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.

This study reports the synthesis and characterization of a series of non-precious metal complexes (C1–C7) and their application in the reduction of isocyanates. Notably, organoaluminium complex C5 effectively catalysed the hydroalkoxylation of isocyanates to urethane products with excellent yields under mild, base-free conditions, demonstrating significant potential in green chemistry. Additionally, organozinc complex C1 was highly efficient in catalyzing the hydroboration of isocyanates, affording formamides and N-methylamines. Mechanistic insights into both transformations were explored through stoichiometric experiments as well as theoretical studies, and plausible catalytic pathways were proposed. This work provides an eco-friendly alternative to traditional precious metal catalysts, advancing sustainable synthetic methodologies in isocyanate reduction.

The reactions of LAlH₂ (L = HC(CMeNAr)₂, Ar = 2,6-ⁱPr₂C₆H₃) (1) with diphenylphosphane oxide [Ph₂P(O)H], diphenylphosphinamide [Ph₂P(O)NH₂], and diaryl/alkyl phosphane [(RO)₂P(O)H (R = Ph, or ⁱPr)] afford their corresponding compounds with compositions LAl(H)OP(Ph)₂ (2), LAl[OP(Ph)₂]₂ (3), LAl{[N(H)P(O)(Ph)₂][OP(Ph)₂]} (4), LAl(OⁱPr)₂ (5), and LAl(OPh)₂ (6), respectively. These reactions probably undergo a process of dehydrogenation coupling, deaminating dehydrogenation coupling, or chain-breaking coupling. It is noteworthy to mention that the reaction of compound 1 with 2 equiv. of Ph₂P(O)NH₂ results in 4 and its isomer 4a, where both the complexes exhibit different coordination towards the aluminium centre. The structure of 4a is very rare, where the aluminium centre connects with a nacnac ligand as well as with N and O to form a six-membered and a four-membered ring, respectively. All the compounds were well characterized by multinuclear NMR spectroscopy, elementary analysis, and single-crystal X-ray diffraction.

Transient Receptor Potential Vanilloid 4 (TRPV4) functions as a multimodally activated Ca$^{2+}$ permeable, non-selective cation channel involved in osmotic and mechanical cue transduction, playing a significant role in joint pain and inflammatory responses in Rheumatoid Arthritis (RA). The deep learning-driven approach has already been applied to the research of TRPV4 and its inhibitors. Large language models have demonstrated remarkable capabilities in knowledge transfer and downstream tasks, performing excellently even on small datasets of inhibitors. A fine-tuned model, TRPV4-Gemma-3B, was used to generate candidate molecules for de novo design of TRPV4 inhibitors. Additionally, our candidate screening protocol employs standard evaluation methods in small molecule drug discovery, including assessments of drug-likeness, synthetic accessibility, and target-ligand binding affinity. By combining affinity data and molecular fragments, we compared the differences between the sample data and the content generated by TRPV4-Gemma-3B. This further validated that our fine-tuned model successfully captured the structural characteristics of TRPV4 inhibitor molecules. We further validated the affinity between the molecules designed by TRPV4-Gemma-3B and the target using molecular dynamics methods, which also strengthened the reliability of our fine-tuned Gemma-based tool in redesigning TRPV4 inhibitors. In future work, further integrating reinforcement learning and multi-agent approaches to incorporate the diverse structural characteristics exhibited by inhibitor molecules into the alignment of large models will also be of significant importance.

As large-scale 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 design of stimulus-responsive materials used in drug delivery over the long term. To optimize large models for extensive dataset processing and comprehensive learning akin to a chemist’s intuition, integrating deeper chemical insights is imperative. Our study initially contrasted the performance of Bigbrid, Gemma, GPT NeoX, etc., specifically focusing on designing photoresponsive drug delivery molecules. We gathered excitation energy data through computational chemistry tools and further explored light-driven isomerisation reaction as a critical mechanism in drug delivery. Our study explored the effectiveness of incorporating human feedback into reinforcement learning to imbue large models with chemical intuition, improving their understanding of relationships involving -N=N- groups in photoisomerisa-tion transitions of light-responsive molecules. Despite progress, the limited availability of specialized domain datasets continues to be a significant challenge in maximizing the performance of large models.