Hierarchical Molecular Language Models (HMLMs)

Hasi Hays, Yue Yu, William J. Richardson · 2025

Read Hierarchical Molecular Language Models (HMLMs) by Hasi Hays, Yue Yu, William J. Richardson (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2512-00696.txt.

Introduction

Hasi Hays, Yue Yu, and William Richardson introduce Hierarchical Molecular Language Models in a 2025 paper on cell biology. Artificial intelligence is changing how biologists decode the networks cells use to signal each other. Their idea treats cell signaling as a specialized language with its own parts and rules.

In that picture, signaling molecules act as tokens, the words of the language. Protein interactions supply the syntax, or grammar, that joins them. The effects that follow supply the semantics, or meaning. Framing signals this way lets language model tools work on cell communication. That could help researchers predict how cells respond when signals change.

Catalog record

CREATORHasi Hays, Yue Yu, William J. Richardson
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE86 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2512-00696

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Hierarchical Molecular Language Models (HMLMs) Hasi Hays, Yue Yu, William J. Richardson Abstract Abstract Artificial intelligence (AI) is reshaping computational and network biology by enabling new approaches to decode cellular communication networks. We introduce Hierarchical Molecular Language Models (HMLMs), a novel framework that models cellular signaling as a specialized molecular language, where signaling molecules function as tokens, protein interactions define syntax, and functional consequences constitute semantics. HMLMs employ a transformer-based architecture adapted to accommodate graph-structured signaling networks through information transducers, mathematical entities that capture how molecules receive, process, and transmit signals. The architecture integrates multi-modal data sources across molecular, pathway, and cellular scales through hierarchical attention mechanisms and scale-bridging operators that enable information flow across biological hierarchies. Applied to a complex network of cardiac fibroblast signaling, HMLMs outperformed traditional approaches in temporal dynamics prediction, particularly under sparse sampling conditions. Attention-based analysis revealed biologically meaningful crosstalk patterns, including previously uncharacterized interactions between signaling pathways. By bridging molecular mechanisms with cellular phenotypes through AI-driven molecular language representation, HMLMs establish a foundation for biology-oriented large language models (LLMs) that could be pre-trained on comprehensive pathway datasets and applied across diverse signaling systems and tissues, advancing precision medicine and therapeutic discovery. Keywords Artificial intelligence, LLMs, Network modeling, Molecular language, Precision medicine † † footnotetext: ∗ Correspondence: [email protected] 1 INTRODUCTION Large language models (LLMs) have demonstrated remarkable capabilities in processing complex sequential and relational data across diverse domains, suggesting that language-based architectures fundamentally capture principles of information processing applicable beyond natural language. We hypothesize that cellular signaling networks, which encode biological information through molecular interactions and temporal dynamics, can be effectively modeled as a specialized form of molecular language amenable to transformer-based learning frameworks. Cellular signaling networks are complex systems that process and relay information essential for cellular responses to environmental cues. Traditional modeling approaches, including ordinary differential equations (ODEs), agent-based models, Boolean networks, and statistical methods, have been integral in understanding these networks but often fail to adequately capture the nuances of context-dependent signaling, crosstalk between pathways, and the temporal dynamics inherent in biological processes. For instance, while ODEs enable detailed mechanistic modeling, they require extensive parameterization that can limit their applicability to larger networks 1 , 2 . Conversely, Boolean networks simplify molecular interactions to binary states, which compromises their granularity 3 , 4 . Bayesian networks, while useful for identifying probabilistic relationships, can struggle to incorporate feedback mechanisms common in signaling pathways 4 . Recent literature emphasizes the essential temporal and spatial dynamics of signaling interactions, necessitating a reassessment of traditional modeling frameworks 5 , 6 . Recent developments in network biology emphasize a hierarchical organization that can better encapsulate the complexity of biological systems, encouraging integration of hierarchical information into network analyses 7 , 8 . This alignment of hierarchical structures with network representation can unveil multi-scale patterns and causal relationships. Moreover, engineering LLMs, especially their application to biological data, has paved the way for novel frameworks. These LLMs, which rely on transformer architectures, demonstrate proficiency in processing complex datasets and generating meaningful predictions, which can be adapted to understand the signaling behaviors in cellular networks 9 , 10 . Current explorations have indicated that biological sequences, analogous to linguistic structures, can leverage LLM methodologies to unravel relationships and emergent biological properties, indicating a promising cross-disciplinary synergy 11 , 12 . In light of this, we propose a novel framework, hierarchical molecular language models (HMLMs), designed to encapsulate the complexity of cellular signaling networks while utilizing advanced LLM capabilities. HMLMs aim to overcome the limitations of previous modeling approaches through multiple innovations. Specifically, HMLMs integrate multi-modal data sources and employ a graph-structured attention mechanism accommodating the intricate topology of signaling networks 13 . By adding temporal embeddings, HMLMs can better match the actual timing of signaling events, which greatly improves the accuracy of their models 8 . The hierarchical representation within HMLMs permits them to learn at various organizational levels, from individual molecular interactions to larger pathway modules, thus enabling the integration of diverse experimental data and comprehensive analysis of signaling dynamics 14 , 15 . As cellular signaling networks are inherently complex, a systems biology approach that integrates both computational and experimental data is crucial for unraveling the intricacies of cell behavior. Combining data from various sources such as genomics, proteomics, nutrigenomic, and metabolomics enhances the ability to build a comprehensive understanding of signaling pathways 16 , 17 , 18 . This data integration allows researchers to explore how different molecular players interact within the context of the entire cellular system, ultimately aiding in the identification of novel therapeutic targets. Leveraging these sophisticated techniques improves the accuracy of predictive models and facilitates the study of dynamic processes that underpin cellular decisions, making it possible to simulate how cells respond to internal and external stimuli over time. Moreover, the exploration of hierarchical representations within signaling networks can illuminate the multi-scale organization of biological systems and their functional implications. Identifying substructures and modules within a network allows for the modeling of emergent behaviors that are essential for understanding complex cellular phenomena. The application of machine learning approaches to these hierarchical models further streamline the analysis of high-dimensional data, yielding new insights into the regulation of signaling pathways and their dysregulation in disease states. By prioritizing a systems-level perspective and fostering collaboration across disciplines, researchers can significantly advance efforts in precision medicine, thus enabling tailored therapeutic strategies that leverage the unique cellular contexts of individual patients. Furthermore, the adaptable nature of computational techniques from language modeling suggests that the principles underlying these approaches can effectively translate into biological contexts. Several studies have successfully applied LLM techniques to various aspects of molecular biology, revealing critical insights that traditional methods often overlook 10 , 19 . The structured representation of data as “molecular language”, characterized by unique tokens for signaling molecules and defined syntax for molecular interactions, provides a framework where complex relationships within signaling networks can be mapped and analyzed 20 . As the realms of computational systems biology and artificial intelligence increasingly converge, HMLMs represent a significant advancement in the modeling of cellular signaling networks. The capacity of these models to predict cellular responses to perturbations such as drug treatments or genetic modifications positions them as valuable tools in therapeutic discovery and personalized medicine 21 , 19 . This study makes several key contributions to the field of systems biology and its advancement in artificial intelligence (AI), including (1) We introduce cellular signaling as a form of molecular language with its unique grammar and semantics. As such, a theoretical foundation of molecular artificial intelligence (MAI) is established, which applies language modeling techniques to signaling networks. (2) We introduce HMLMs, a new computational architecture that adapts the transformer architecture to model signaling networks as information-processing systems across multiple scales. (3) We introduuce molecular-based information transducer mechanism to effectively capture context-dependent signaling behavior of signaling networks. (4) We show how HMLMs can be used to gain mechanistic insights into signaling dynamics, identify critical network nodes, and predict the effects of perturbations, facilitating the development of targeted therapeutic interventions. (5) We propose HMLMs as a potential model biology-oriented LLMs. In conclusion, while traditional signaling network modeling approaches provide a foundational understanding of cellular processes, the integration

Read the complete text