A Multi-task Large Reasoning Model for Molecular Science

Abstract

Advancements in artificial intelligence for molecular science are necessitating a paradigm shift from purely data-driven predictions to knowledge-guided computational reasoning. Existing molecular models are predominantly proprietary, lacking general molecular intelligence and generalizability. This underscores the necessity for computational methods that can effectively integrate scientific logic with deep learning architectures. Here we introduce a multi-task large reasoning model designed to emulate the cognitive processes of molecular scientists through structured reasoning and reflection. Our approach incorporates multi-specialist modules to provide versatile molecular expertise and a chain-of-thought (CoT) framework enhanced by reinforcement learning infused with molecular knowledge, enabling structured and reflective reasoning. Systematic evaluations across 10 molecular tasks and 47 metrics demonstrate that our model achieves an average 50.3% improvement over the base architecture, outperforming over 20 state-of-the-art baselines, including ultra-large-parameter foundation models, despite using significantly fewer training data and computational resources. This validates that embedding explicit reasoning mechanisms enables high-efficiency learning, allowing smaller-scale models to surpass massive counterparts in both efficacy and interpretability. The practical utility of this computational framework was validated through a case study on the design of central nervous system (CNS) drug candidates, illustrating its capacity to bridge data-driven and knowledge-integrated approaches for intelligent molecular design.

Figures (6)

• Figure 1: Overview of the reasoning framework. (a) Current LLMs in multi-task molecular science tasks, showcasing their core capabilities and inherent challenges. (b) Molecular multi-task reasoning framework, detailing the process from user query to inference, featuring tokenization and embedding, specialist selection via a router, and a multi-specialist layer within a pre-trained LLM. This framework unifies diverse molecular tasks through data synergy and embeds chemical logic into CoT reasoning for science-grounded outputs, with an example showcasing a text-based molecular generation task.
• Figure 2: Comprehensive evaluation of model performance and synergy. (a) Overall performance against baselines, comparing our model with over ten representative LLMs across core metrics as detailed in (b), with all radar plot metrics normalized. (b) Detailed metric values across the ten tasks. (c) Ablation analysis of model architectures and training strategies, contrasting experimental setups to validate the efficacy of data synergy and specialist synergy.
• Figure 3: Post-training adaptations in specialist modules. (a) Molecular representation changes induced by model training, visualized via UMAP dimensionality reduction on sampled molecules from ten tasks processed through each specialist. (b) L2 norm variation of specialists' weights relative to initial weights, quantifying adaptation degrees. (c) Weight difference analysis between the molecule captioning specialist and scientific task specialists for text generation tasks. Heatmaps show 2D histograms of merged LoRA A/B weights per layer, with y-axis as sorted layers (e.g., Layer_n_q), x-axis as binned weights [-0.05, 0.05], and cell intensities as probability densities (darker = higher density, estimated via 50-bin histograms with units of 1/weight_value). Left: absolute density for captioning specialist (linear norm); right: signed density differences $\Delta\rho$ (symmetric norm).
• Figure 4: Case study: the pipeline for CNS drug candidate development. (a) Our model integrates multi-task reasoning for CNS drug screening and analysis. (b) Case: Text-based molecule generation for designing a novel CNS-penetrant muscarinic agonist. (c) Case: Molecular property prediction, exemplified by the BBBP task. Here, the A$\to$C$\to$I chain accounts for 19.7% and A$\to$C$\to$E$\to$I for 26.0% of reasoning paths, illustrating distribution via bar charts of multi-step inference flows. (d) Case: Retrosynthesis prediction, identifying potential reactants for the target molecule.
• Figure 5: Overview of dataset construction and feature analysis. The figure illustrates the dataset construction process (a), feature coverage analysis of instruction and curated data (b), and the molecular CoT annotation and denoising workflow (c), with detailed distributions and workflows detailed in the text.