SyntheFluor-RL: AI-Driven Fluorophore Design

• SyntheFluor-RL is a generative AI platform that automates the discovery of synthesizable small-molecule fluorophores using reinforcement learning and graph neural networks.
• It integrates a multi-objective reward system with expanded reaction templates to dynamically balance photophysical properties and synthetic feasibility.
• The system identifies fluorophores surpassing established dyes, validated through rigorous experimental synthesis and live-cell imaging.

SyntheFluor-RL is a generative artificial intelligence platform designed to automate the discovery of readily synthesizable small-molecule fluorophore scaffolds. Developed to address the frequent synthetic inaccessibility of AI-generated fluorophores, SyntheFluor-RL uses reinforcement learning (RL), an expanded reaction chemistry template set, and @@@@3@@@@ (GNN) property predictors to discover compounds that are both synthetically tractable and possess desirable photophysical characteristics. The system integrates reaction-aware molecular construction with multi-objective optimization guided by dynamically balanced photophysical and synthetic metrics, and has yielded fluorophores that surpass established dyes in key optical parameters (Sayana et al., 12 Jan 2026).

1. Reinforcement Learning Design and Synthesis Constraints

SyntheFluor-RL’s molecular generation process is formalized as a sequential decision-making problem over an overview tree $T$, with each node $N$ representing a combination of Enamine REAL Space building blocks. The action space comprises two coupled decisions: (a) selection of a chemical reaction from a library of 70 reaction templates (including 13 from SyntheMol-RL and 57 for diverse ring-forming transformations such as Suzuki–Miyaura, amide coupling, heterocycle closures), and (b) choice of building blocks to react. The reaction and building block selection at each node is governed by a value/policy function realized as an MLP-Morgan fingerprint model: $V(N)$ computes a scalar value from concatenated Morgan fingerprints of building blocks and four water solvent descriptors (SP, SdP, SA, SB). Node sampling in rollouts is probabilistic, with selection weights proportional to $\exp(V(N)/\tau)$, where $\tau$ is an automatically tuned exploration parameter.

2. Multi-objective Reward Function and Dynamic Balancing

The reward (scoring) function for SyntheFluor-RL assigns a scalar $r(m)$ to each generated molecule $m$ as a dynamically weighted sum of four predictors:

$r(m) = \sum_{k=1}^4 w_k M_k(m)$

where:

• $M_1(m) = p_{\mathrm{PLQY}>0.5}$: probability that photoluminescence quantum yield (PLQY) exceeds 0.5 (Chemprop-Morgan classifier output)
• $M_2(m) = \lambda_{\rm abs}(m)$: predicted absorption maximum (nm)
• $M_3(m) = \lambda_{\rm em}(m)$: predicted emission maximum (nm)
• $M_4(m) = S_{\pi}(m)$: size of the largest connected sp² network (integer, DFS algorithm)

The weights $\{w_k\}$ are updated dynamically in response to rolling success rates $s_k$ for each objective, defined as the proportion of recent molecules meeting designated thresholds: $p_{\mathrm{PLQY}} \ge 0.5$, $\lambda_{\rm abs} \in [420, 750]$ nm, $\lambda_{\rm em} \in [420, 750]$ nm, $S_\pi \ge 12$. Weights are increased for under-optimized objectives (i.e., with low $s_k$), prioritizing harder-to-achieve targets as generation proceeds. The precise update mechanism is not disclosed in closed form (Sayana et al., 12 Jan 2026).

3. Chemprop-Morgan Property Predictor Architectures

SyntheFluor-RL relies on Chemprop-Morgan GNN models for both binary classification (PLQY) and regression (absorption, emission), trained on curated ChemFluor datasets:

• GNN: 3-step message passing, yielding a 300-dimensional graph embedding.
• Input: concatenation of the GNN output, a 2,048-bit radius-2 Morgan fingerprint, and four solvent descriptors (SP, SdP, SA, SB).
• Output: a single-hidden-layer MLP, with sigmoid (PLQY classifier) or linear (absorption, emission regressors) activation.
• Training data:
  • PLQY: 3,055 molecule–solvent pairs (binary)
  • Absorption: 4,202 pairs (regression)
  • Emission: 4,333 pairs (regression)
• Loss functions: binary cross-entropy (PLQY); mean squared error (absorption/emission)
• Validation:
  • PLQY ROC-AUC: $0.895 \pm 0.019$
  • Absorption MAE: $13.12 \pm 1.20$ nm
  • Emission MAE: $18.95 \pm 0.99$ nm

These Chemprop-Morgan GNNs are used in the RL reward; a computationally faster MLP-Morgan model (no GNN) is deployed for value estimation during roll-outs.

4. Generation Through Filtering and Scaffold Selection

The molecule generation pipeline encompasses the following:

• Approximately 10,000 RL roll-outs over $16$ h $38$ m on 32 CPUs and 1 GPU, resulting in 11,590 synthetic candidates using 18 of 70 available reactions (notably, five of them newly introduced ring-forming types).
• Four-stage filtering cascade:
  1. $S_\pi \ge 12$ → 6,111 molecules (from 11,590)
  2. $p_{\mathrm{PLQY}} > 0.5$ → 1,855 molecules
  3. $\lambda_{\rm abs} \in [420, 750]$ nm → 1,834 molecules
  4. $\lambda_{\rm em} \in [420, 750]$ nm → 631 molecules

• Diversity selection by k-means clustering (Tanimoto on Morgan fingerprints, $k=100$), manual exemplar selection yields 52 distinct structures.

• Enamine availability screen: 34 candidates found commercially accessible.
• Final quantum chemical filter: TD-DFT at B3LYP/3-21G* level with SCRF water, oscillator strength $f_{\rm osc} > 0.01$ yields 19 scaffolds for synthesis.

5. Experimental Synthesis and Photophysical Characterization

Among the 19 selected scaffolds, 14 were synthesized by Enamine; one decomposed, leaving 13 tested. Specific isolated yields and stepwise synthetic routes are not reported. Photophysical characterization was performed in chloroform at 10 mM stock concentration:

• Excitation/emission spectra: Horiba Fluorolog 3 (1 nm steps, 4 nm slit widths, 0.1 s integration).
• Quantum yield: relative to quinine sulfate standard ($\Phi_{\rm std} = 0.62$), 330 nm excitation, matched absorbance ≈ 0.077; Compound 13 yielded $\Phi_f = 0.62$.
• Extinction coefficient: Beckman DU 640, $ε \approx 6,000~\mathrm{M}^{-1}\mathrm{cm}^{-1}$ for Compound 13.
• Fluorescence lifetimes: TCSPC, amplitude-weighted means—Compound 13: 11.5 ns (notably long for blue dyes), Compound 2: 1.8 ns, Compound 11: 1.5 ns.
• Stokes shift: 97 nm (Compound 13, Ex$_{\rm max}$ = 363 nm, Em$_{\rm max}$ = 460 nm)
• Live-cell imaging (HEK293): dose-response uptake (0–10 μM, DAPI channel, Echo Revolve microscope); Compound 13 conferred high, dose-dependent contrast.

6. Comparative Performance and Methodological Advances

SyntheFluor-RL incorporates several distinct advances over previous generative dye design frameworks:

• Compared to ChemTS-based approaches (DNMG, Sumita et al., 2022), which iterate via Random Forest models and in-loop TD-DFT, yielding 3,643 candidates and a single validated scaffold over ~5 days (1,024 CPUs), SyntheFluor-RL’s workflow, leveraging neural GNN predictors and relegating quantum chemistry to a downstream filter stage, produces 11,590 candidates and selects 19 experimental scaffolds within ~16.5 h on 32 CPUs + 1 GPU. Three distinct, bright scaffolds are experimentally validated (Sayana et al., 12 Jan 2026).
• The FLAME system (REINVENT 4 + FLSF GNN, Zhu et al., 2025) generated ~1 million molecules, with post hoc synthetic accessibility checks and a single scaffold validated. By contrast, SyntheFluor-RL enforces synthetic tractability throughout by restricting the action space to known reaction templates and available building blocks—ensuring 100% of generated candidates are, by construction, synthesizable.
• Lead scaffold (benzothiadiazole-based, Compound 13) achieves a PLQY of 0.62, Stokes shift of 97 nm, and an excited-state lifetime of 11.5 ns—figures that exceed those reported for commercial blue-emitting dyes, including DAPI, Hoechst, Atto 425, and Alexa 405.
• Eight out of 14 synthesized candidates required adoption of the newly introduced ring-forming reactions, explicitly demonstrating the impact of expanding the reaction template library on structural diversity and synthetic feasibility.

7. Significance and Future Directions

SyntheFluor-RL exemplifies rigorous synthesis-constrained molecular generation, integrating RL over reaction template libraries, GNN-based photophysical property optimization, and real-time balancing of multiple objectives. Its demonstrated ability to deliver novel, experimentally validated fluorophores with superior Stokes shift and excited-state lifetime marks a significant methodological advance relative to prior generative systems that decouple synthesis planning and property targeting. A plausible implication is that dynamic reward balancing and explicit synthetic constraints during generation can be generalized to other domains of functional molecule design where experimental realization is critical (Sayana et al., 12 Jan 2026).