Generating readily synthesizable small molecule fluorophore scaffolds with reinforcement learning

TL;DR

SyntheFluor-RL is developed, a generative AI model that employs known reaction libraries and molecular building blocks to create readily synthesizable fluorescent molecule scaffolds via reinforcement learning to identify synthetically accessible fluorophores for further development.

Abstract

Developing new fluorophores for advanced imaging techniques requires exploring new chemical space. While generative AI approaches have shown promise in designing novel dye scaffolds, prior efforts often produced synthetically intractable candidates due to a lack of reaction constraints. Here, we developed SyntheFluor-RL, a generative AI model that employs known reaction libraries and molecular building blocks to create readily synthesizable fluorescent molecule scaffolds via reinforcement learning. To guide the generation of fluorophores, SyntheFluor-RL employs a scoring function built on multiple graph neural networks (GNNs) that predict key photophysical properties, including photoluminescence quantum yield, absorption, and emission wavelengths. These outputs are dynamically weighted and combined with a computed pi-conjugation score to prioritize candidates with desirable optical characteristics and synthetic feasibility. SyntheFluor-RL generated 11,590 candidate molecules, which were filtered to 19 structures predicted to possess dye-like properties. Of the 19 molecules, 14 were synthesized and 13 were experimentally confirmed. The top three were characterized, with the lead compound featuring a benzothiadiazole chromophore and exhibiting strong fluorescence (PLQY = 0.62), a large Stokes shift (97 nm), and a long excited-state lifetime (11.5 ns). These results demonstrate the effectiveness of SyntheFluor-RL in the identification of synthetically accessible fluorophores for further development.

Figures (12)

• Figure 1: SyntheFluor-RL pipeline overview. The workflow begins with training set processing followed by property predictor model development. Next, SyntheFluor-RL generates molecules using the trained property predictors within a chemical space of synthesizable molecules. These molecules are then filtered based on calculated and predicted properties, structural diversity, and novelty. Finally, selected molecules were synthesized and experimentally validated.
• Figure 2: Property Prediction Model Development. A) Distributions of photoluminescence quantum yield (PLQY), absorption wavelengths, and emission wavelengths in the ChemFluor dataset. Red lines represent cutoff values for PLQY classification model ($PLQY > 0.5$ considered fluorescent) and wavelengths considered to be in the visible spectra (420 nm to 770 nm) for absorption and emission. B) Visualizations of graph neural network architectures combined with either Morgan fingerprints or RDKit fingerprints (top two) and MLP architectures using Morgan or RDKit fingerprints (bottom two). C) ROC curves and PRC curves for PLQY classification models. D) Plots of predicted absorption versus actual absorption for the absorption regression model using the Chemprop-Morgan architecture, and predicted emission vs actual emission for the emission regression model using the Chemprop-Morgan architecture. E) Visualization of sp$^2$ network algorithm. Atoms and bonds highlighted in red are members of the largest connected network of sp$^2$-hybridized atoms in the molecule.
• Figure 3: SyntheFluor-RL development and performance. A) Schematic of the SyntheFluor-RL reinforcement learning algorithm. Step 1 shows the selection of building blocks with intermediate scoring conducted by four MLP-Morgan models (PLQY (blue), absorption (green), emission (purple), for sp$^2$ (red)), and the pairing of the final selected building blocks via Reaction 2718 to create the target molecule. Step 2 shows the evaluation of the candidate model via four models (Chemprop-Morgan for PLQY (light blue), absorption (light green), and emission (light purple), and the sp$^2$ algorithm for sp$^2$ network size (light red)). Step 3 shows how the scores from Step 2 are used to update the corresponding MLP models in Step 1 and re-weight the building blocks for the next rollout. B) An example of a non-ring forming reaction in the original set of 13 reactions (top), and a ring-forming reaction in the extended set of 70 reactions (bottom). C) Distribution of PLQY probabilities (left) and histogram of sp$^2$ network sizes (right) on the generated molecules versus a random sample of 10,000 molecules in the REAL Space.
• Figure 4: Molecule selection for synthesis. A) Histogram of intra-cluster and inter-cluster Tanimoto similarities for molecules separated into 100 clusters via K-means. B) t-SNE representation of ChemFluor training set molecules, randomly selected molecules in REAL Space, generated molecules, and the final 19 selected molecules. C) Histogram of calculated oscillator strength in the 52 selected molecules. Red line shows the 0.01 cutoff for oscillator strength.
• Figure 5: Experimental validation of fluorescent properties. A) Emission spectra of 13 synthesized molecules compared to the quinine sulfate standard at 10 mM. B) Structure of the brightest compound (13) at 10 mM in chloroform under UV lamp excitation. C) Normalized excitation and emissions spectrum of compound 13 (Exmax = 363 nm, Emmax = 460 nm). D) Fluorescence lifetime of compound 13 obtained from time-correlated single photon counting. The red line is a double exponential fit with mean lifetime of 11.55 s.