Large Language Model‐Assisted Prediction and Virtual Screening of Near‐Infrared‐II Fluorescent Molecules

ABSTRACT Fluorescence imaging in the second near‐infrared window (NIR‐II, 1000–1700 nm) has attracted increasing attention in biomedical applications owing to its deep tissue penetration, high spatial resolution, and low background autofluorescence. However, the rational optimization of high‐performance NIR‐II fluorophores remains challenging because of the complex relationship between molecular structure and optical properties. In this study, we integrated existing fluorescence datasets with NIR‐II fluorophore…

By the Truvace deskOctober 17, 20262 min read1 sourcePeer-reviewedTRV-2026-1663· cite & verify

Large Language Model‐Assisted Prediction and Virtual Screening of Near‐Infrared‐II Fluorescent Molecules — Fluorescence Imaging 05 by Liat Goldshaid, Efrat Rubinstein, Alexander Brandis, Dadi Segal, Noa Leshem, Ori Brenner, Vyacheslav Kalchenko, Doron Eren, Tamar Yecheskel, Yoseph Salitra, Yoram Salomon and Avigdor Scherz. CC BY-SA 2.0 · https://creativecommons.org/licenses/by-sa/2.0

In brief

Researchers assembled a dataset that merged existing fluorescence data with NIR-II fluorophore data, spanning emission wavelengths from visible light to the NIR-II window. They built a large language model framework to predict each molecule's maximum emission wavelength and compared it with conventional machine learning and deep learning models.

The model then guided a virtual screen to rank candidate fluorophores, and two were tested experimentally. The work suggests language models can speed up the search for imaging dyes, a task made hard by the tangled link between molecular structure and optical properties. The abstract reports only competitive, not superior, accuracy, so the size of the advantage remains unclear.

Main points

The dataset combined existing fluorescence data with NIR-II fluorophore data, covering emission from the visible region to the NIR-II window.
The LLM framework was benchmarked against conventional machine learning and deep learning baselines.
The LLM achieved competitive, not clearly superior, predictive performance relative to those baselines.
Virtual screening led to two candidates that were characterized experimentally in vitro and in vivo.

The gain

An LLM-based model predicted molecular emission wavelengths competitively, and its guided virtual screening identified two NIR-II fluorophore candidates that showed favorable imaging in vitro and in vivo.

The rundown

NIR-II fluorescence imaging (1000–1700 nm) is valued in biomedical applications for deep tissue penetration, high spatial resolution and low background autofluorescence. Designing high-performance fluorophores remains difficult because molecular structure relates to optical properties in complex ways.

The authors trained an LLM-based framework on a combined dataset to predict maximum emission wavelength, then used the model to prioritize candidate NIR-II fluorophores through virtual screening.

Two candidates were synthesized or obtained and experimentally characterized, showing favorable imaging performance in vitro and in vivo. The authors describe this as an efficient computational strategy for candidate prioritization and molecular optimization.