Background: As large language models continue to expand in size and diversity, their substantial potential and the relevance of their applications are increasingly being acknowledged. The rapid advancement of these models also holds profound implications for the long-term design of stimulus-responsive materials used in drug delivery. Methods: The large model used Hugging Face’s Transformers package with BigBird, Gemma, and GPT NeoX architectures. Pre-training used the PubChem dataset, and fine-tuning used QM7b. Chemist instruction training was based on Direct Preference Optimization. Drug Likeness, Synthetic Accessibility, and PageRank Scores were used to filter molecules. All computational chemistry simulations were performed using ORCA and Time-Dependent Density-Functional Theory. Results: To optimize large models for extensive dataset processing and comprehensive learning akin to a chemist’s intuition, the integration of deeper chemical insights is imperative. Our study initially compared the performance of BigBird, Gemma, GPT NeoX, and others, specifically focusing on the design of photoresponsive drug delivery molecules. We gathered excitation energy data through computational chemistry tools and further investigated light-driven isomerization reactions as a critical mechanism in drug delivery. Additionally, we explored the effectiveness of incorporating human feedback into reinforcement learning to imbue large models with chemical intuition, enhancing their understanding of relationships involving -N=N- groups in the photoisomerization transitions of photoresponsive molecules. Conclusions: We implemented an efficient design process based on structural knowledge and data, driven by large language model technology, to obtain a candidate dataset of specific photoswitchable molecules. However, the lack of specialized domain datasets remains a challenge for maximizing model performance.

Photoresponsive drug delivery stands as a pivotal frontier in smart drug administration, leveraging the non-invasive, stable, and finely tunable nature of light-triggered methodologies. The generative pre-trained transformer (GPT) has been employed to generate molecular structures. In our study, we harnessed GPT-2 on the QM7b dataset to refine a UV-GPT model with adapters, enabling the generation of molecules responsive to UV light excitation. Utilizing the Coulomb matrix as a molecular descriptor, we predicted the excitation wavelengths of these molecules. Furthermore, we validated the excited state properties through quantum chemical simulations. Based on the results of these calculations, we summarized some tips for chemical structures and integrated them into the alignment of large-scale language models within the reinforcement learning from human feedback (RLHF) framework. The synergy of these findings underscores the successful application of GPT technology in this critical domain.

In our effort to more reliably leverage large language models for designing effective photoresponsive molecular materials for drug delivery, we developed the Explanation CPO (XCPO), an end-to-end reinforcement learning fine-tuning workflow. The maturity of RLHF techniques has greatly stimulated the flourishing application of general-purpose LLMs across various scientific domains. Compared with representative RLHF approaches like CPO, our XCPO emphasizes the potential and effectiveness of a fully AI-driven, end-to-end fine-tuning process. In this workflow, the preference dataset for RLFT is derived from the combined output of AI models and XAI tools, whereas traditional RLHF pipelines require domain researchers to identify key features based on their expertise. This streamlined improvement via integration with XAI tools has not been previously reported. Based on our machine learning evaluation and computational chemistry validation, our end-to-end XCPO approach outperformed CPO by over 27\% on the MoE-based Qwen3Moe architecture, while the designed photoresponsive molecular materials aligned well with chemical intuition. Our end-to-end workflow compensates for the subjective influence of chemists' intuitive experience, enabling the model to iterate within a fully AI-driven process.

Photoresponsive drug delivery stands as a pivotal frontier in smart drug administration, leveraging the non- invasive, stable, and finely tunable nature of light-triggered methodologies. The Generative Pre-trained Transformer (GPT) has been employed for generating molecular structures. In our study, we harnessed GPT-2 on the QM7b dataset to refine a UV- GPT model with adapters, enabling the generation of molecules responsive to UV light excitation. Utilizing the Coulomb matrix as a molecular descriptor, we predicted the excitation wavelengths of these molecules. Furthermore, we validated the excited state properties through Quantum chemical simulations. The synergy of these findings underscores the successful application of GPT technology in this critical domain.

The combination of parametric quantum circuits and density matrix coding can significantly reduce the number of parameters in artificial neural networks. The reduction in the number of model parameters helps to improve the commu- nication efficiency when training deep learning models under federated learning architectures. In this study, we showcase the enhanced communication efficiency achieved in federated learning by utilizing quantum neural networks in the context of the molecular inverse synthesis task within reinforcement learning. Specifically, we consider the federated learning task on the reinforcement learning-based retrosynthesis. we adopted the USPTO-50k chemical reaction dataset. the MLP and quantum neural network are used as the agent of the reinforcement learning algorithm, respectively. Enhancements in communica- tion efficiency stem from the capacity of encoded quantum states within quantum neural networks to effectively represent data. All instances are additionally situated within the framework of ligand molecules associated with Tau proteins.

We propose a novel Deep Active Learning (DeepAL) model-3D Wasserstein Discriminative UNet (WD-UNet) for reducing the annotation effort of medical 3D Computed Tomography (CT) segmentation. The proposed WD-UNet learns in a semi-supervised way and accelerates learning convergence to meet or exceed the prediction metrics of supervised learning models. Our method can be embedded with different Active Learning (AL) strategies and different network structures. The model is evaluated on 3D lung airway CT scans for medical segmentation and show that the use of uncertainty metric, which is parametrized as an input of query strategy, leads to more accurate prediction results than some state-of-the-art Deep Learning (DL) supervised models, e.g.,3DUNet and 3D CEUNet. Compared to the above supervised DL methods, our WD-UNet not only saves the cost of annotation for radiologists but also saves computational resources. WD-UNet uses a limited amount of annotated data (35% of the total) to achieve better predictive metrics with a more efficient deep learning model algorithm.