mCLM: A Modular Chemical Language Model that Generates Functional and Makeable Molecules

TL;DR

mCLM addresses the gap between in silico molecule design and practical synthesis by tokenizing molecules into function-infused, automation-friendly building blocks and jointly modeling them with natural language in a bilingual, multimodal Transformer. It introduces a synthesis-aware vocabulary and a two-tokenization strategy, enabling automated assembly on modular synthesis platforms and grounding reasoning in real-world makeability. The model is trained with a unified loss over both modalities and employs critical chemical reasoning to iteratively refine designs toward multi-objective targets, validated through ADMET oracle evaluations, FDA-drug improvements, and case studies on fallen angels. The results demonstrate high validity and makeability, improved property profiles, and a scalable pathway to AI-guided, automated laboratory discovery with potential for autonomous, iterative drug and material development, while outlining future multimodal and systems-level enhancements.$\mathcal{L}=H(P(\mathbf{x}), P_\theta(\mathbf{x}))$.$

Abstract

Despite their ability to understand chemical knowledge, large language models (LLMs) remain limited in their capacity to propose novel molecules with desired functions (e.g., drug-like properties). In addition, the molecules that LLMs propose can often be challenging to make, and are almost never compatible with automated synthesis approaches. To better enable the discovery of functional small molecules, LLMs need to learn a new molecular language that is more effective in predicting properties and inherently synced with automated synthesis technology. Current molecule LLMs are limited by representing molecules based on atoms. In this paper, we argue that just like tokenizing texts into meaning-bearing (sub-)word tokens instead of characters, molecules should be tokenized at the level of functional building blocks, i.e., parts of molecules that bring unique functions and serve as effective building blocks for real-world automated laboratory synthesis. This motivates us to propose mCLM, a modular Chemical-Language Model that comprises a bilingual language model that understands both natural language descriptions of functions and molecular blocks. mCLM front-loads synthesizability considerations while improving the predicted functions of molecules in a principled manner. mCLM, with only 3B parameters, achieves improvements in synthetic accessibility relative to 7 other leading generative AI methods including GPT-5. When tested on 122 out-of-distribution medicines using only building blocks/tokens that are compatible with automated modular synthesis, mCLM outperforms all baselines in property scores and synthetic accessibility. mCLM can also reason on multiple functions and iteratively self-improve to rescue drug candidates that failed late in clinical trials ("fallen angels").

Figures (13)

• Figure 1: mCLM adopts a modular chemical vocabulary, which uses synthesis robot-friendly molecular building blocks as tokens together with natural language tokens. Compared with using natural language names, SMILES strings, or holistic embeddings for whole molecules, this building block level of tokenization better enables the prediction of compounds with improved properties, and guarantees automated synthesizability a priori, thus building a direct link between the digital and physical worlds. This approach stands to substantially enable AI-guided discovery of new small molecules with targeted functions.
• Figure 2: An overview of the mCLM. It is a multimodal chemical-language model that tokenizes molecules into synthesis robot-friendly building blocks, thus creating a direct link between the digital and physical worlds. After being trained on datasets consisting of properties, functions and synthesis data, the mCLM can conduct critical chemical reasoning through an iterative refinement process.
• Figure 3: An overview of the tokenization process. A functional molecule is first processed by the synthesis-guaranteed tokenizer to produce a set of building blocks compatible with automated modular synthesis. These blocks are then evaluated via a structure coverage check to determine whether they fully reconstruct the original molecule. If coverage is complete, the blocks are used directly for pretraining. Otherwise, the molecule is reprocessed using a rule-based tokenizer to ensure full representation for training purposes.
• Figure 4: Four of the most frequent mCLM-proposed modifications to improve DILI in FDA-approved drugs.
• Figure 5: Examples of fallen angel property modification.