Non-small cell lung cancer (NSCLC) accounts for approximately 85% of all lung cancer cases and remains a leading cause of cancer-related death worldwide. While targeted therapies against specific mutations (such as EGFR inhibitors like Gefitinib and ALK inhibitors like Crizotinib) have improved outcomes for patients who harbor those genetic changes, many patients lack these targetable mutations, and resistance to existing treatments remains a pervasive problem.
STAT3 (Signal Transducer and Activator of Transcription 3) is a protein that acts as a molecular switch controlling whether genes involved in cell survival, proliferation, immune evasion, and blood vessel growth are turned on. In many cancers including NSCLC, STAT3 is abnormally active, making cancer cells grow faster, resist cell death, and evade the immune system. STAT3 overactivation is also linked to poor patient prognosis and resistance to conventional chemotherapy.
Phosphorylation at position Tyr705 is the key activation event for STAT3. When a cytokine signal (like IL-6) triggers this phosphorylation, two STAT3 molecules pair up (dimerize) and move into the cell nucleus to switch on cancer-promoting genes. Blocking this phosphorylation step, therefore, cuts off the entire downstream cascade of tumor-promoting effects.
Existing STAT3 inhibitors such as napabucasin, stattic, and OPB-51602 have shown promise in clinical trials but face challenges including poor bioavailability (the amount of drug that reaches the target), off-target effects on healthy tissues, and the development of resistance. This creates a need for new, more selective inhibitors that can be reliably delivered and tolerated in patients.
Traditional drug discovery relies on scientists manually designing molecules based on chemical intuition, then testing them one by one - a slow, expensive process that explores only a small fraction of possible chemical structures. Machine learning offers a fundamentally different approach: learning from existing drug molecules to generate entirely new, optimized structures automatically.
Generative deep learning (GDL) uses artificial neural networks to learn the statistical patterns of known drug molecules and generate novel molecules with desired properties. The approach in this study used a conditional recurrent neural network (cRNN) - a type of network that processes sequences, well-suited to the string-based representations (called SMILES) used to describe chemical structures in computers.
The workflow in this study proceeded in stages: first generating a large virtual library of novel molecules using the AI model, then filtering this library computationally using drug-like criteria, then docking the top candidates into the STAT3 protein structure to predict binding, and finally synthesizing and testing the most promising compounds in cancer cell experiments.
Transfer learning was the key technique enabling the AI model to generate STAT3-targeted molecules. The model was first trained on 12 million diverse drug-like compounds from the ZINC database to learn general chemical language, then fine-tuned on 411 known STAT3 inhibitors from the ChEMBL database to steer molecule generation toward STAT3-relevant chemical space.
The generative model used SMILES notation - a text-based system that encodes molecular structure as a linear string - as its input and output format. A bidirectional long short-term memory (LSTM) network served as a feature extractor, reading the input molecule's SMILES string to create a compressed representation, which then fed into a second LSTM that generated new SMILES strings one character at a time. A softmax layer ensured each generated character was chemically valid according to molecular grammar rules.
The generation process produced 39,477 molecules, of which 27,863 survived initial filtering to remove duplicates, unstable structures, and molecules that could not be converted to 3D conformations. This cleaned library then underwent drug-likeness screening based on revised Lipinski rules (molecular weight under 700 Da, LogP under 5, limited hydrogen bond donors and acceptors), yielding candidates with predicted favorable absorption and bioavailability.
Molecular docking was performed using a hierarchical two-step approach. First, AutoDock Vina rapidly screened all compounds against the STAT3 protein structure (obtained from PDB ID: 6nuq), ranking them by predicted binding energy and selecting the top 10%. Then the CYSCORE and FitDock tools refined this selection further, using more precise binding affinity calculations and structural similarity clustering to identify chemically diverse high-affinity candidates.
MM/GBSA calculations (Molecular Mechanics/Generalized Born Surface Area) were used as a final computational refinement step, providing more physically accurate estimates of binding free energy than docking alone. From the top-ranked 95 molecules, compounds containing a benzotriazole chemical scaffold were prioritized for synthesis and biological testing, yielding the final lead compounds HG106 and HG110.
Cell viability assays were performed using the SRB (sulforhodamine B) method across four NSCLC cell lines (NCI-H441, NCI-H1299, NCI-H522, A549) and two normal cell lines (MRC-5 lung fibroblasts and GES-1 gastric epithelial cells). Cancer drugs ideally kill cancer cells while sparing normal cells - testing on normal cell lines allows researchers to assess selectivity and potential toxicity before moving to animal or human studies.
Colony formation assays tested whether the compounds could prevent cancer cells from proliferating over longer periods (7-10 days) at low concentrations (0.5 and 1 micromolar). Cells that survive a short treatment can still form colonies representing long-term growth capacity. This assay is considered more clinically relevant than short-term viability tests because it measures sustained anti-proliferative activity.
Apoptosis assays used flow cytometry with Annexin V and propidium iodide staining to measure the percentage of cells undergoing programmed cell death after drug treatment. Apoptosis is the preferred mechanism of cancer cell killing because it is an orderly, non-inflammatory process. Western blot analysis then confirmed which apoptosis-regulating proteins were altered, specifically measuring Bax (pro-death), Bcl-2 (pro-survival), and caspase-3 (an apoptosis executor enzyme).
STAT3 pathway inhibition was confirmed by two techniques: Western blot measured the level of phosphorylated STAT3 (p-STAT3 at Tyr705) after drug treatment, and immunofluorescence microscopy visualized whether STAT3 protein moved into the cell nucleus after IL-6 stimulation. Nuclear translocation is required for STAT3 to activate its cancer-promoting gene targets, so blocking it effectively shuts down the pathway.
UMAP visualization of the chemical space occupied by the generated molecules compared to the source (ZINC) and target (known STAT3 inhibitors) datasets showed a clear gradient: generated molecules (gray dots) distributed between the source molecules (red) and the target STAT3 inhibitors (blue). This confirmed the model successfully learned to bridge these two chemical spaces, generating molecules with STAT3 inhibitor-like properties.
Scaffold analysis revealed that 98.3% of the generated molecules contained chemical scaffolds not present in either the source or target training datasets. This high degree of novelty is critical for drug discovery - truly new scaffold types may have different toxicity profiles, better bioavailability, or overcome existing resistance mechanisms compared to previously known inhibitor classes.
The physicochemical properties of the generated molecules closely matched the distribution of known STAT3 inhibitors in terms of molecular weight, lipophilicity (LogP), drug-likeness scores (QED), and water solubility (LogS). This demonstrates that the generative model learned not just structural patterns but also the pharmacokinetic property profiles associated with effective STAT3 inhibitors.
From the final screening, 90 compounds were synthesized and tested in cell-based assays. HG106 and HG110, both built on a benzotriazole scaffold, emerged as the top performers - significantly inhibiting the growth of multiple NSCLC cell lines while showing minimal toxicity to normal cell lines, making them suitable lead compounds for further development.
Cell viability testing showed that both HG106 and HG110 potently suppressed the growth of all four tested NSCLC cell lines (NCI-H441, NCI-H1299, NCI-H522, A549) in a dose-dependent manner. Critically, minimal toxicity was observed in MRC-5 normal lung fibroblasts and GES-1 gastric epithelial cells, indicating selectivity for cancer cells over normal tissues - a key requirement for a safe therapeutic agent.
Colony formation assays confirmed that at just 1 micromolar concentration - a clinically achievable level in tissues - both compounds significantly suppressed the long-term proliferative capacity of all tested NSCLC cell lines. Soft-agar colony formation assays yielded consistent results, adding confidence that the compounds can suppress cancer cell growth under conditions that mimic the three-dimensional tissue environment.
Apoptosis induction was confirmed by flow cytometry showing a significantly elevated proportion of apoptotic NCI-H441 cells after treatment with HG106 or HG110 compared to controls. Western blot analysis showed the molecular signature of apoptosis: increased levels of Bax and caspase-3 (pro-death proteins) and decreased Bcl-2 (a survival protein). This demonstrates that the compounds kill cancer cells through the natural programmed cell death pathway rather than nonspecific toxicity.
STAT3 pathway inhibition was confirmed by showing that both compounds dose-dependently reduced phosphorylated STAT3 (p-STAT3 at Tyr705) levels in both H441 and H1299 cells. Immunofluorescence imaging further showed that HG110 prevented STAT3 from moving into the nucleus even when cells were stimulated with IL-6, the cytokine that normally activates STAT3. This confirms HG110's mechanism of action is directly targeting STAT3 phosphorylation.
Molecular dynamics (MD) simulations ran for 200 nanoseconds each for HG106 and HG110 bound to STAT3, mimicking the physical movement of atoms over time at physiological temperature and salt concentration. These simulations can reveal whether a drug molecule remains stably bound to its target protein or drifts away, and they identify precisely which amino acid residues form the key interactions.
HG106 docking analysis showed a binding score of -8.5 kcal/mol, with the molecule forming hydrogen bonds with residues Glu638 and Gln644, and hydrophobic contacts with Trp623, Val637, Tyr640, Tyr657, and Ile659 in the STAT3 binding pocket. MD simulation confirmed the complex stabilized within approximately 10 nanoseconds, with the key hydrogen bond between the methoxyl group and Gln644 remaining stable throughout the simulation.
HG110 showed superior binding stability compared to HG106. It formed additional stable hydrogen bonds with Gln644 and Tyr657, and its RMSD (a measure of structural deviation over time) of 0.16 angstroms was lower than HG106's 0.19 angstroms, indicating less movement and tighter binding. The free energy landscape showed only a single energy minimum for HG110, compared to three for HG106, confirming that HG110 adopts a single stable conformation in the binding pocket.
Quantum chemical calculations using density functional theory (DFT) provided additional evidence that HG110's mechanism is non-covalent. HG110's higher HOMO-LUMO energy gap (9.63 eV versus 9.48 eV for HG106) indicates lower chemical reactivity, meaning HG110 does not form permanent chemical bonds with the protein. This is favorable because irreversible covalent inhibitors can sometimes have unpredictable off-target effects.
HG110 represents a novel STAT3 inhibitor scaffold derived from benzotriazole chemistry, a chemical framework not previously reported as a primary scaffold for STAT3 inhibition. The compound's ability to inhibit STAT3 phosphorylation at Tyr705, block nuclear translocation, and induce apoptosis through caspase-3 activation positions it as a promising candidate for further development toward clinical use in NSCLC.
Compared to existing treatments, STAT3 inhibitors offer potential advantages: they target a central oncogenic pathway implicated in both tumor progression and resistance to other therapies. EGFR and ALK inhibitors work only in patients with specific mutations, while immune checkpoint inhibitors (though revolutionary) are ineffective in a significant proportion of patients. STAT3 inhibition could potentially benefit a broader patient population and work in combination with these existing approaches.
The generative deep learning approach demonstrated its value by producing molecules with 98.3% novel scaffolds, many of which would likely never have been considered through conventional medicinal chemistry. This chemical space exploration is particularly important for STAT3, where known inhibitors have struggled with bioavailability and off-target effects - novel scaffolds may overcome these limitations.
Limitations acknowledged by the authors include that the study was primarily in vitro (cell culture experiments), and while in vivo xenograft experiments are mentioned in the conclusion, detailed in vivo efficacy data are not fully presented. Further optimization of pharmacokinetic properties - how the drug is absorbed, distributed, metabolized, and excreted in the body - will be needed before HG110 could advance to clinical trials.
This study successfully demonstrates that generative deep learning can efficiently identify novel, active drug candidates targeting STAT3 in NSCLC. By combining AI-driven molecule generation with systematic computational and experimental validation, the researchers identified HG110 as a potent and selective STAT3 phosphorylation inhibitor with confirmed anti-tumor activity in cell culture.
HG110 mechanistically inhibits STAT3 phosphorylation at Tyr705, prevents STAT3's nuclear translocation, and triggers cancer cell death through the caspase-3 apoptosis pathway. These actions collectively suppress multiple cancer-promoting processes that STAT3 controls, including proliferation, survival, and immune evasion.
Future research priorities identified by the authors include: optimizing HG110's pharmacokinetic properties (absorption, half-life, metabolic stability) to improve clinical applicability; testing combinations with immune checkpoint inhibitors and other targeted therapies; and conducting thorough in vivo animal studies followed by preclinical toxicology assessment before any consideration of clinical trials.
The broader implication of this work is demonstrating that the combination of generative AI for molecule design, computational screening for candidate prioritization, and rigorous cell biology validation is an efficient and productive drug discovery pipeline - one that could be applied to other cancer targets beyond STAT3 and other cancer types beyond NSCLC.