SIGMA: Structure-Invariant Generative Molecular Alignment for Chemical Language Models via Autoregressive Contrastive Learning

2026-03-26

AI Summary

This work addresses the modality mismatch between graph-structured molecules and their linear string representations, which arises when structurally equivalent molecules yield divergent generation trajectories due to arbitrary linearization. To resolve this, the authors propose a structure-invariant molecular alignment mechanism that leverages autoregressive contrastive learning to align prefix hidden states sharing common suffixes, along with a token-level contrastive objective that enables the model to recognize molecular geometric symmetries while preserving efficient sequential generation. During inference, they introduce Isomorphic Beam Search (IsoBeam) to prune isomorphic decoding paths. Without altering the linear representation format, the method significantly enhances both structural fidelity and generation diversity, outperforming strong baselines across standard benchmarks and multi-objective optimization tasks with improved sample efficiency.

Technology Category

• Machine Learning: Deep Generative Models & Autoencoders
• Search and Optimization: Sampling/Simulation-based Search
• Natural Language Processing: Generation

Application Category

• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs
• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs

Abstract

Linearized string representations serve as the foundation of scalable autoregressive molecular generation; however, they introduce a fundamental modality mismatch where a single molecular graph maps to multiple distinct sequences. This ambiguity leads to \textit{trajectory divergence}, where the latent representations of structurally equivalent partial graphs drift apart due to differences in linearization history. To resolve this without abandoning the efficient string formulation, we propose Structure-Invariant Generative Molecular Alignment (SIGMA). Rather than altering the linear representation, SIGMA enables the model to strictly recognize geometric symmetries via a token-level contrastive objective, which explicitly aligns the latent states of prefixes that share identical suffixes. Furthermore, we introduce Isomorphic Beam Search (IsoBeam) to eliminate isomorphic redundancy during inference by dynamically pruning equivalent paths. Empirical evaluations on standard benchmarks demonstrate that SIGMA bridges the gap between sequence scalability and graph fidelity, yielding superior sample efficiency and structural diversity in multi-parameter optimization compared to strong baselines.

Problem

• trajectory divergence
• molecular representation
• linearization ambiguity
• structural equivalence
• autoregressive generation

Innovation

• autoregressive contrastive learning
• molecular string representation
• structure invariance
• isomorphic beam search
• trajectory divergence