Unveiling Latent Knowledge in Chemistry Language Models through Sparse Autoencoders

2025-12-08

AI Summary

Chemical language models (CLMs) remain largely opaque “black boxes,” hindering understanding of how they internally represent chemical knowledge. Method: To address this, we introduce sparse autoencoders (SAEs)—a first application in CLM interpretability—to systematically decode hidden-layer activation patterns of the foundational model SMI-TED across diverse molecular datasets. Contribution/Results: Our approach successfully identifies numerous semantically coherent, interpretable features—including molecular structural motifs, physicochemical properties, and pharmacological categories—establishing a direct, human-understandable mapping between neural activations and domain-specific chemical concepts. This confirms CLMs’ intrinsic capacity to encode established chemical priors and yields the first general-purpose knowledge discovery framework tailored for chemistry AI. By bridging neural representations with expert-validated chemical semantics, our work establishes a novel paradigm for developing trustworthy, interpretable large-scale chemical models.

Technology Category

• Machine Learning: Large Multimodal Models (LMMs)
• Natural Language Processing: (Large) Language Models
• Knowledge Representation and Reasoning: Common-Sense Reasoning

Application Category

• Search and Retrieval-Augmented AI: Large language models for search
• Semantics and Knowledge: Data modeling to support human-machine intelligence, including LLMs agents, intelligent system behavior, explanations, and user-friendly interactions
• Web Mining and Content Analysis: Large pretrained models with web data

Abstract

Since the advent of machine learning, interpretability has remained a persistent challenge, becoming increasingly urgent as generative models support high-stakes applications in drug and material discovery. Recent advances in large language model (LLM) architectures have yielded chemistry language models (CLMs) with impressive capabilities in molecular property prediction and molecular generation. However, how these models internally represent chemical knowledge remains poorly understood. In this work, we extend sparse autoencoder techniques to uncover and examine interpretable features within CLMs. Applying our methodology to the Foundation Models for Materials (FM4M) SMI-TED chemistry foundation model, we extract semantically meaningful latent features and analyse their activation patterns across diverse molecular datasets. Our findings reveal that these models encode a rich landscape of chemical concepts. We identify correlations between specific latent features and distinct domains of chemical knowledge, including structural motifs, physicochemical properties, and pharmacological drug classes. Our approach provides a generalisable framework for uncovering latent knowledge in chemistry-focused AI systems. This work has implications for both foundational understanding and practical deployment; with the potential to accelerate computational chemistry research.

Problem

• Uncover interpretable features in chemistry language models
• Analyze activation patterns across diverse molecular datasets
• Identify correlations between latent features and chemical knowledge domains

Innovation

• Sparse autoencoders extract interpretable latent features
• Analyze activation patterns across diverse molecular datasets
• Correlate latent features with chemical knowledge domains