Large language models accelerated health informatics screening framework predicts infrared triggered drug delivery...

Metal-doped hydroxyapatite (HA) is an excellent material for infrared (IR)-induced on-demand drug delivery applications. However, conventional methods for predicting the optoelectronic and photothermal properties of metal-doped hydroxyapatite require intensive computational efforts using density functional theory (DFT). In this work, we show that fine-tuned large language models, namely Darwin 1.1.5 (a question- answering and multitask fine-tuned LLaMA-7 B model with 28,000 scientific QA pairs and 21 FAIR datasets, including NagasawaOPV and QM 9 datasets with SMILES encoding) and an optical property predictor based on T5, can achieve state-of-the-art accuracy with MAD and RMSE of 0.72 eV and 0.69 eV for bandgap prediction, respectively, compared to PBE- DFT (MAD= 1.65 eV) and GNN predictors (RMSE ≈ 0.85 eV), while reducing the time cost by an estimated 50-70%. The proposed system operates as an evidence-based characterization tool and decision support system, proven by predicting a monotonically decreasing trend in bandgap as the doping concentration increases: the Ag-HA series goes from 4.312 eV (0.25 mol %) to 3.983 eV (0.75 mol %); the Au-HA series from 4.300 to 3.950 eV; the Ti- HA series from 4.000 to 3.600 eV; the Zn-HA series from 4.680 to 4.400 eV; and the Mg-HA series from 4.600 to 4.500 eV, all consistent with hybrid-functional DFT benchmarks. T5 simulations show quadratically increasing absorption between 808 nm (0.5-1.3 a.u. for 1-2 mol % Ag and Au) and plasmonic SPR peaks at 420-432 nm. The predicted photothermal conversion efficiency is 34.7 ± 2% for Ag-HA, while at 445 nm it is 18.8%. On the other hand, NIR-induced tetracycline release from Ag-HA reaches 39.21 ± 2.2.5% at 10 min, which is 2.15 times that of passive diffusion (Computational estimate yet to be verified experimentally). All dopants exhibit substitutional doping at Ca²⁺ or PO₄³⁻ sites in the P 6₃/m hexagonal structure (a = b = 9. 418 Å, c = 6. 884 Å). The entire pipeline enables quick screening of candidate structures without incurring costs for DFT computations or experiments. Our work fully validates LLM predictions, discusses its validity domain and uncertainty sources, and positions it as a decision-making tool for further synthesis and characterization. Our results demonstrate that LLM-accelerated screening is a robust, reproducible alternative to DFT studies of next-generation IR-responsive nanomedicines for oncology and infectious diseases.