasets: USPTO-50K

Benchmarks: USPTO-50K_T (label shift, retro-template and minimal-template splits) USPTO-50K_M (covariate shift, molecular size and scaffold splits)

Metrics: top-k accuracy (k=1,3,5,10)

Baselines: GLN Molecular Transformer GraphRetro RetroComposer MHN

Builds on: Invariant Risk Minimization Molecular Transformer GLN GraphRetro RetroComposer

Closest papers by meaning

• RetroBridge: Modeling Retrosynthesis with Markov Bridges (ICLR 2024 · similarity 0.90)
• GDiffRetro: Retrosynthesis Prediction with Dual Graph Enhanced Molecular Representation and Diffusion Generation (AAAI 2025 · similarity 0.90)
• Advancing Retrosynthesis with Retrieval-Augmented Graph Generation (AAAI 2025 · similarity 0.90)
• Trustworthy Retrosynthesis: Eliminating Hallucinations with a Diverse Ensemble of Reaction Scorers (NeurIPS 2026 · similarity 0.89)
• LLM-Augmented Chemical Synthesis and Design Decision Programs (ICML 2025 · similarity 0.89)
• RETRO SYNFLOW: Discrete Flow-Matching for Accurate and Diverse Single-Step Retrosynthesis (NeurIPS 2025 · similarity 0.88)