📄 chemistry

MolGPT-assisted exploration of VEGFR2-related chemical space through virtual screening, experimental bridging, and molecular dynamics simulation

This study presents a comprehensive workflow integrating MolGPT-based generative design, machine learning, and computational simulations to prioritize VEGFR2 inhibitors, which was experimentally validated by testing accessible surrogate compounds that ultimately failed to confirm the predicted activity or binding mechanisms.

Original authors: Taiying Li, Shuning Diao, Yuxiang He, Shengzhen Hou, Jinping Zhang

Published 2026-10-07

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Taiying Li, Shuning Diao, Yuxiang He, Shengzhen Hou, Jinping Zhang

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Cancer is a disease of uncontrolled growth, but tumors cannot grow large without a supply line. Once a tumor reaches a few millimeters in size, it must trigger a switch to build its own network of blood vessels, a process known as angiogenesis. This new network delivers the oxygen and nutrients the cancer needs to survive and spread. One of the primary signals that tells blood vessels to grow is a protein called VEGFR2. Because this receptor is so central to the process, scientists have long tried to block it with drugs to starve tumors. However, the drugs currently available often act like a sledgehammer, hitting other targets and causing side effects, or the cancer eventually finds a way to bypass the blockage. The search continues for new, more precise molecules that can stop this growth signal without the collateral damage.

In a recent study, researchers from Shandong University of Traditional Chinese Medicine attempted to find these new molecules not by looking through existing shelves of chemicals, but by asking a computer to imagine them from scratch. They used a type of artificial intelligence called MolGPT, which is trained on the language of chemistry. Just as a language model can write a new sentence by predicting the next word based on what came before, this model predicts the next part of a chemical structure to create entirely new molecules that have never existed before. The goal was to generate thousands of these unique, theoretical structures and then filter them down to find the ones most likely to block VEGFR2.

The team started by generating 5,000 new molecular structures. After cleaning up the list to remove duplicates and invalid shapes, they were left with 4,992 unique candidates. To decide which of these new shapes might work, they ran them through a series of computer tests. First, a predictive model estimated how likely each molecule was to be active against the target. This narrowed the field down to 522 candidates. Next, they used a method called molecular docking, which simulates how a molecule fits into the specific pocket of the VEGFR2 protein, like a key trying to fit into a lock. This step selected nine molecules that appeared to fit the best. Finally, the researchers checked these nine for drug-like properties, such as whether they would be safe for the body to process and whether they could be easily made in a lab. This process left them with four top candidates.

Here, the project hit a practical wall. The four best molecules were so new and complex that they were not available for purchase, and making them from scratch in a lab would take a long time. To keep the research moving, the team employed a clever workaround. Instead of waiting to make the perfect new molecules, they searched for commercially available compounds that looked somewhat similar to the four top candidates. They found six such "surrogate" compounds in a natural product library. These surrogates were not the exact molecules the computer had designed, but they shared enough structural features to serve as a test run. The researchers then tested these six real-world compounds in a lab to see if they could actually block the VEGFR2 protein.

The results of this experimental test were largely negative, but they provided a clear picture of the challenges in this field. Five of the six surrogate compounds showed no significant ability to block the target. One compound, levomefolic acid, showed a strange and inconsistent reaction. At a concentration of 20 micromolar, it blocked the protein by about 56 percent, but when the concentration was doubled to 40 micromolar, the blocking effect dropped to just 12 percent. This erratic behavior meant the researchers could not calculate a reliable measure of its strength. Because the connection between the original computer-generated molecule and this surrogate was weak, the team could not claim that the new design worked.

To understand what might be happening, the researchers turned to computer simulations to watch how the levomefolic acid molecule behaved when it was near the protein. They simulated the interaction over a period of time to see if the molecule held on tightly or drifted away. The simulations showed that while the molecule did interact with the protein, it was less stable than a known drug used as a reference. The computer models suggested that the molecule formed some helpful connections but also faced significant resistance from the surrounding water and environment, making it a less efficient binder than expected. The study concluded that while the computer workflow successfully generated new ideas and filtered them down to a few possibilities, the bridge between the computer's imagination and the lab's reality remained unproven.

Ultimately, this research serves as a demonstration of a new workflow rather than a discovery of a new cure. The team successfully showed that artificial intelligence can generate vast numbers of novel chemical structures and that a multi-step filtering process can narrow them down to a manageable few. However, the experiment also highlighted a significant gap: the molecules the computer designs are often so different from what is available in the real world that finding a suitable substitute is difficult, and even when a substitute is found, it may not behave the way the computer predicted. The study suggests that while AI is a powerful tool for exploring the vast landscape of possible chemicals, the path from a digital design to a working medicine still requires direct synthesis and testing of the exact molecules the computer creates. The journey from a new idea to a new drug remains a complex challenge, where the promise of artificial intelligence must still be grounded in the hard work of experimental verification.

Technical Summary: MolGPT-assisted exploration of VEGFR2-related chemical space

Problem Statement
Vascular endothelial growth factor receptor 2 (VEGFR2) is a critical target for anti-angiogenic cancer therapy. However, existing inhibitors (e.g., sorafenib, sunitinib) often suffer from off-target toxicity, acquired resistance, and limited scaffold diversity. While deep learning-based generative models offer a pathway to explore chemical space beyond traditional compound libraries, a significant bottleneck remains: translating de novo-generated molecular structures into experimentally testable compounds. Prioritized molecules are frequently unavailable commercially and difficult to synthesize rapidly, creating a gap between computational design and experimental validation.

Methodology
The study proposes a multistage candidate-prioritization workflow designed to bridge generative AI with exploratory experimental assessment. The workflow consists of the following stages:

Molecular Generation: A MolGPT model (Transformer decoder-based) was used to generate 5,000 de novo small molecules. The model was not fine-tuned specifically for VEGFR2 but served as a general-purpose generator to ensure structural novelty and chemical diversity.
Chemical Space Analysis: The generated molecules were analyzed against a reference dataset of 5,564 VEGFR2-related compounds (2,841 active, 2,723 inactive) using Uniform Manifold Approximation and Projection (UMAP) and Extended Connectivity Fingerprints (ECFP4) to assess structural overlap and novelty.
Quality Assessment: The Molecular Sets (MOSES) benchmark was employed to evaluate the validity, uniqueness, novelty, and diversity of the generated library.
Virtual Screening:
- Activity Prediction: A pre-trained Graph Convolutional Network (GCN) assigned activity scores to the 4,992 unique valid structures. Molecules with a score > 0.5 were retained.
- Molecular Docking: The retained candidates were docked into the VEGFR2 ATP-binding pocket (PDB ID: 4ASE) using AutoDock Vina.
- ADMET Assessment: Candidates were filtered based on Lipinski's Rule of Five, predicted ADMET properties (using ADMETlab 3.0), and drug-likeness (QED).
Similarity Matching (Bridging Strategy): Recognizing that prioritized de novo molecules were not commercially available, the workflow searched the TargetMol Natural Product Library for structurally similar surrogate compounds. This step aimed to find experimentally accessible molecules that might share local structural features or pharmacophores with the generated candidates.
Experimental Validation: Six similarity-matched surrogate compounds were subjected to in vitro VEGFR2 kinase inhibition assays.
Computational Post-Analysis: For the surrogate showing the most notable (though limited) response, molecular dynamics (MD) simulations (100 ns) and MM–PBSA binding energy analyses were performed to generate structural hypotheses regarding binding modes and complex stability.

Key Results

Generation and Screening: MolGPT generated 4,992 unique valid structures. The MOSES benchmark indicated high validity (1.000) and uniqueness (0.998), with a novelty score of 0.824. Chemical space analysis revealed that the generated molecules had low structural similarity to the reference dataset (mean Tanimoto similarity $T_{cmax} \approx 0.30$ ), confirming the exploration of novel chemical space.
Prioritization: GCN screening retained 522 candidates. Molecular docking and ADMET filtering further narrowed this to four prioritized candidates (Cpd4, Cpd6, Cpd7, Cpd9).
Surrogate Matching: Six commercially available surrogate compounds were identified for the four prioritized candidates. The Tanimoto similarity between generated molecules and surrogates was generally low (0.25–0.35), with the highest being 0.57 (Cpd4 to prazosin).
Experimental Outcomes:
- Five of the six surrogates (prazosin, trimetozine, metacetamol, paroxypropione, tectoquinone) showed no significant inhibition ( $IC_{50} > 40,000$ nM).
- Levomefolic acid (matched to Cpd9) exhibited a non-monotonic concentration–response profile, with mean inhibition peaking at 56.6% at 20 µM and dropping to 12.8% at 40 µM. Consequently, the reported $IC_{50}$ of 9,038 nM was deemed an exploratory estimate rather than a reliable quantitative measure of potency.
Computational Analysis of Levomefolic Acid: MD simulations and MM–PBSA analysis suggested that while levomefolic acid formed a stable complex with favorable electrostatic contributions, its estimated binding energy ( $\Delta G_{bind} \approx -14.2$ kcal/mol) was significantly less favorable than the reference inhibitor sorafenib ( $\Delta G_{bind} \approx -35.6$ kcal/mol). The complex showed larger conformational fluctuations (RMSD) compared to sorafenib.

Significance and Claims
The paper positions this study primarily as a methodological exploration of a computational-to-experimental prioritization workflow rather than a successful discovery of a new VEGFR2 inhibitor.

Workflow Validation: The study demonstrates a practical strategy for connecting generative molecular design with the experimental assessment of accessible compounds via similarity matching. It highlights the feasibility of using generative AI to expand chemical space beyond traditional libraries.
Limitations of Similarity Matching: The results underscore the limitations of 2D similarity matching (ECFP4/Tanimoto) as a bridge to experimental validation. The low structural similarity between the generated candidates and their surrogates, combined with the lack of reproducible inhibitory activity in the surrogates, indicates that activity transfer cannot be assumed based on 2D fingerprints alone.
Modest Conclusions: The authors explicitly state that the study does not validate the VEGFR2 inhibitory activity of the generated molecules (Cpd4, Cpd6, Cpd7, Cpd9) nor does it confirm that the similarity-matching strategy successfully transferred activity. The experimental data for levomefolic acid were inconclusive due to non-monotonic behavior.
Future Directions: The authors conclude that future work must prioritize the direct synthesis and biological testing of the generated candidates, the use of 3D shape and pharmacophore similarity for matching, and the empirical calibration of the GCN model's applicability domain.

In summary, the paper offers a structured framework for integrating generative AI with experimental screening but concludes that significant methodological refinements and direct experimental validation are required before such workflows can reliably yield new therapeutic candidates.

Technical Summary: MolGPT-assisted exploration of VEGFR2-related chemical space

More like this