MPPReasoner: Injecting Chemical Reasoning into Multimodal Large Models to Reshape the Paradigm of Molecular Property Prediction

In the fields of drug discovery and materials science, molecular property prediction is a key link in accelerating R&D processes. Traditional machine learning methods rely on large amounts of labeled data, while general-purpose large language models lack professional chemical reasoning capabilities. MPPReasoner is built on Qwen2.5-VL-7B-Instruct, systematically integrating chemical reasoning into multimodal large models through a two-stage training framework, and has shown exceptional performance on multiple benchmark datasets, opening up a new technical path for molecular property prediction.

Accurate prediction of molecular properties is a core requirement in drug discovery and materials science R&D. Traditional machine learning relies on large amounts of labeled data, and general-purpose large language models lack professional chemical reasoning capabilities, making it difficult to effectively correlate molecular structure and property relationships—thus the MPPReasoner project was born.

The core innovation of MPPReasoner lies in its multimodal input processing and reasoning enhancement design. The model simultaneously receives SMILES strings (serialized chemical information) and 2D molecular images (spatial structural relationships) to understand molecular characteristics from multiple dimensions; based on the Qwen2.5-VL-7B-Instruct base model, it introduces few-shot examples with Tanimoto similarity retrieval to improve prediction accuracy and interpretability.

MPPReasoner adopts a two-stage training approach: 1. Supervised Fine-Tuning (SFT): Trained with 16,000 carefully selected reasoning trajectories to enable the model to learn step-by-step analysis like a chemical expert; 2. Reinforcement Learning (RLPGR framework): A three-layer reward structure (basic layer: answer correctness and format; reasoning layer: logical consistency and comparative analysis; chemical layer: application of chemical principles and structural analysis) to ensure professionalism and rigor.

The project built an evaluation system with 8 datasets (4 in-domain: BACE, BBBP, SIDER, HIV; 4 out-of-domain: Bioavailability, CYP2C9_V, CYP2D6_V, AMES) to test accuracy, generalization, and robustness. Using the ROC-AUC metric, preliminary results show: BACE reaches 0.9090, BBBP reaches 0.7436, and the reasoning process can be verified by experts, enhancing the credibility of the results.

MPPReasoner has a wide range of application scenarios: in drug discovery, screening candidate compounds for blood-brain barrier penetration, toxicity, etc.; in materials science, predicting bioavailability and metabolic stability, which can shorten R&D cycles and reduce costs. Deployment requires 8 NVIDIA A100 80GB GPUs (minimum 4), and at least 100GB of storage, reflecting the resource challenges of large models in scientific computing.

Current limitations: It mainly handles 2D molecular representations, with insufficient prediction of 3D conformational properties; training data comes from public datasets, so prediction accuracy for rare/novel molecules may decrease. Future directions: Integrate 3D structural information, develop lightweight models to reduce deployment costs, expand training data to cover a wider chemical space, and realize closed-loop optimization of prediction-experiment-feedback.

MPPReasoner is a significant progress in the intersection of AI and chemistry. By injecting chemical reasoning capabilities, it improves prediction accuracy and provides interpretable and verifiable methods. With the popularization of computing resources and algorithm optimization, such technologies are expected to accelerate the scientific discovery process in drug discovery and materials science.