AI-Driven Molecular Design and Docking of Ferroptosis Inhibitors for Parkinson’s Disease

ABSTRACT

Ferroptosis is a specific form of regulated cell death that occurs in the human brain. It is peroxidation on a molecular level that is dependent on iron abundance. Due to its neurodegenerative properties, ferroptosis plays a major role in the development of Parkinson’s disease. Current treatments focus on symptom suppression rather than addressing underlying molecular damage. This research develops a computational pipeline combining machine learning and molecular docking to identify ferroptosis inhibitors targeting the root cause of neuronal death. The researcher trained a machine learning model using Variational Autoencoders (VAEs) and Recurrent Neural Networks (RNNs) on 5,000-10,000 known ferroptosis inhibitors to generate novel molecular structures. Five lead compounds (Molecules 10, 18, 19, 21, and 22) demonstrated exceptional drug-like properties including zero Lipinski violations, bioavailability scores of 0.55-0.85, favorable blood-brain barrier penetration, and optimized lipophilicity and solubility profiles. Then these AI-generated molecules and screened ~2,647 FDA-approved drugs were examined through molecular docking against ferroptosis proteins 7P9V and 1SUV. Machine learning models achieved cross-validation R² scores exceeding 0.78, identifying 147 strong binders (≤-9.0 kcal/mol) from FDA libraries with 36 optimal candidates from 1SUV screening and 16 from 7P9V screening. Molecular complexity and accessible surface area emerged as key binding determinants. This integrated approach provides both immediate drug repurposing opportunities and supported novel scaffolds for future development, significantly accelerating drug discovery while offering effective treatment and prevention pathways for ferroptosis-related neurodegeneration.

INTRODUCTION.

Ferroptosis is a form of regulated cell death that occurs in the human brain, driven by iron-dependent lipid peroxidation at the molecular level. Ferroptosis plays a major role in the development of Parkinson’s disease by programming the cell death of dopaminergic motor neurons. Many research studies indicate that about two-thirds of patients with Parkinson’s disease show significant signs of ferroptosis, including iron accumulation in the substantia nigra, increased lipid peroxidation, glutathione depletion, and mitochondrial dysfunction [1]. Overall, these findings highlight ferroptosis as a critical therapeutic target, encouraging this computational approach to identifying novel inhibitors.

A machine learning model combining Variational Autoencoders (VAEs) and Recurrent Neural Networks (RNNs) was trained on 5,000-10,000 known ferroptosis inhibitors and generated 89 novel molecular structures. From these, five lead compounds (Molecules 10, 18, 19, 21, and 22) demonstrated optimal drug-like properties including zero Lipinski violations, favorable blood-brain barrier penetration potential, and bioavailability scores of 0.55-0.85. However, these computationally-generated molecules required experimental validation through binding studies to confirm their therapeutic potential.

We build on this ML foundation and target two critical ferroptosis proteins: SLC7A11 (PDB ID: 7P9V), which regulates cellular iron metabolism and homeostasis, and TFRC (PDB ID: 1SUV), which directly executes the ferroptosis cell death pathway. Using these targets, molecular docking was performed for various FDA-approved drugs and all of the drugs generated through the ML model [2, 3]. This would help us in identifying if the molecules obtained from the ML model have good binding properties with the target proteins and what particular chemical properties are essential for binding with the respective proteins of interest. These proteins represent strategic therapeutic targets in Parkinson’s disease, particularly affecting dopaminergic neurons in the substantia nigra [4]. Traditional drug discovery processes require 10-15 years and billions of dollars to bring a single compound to market, creating an urgent need for accelerated computational approaches that our work addresses [5]. However, by integrating machine learning based drug generation with molecular docking, this pipeline significantly narrows the search space for viable Parkinson’s disease therapeutics. This offers a faster and more cost-effective pathway from computational prediction to potential clinical application.

Here we test if a machine learning model trained on known ferroptosis inhibitors can generate novel molecular structures with most effective combinations for treating ferroptosis and Parkinson’s disease. We support these findings by testing the docking behavior of specific generated molecules, the ones satisfying certain criteria such as bioavailability and the potential to cross the blood brain barrier, with the aforementioned proteins.

Training a machine learning model involves utilizing a multitude of ferroptosis inhibitors to generate the most effective novel ferroptosis and Parkinson’s treatments. Targeting ferroptosis provides an alternative treatment option to the status quo treatment options of symptom management and suppression that do not address the underlying causes of the neural damage on the molecular level. This research also advances current knowledge on neurodegenerative diseases and the relations between ferroptosis and Parkinson’s disease. This novel treatment addresses the underlying causes of the neural damage on the molecular level, increasing the long-term efficiency and effectiveness of the treatment in completely treating and preventing the damage.

MATERIALS AND METHODS.

Materials and Computing Specifications.

No consumable materials were used throughout the course of the experiment. All computational work was performed on a Dell Intel Core 7 Desktop Computer equipped with GPU: NVIDIA GeForce MX570 A, CPU: Intel Core 7 150U, and Ram Capacity: 16.0 GB.

Machine Learning-Based Molecular Generation.

Proper safety techniques including proper posture and eye strain prevention techniques were practiced while performing this experiment. A main project directory was created in Visual Studio Code with subdirectories for models, data, and results. Utility modules were developed for SMILES conversions and encoding. The VAE model was initialized with batch size (128), unit size (512), RNN layers (3), and sequence length (120).

Training data was split 75/25 with batch generation and loss calculation. The training process loaded SMILES strings, converted them to one-hot encoding, initialized model parameters, and ran training loops. Generated molecules underwent optimization through parameter tuning and validation testing. For data analysis, SMILES strings were input into SwissADME to generate structural images, calculate properties, and create egg plot models for evaluation of drug-likeness criteria [6].

Molecular Docking.

Protein crystal structures for 1SUV and 7P9V were obtained from the Protein Data Bank and cleaned by removing water molecules and non-essential heteroatoms and prepared with hydrogen atom addition. Active binding sites were identified through literature review, with grid boxes (20×20×20 Ångstroms) centered on binding sites and exhaustiveness level set to 8 [6]. Two compound libraries were assembled: 2,647 FDA-approved drugs from DrugBank and 89 AI-generated molecules. Virtual screening processed each compound against both protein targets using AutoDock Vina [7], calculating binding affinities in kcal/mol and recording optimal binding poses, resulting in 2,808 total docking simulations. Safety protocols included 20-second eye rest breaks every 20 minutes, proper ergonomic posture, and full breaks every hour. Comprehensive molecular descriptors (87 properties including molecular weight, surface area, complexity, LogP, and topological indices) were computed using RDKit. ADMET properties were predicted computationally, including blood-brain barrier penetration and aqueous solubility. Drug-likeness was assessed using Lipinski’s Rule of Five, Veber rules, and QED scores (threshold QED > 0.5). Structure-activity relationship analysis calculated Pearson correlation coefficients between molecular properties and binding affinities with statistical significance testing (p < 0.05), identified activity cliffs by finding structurally similar compound pairs with large binding differences, and performed molecular scaffold analysis to identify privileged structures. Machine learning validation employed Random Forest regression models with rigorous 5-fold cross-validation, using 100 decision trees with maximum depth of 10 levels. Performance metrics included R² and mean absolute error (MAE). Comparative analysis between FDA-approved drugs and AI-generated molecules employed statistical t-tests to compare predicted binding affinity distributions and identify dual-target inhibitors showing strong binding to both proteins. Final candidate ranking utilized composite scoring weighted 50% binding affinity, 30% drug-likeness, and 20% ADMET favorability.

RESULTS.

AI-Generated Molecular Structures.

The machine learning algorithm successfully generated 89 novel molecular structures from training with known ferroptosis inhibitors. Systematic filtering based on drug-likeness criteria identified five lead compounds with exceptional pharmaceutical properties representing the most promising candidates for further development (Figure 1).

The five lead compounds demonstrated remarkable consistency in meeting the drug-likeness criteria (Table 1). All molecules achieved zero Lipinski violations, confirming optimal oral bioavailability potential. Zero Ghose values indicated significant absorption and membrane activity profiles, while zero Veber violations confirmed optimal polarity and flexibility for drug formulation. Zero Egan violations demonstrated favorable blood-brain barrier penetration potential (LogP <5.88, PSA <131Å), critical for central nervous system drug delivery. Muegge scores remained below 5 across all candidates, confirming structural and pharmacokinetic optimization. Bioavailability scores ranged from 0.55 to 0.85, with Molecule 19 achieving the highest score of 0.85.

Table 1. Most effective predicted novel molecular structures. SMILES Codes provide the correct structural representation of the molecule in a string of text. The molecular formula shows each molecules elemental composition and the number of each type of elements. Lipophilicity (ILogP) ensures bioavailability in the environment with an optimal range of 2-5. Synthetic accessibility shows the molecules measured based on ease of facility production for drug usage optimally ≤ 3. ESOL solubility shows the optimal solubility range to improve absorption potential, and should be between 0.1-100 mg/ml.
Molecule	SMILES Codes	Formula	Lipophilicity (iLOGP)	Synthetic Accessibility	ESOL Solubility (mg/ml)
10	C#CCC(C(C)(C)C)CC(=O)c1ccccc1	C16H20O	3.01	2.65	0.0388
18	CC(NC(C(=O)O)C)Cc1cccnc1	C11H16N2O2	1.68	2.46	90.9
19	Cc1cccc(c1)N(C(=O)O)C	C9H11NO2	1.69	1.29	1.08
21	CC(CNC(=O)c1ccccc1)N	C10H14N2O	1.81	1.42	9.33
22	O=C(C1OC1Cc1ccccc1)c1ccccc1	C16H14O2	2.46	2.66	0.0548

The egg plot model in Figure 2 confirmed that all five molecules occupied favorable regions for human intestinal absorption (HIA) and blood-brain barrier (BBB) permeability, with appropriate P-glycoprotein substrate profiles i.e. PGP-. PGP- status is optimal as P-glycoprotein actively pump out PGP+ substrates from the brain cells. Having compounds of the PGP- category would ensure a higher concentration of them accumulating in the brain cells.

Clear structure-property relationships emerged across the molecular set (Figure 3). All molecules displayed moderate lipophilicity with Molecule 10 exhibiting the highest values. An inverse relationship between polarity and lipophilicity was evident, where higher LIPO values consistently corresponded to lower POLAR values. The FLEX parameter remained relatively consistent (0.2-0.4) across all molecules, indicating similar structural integrity. Most maintained moderate size parameters (0.5-0.6) with Molecule 21 as an outlier at 0.7. Consistently low polarity and insolubility values suggested optimal membrane permeability. Molecule 19 showed unique high unsaturation (0.8), while Molecule 18 demonstrated the most balanced profile across all parameters.

Molecular Docking and Binding Analysis.

Enhanced screening processed 2,808 total docking simulations across four configurations: 1SUV-FDA (986 compounds), 1SUV-AI (77 compounds), 7P9V-FDA (1,669 compounds), and 7P9V-AI (76 compounds). Binding affinity distributions revealed strong binders in both FDA-approved and AI-generated libraries for both protein targets (Figure 4).

Binding Affinity and Drug-Likeness Summary Statistics.

Statistical analysis revealed FDA approved drugs substantially outperformed AI generated molecules in binding strength. For 1SUV, FDA-approved drugs identified 83 strong binders (mean -6.79 ± 1.76 kcal/mol, best -11.13 kcal/mol) compared to zero strong binders for AI molecules (mean -6.68 ± 0.89 kcal/mol, best -8.45 kcal/mol). For 7P9V, FDA drugs identified 64 strong binders (mean -6.39 ± 1.56 kcal/mol, best -10.65 kcal/mol) versus zero for AI molecules (mean -6.17 ± 1.00 kcal/mol, best -8.32 kcal/mol). Conversely, AI generated molecules demonstrated superior mean QED scores (0.619-0.625 vs. 0.558-0.565), indicating successful drug-likeness optimization. Filtering for optimal candidates meeting binding affinity (≤-9.0 kcal/mol), drug-likeness (QED > 0.55), and ADMET criteria yielded 36 optimal FDA candidates from 1SUV screening and 16 from 7P9V screening, compared to zero AI candidates meeting the binding affinity threshold.

DISCUSSION.

This investigation successfully combined machine learning molecular generation with molecular docking to identify potential ferroptosis inhibitors for Parkinson’s disease. The VAE-RNN model generated 89 chemically valid molecules with zero violations across drug-likeness criteria, with five lead compounds demonstrating exceptional pharmaceutical properties. The diversity in physicochemical properties (lipophilicity 1.68-3.01, solubility 0.0388-90.9 mg/mL) enables patient-specific optimization.

Molecular docking revealed FDA-approved drugs substantially outperformed AI molecules in binding strength (83 and 64 strong binders versus zero), reflecting decades of pharmaceutical optimization. However, AI molecules showed superior drug-likeness (QED 0.619-0.625 vs. 0.558-0.565), suggesting complementary advantages: FDA drugs represent immediate clinical translation (3-5 years versus typical 10-15 years), while AI molecules provide novel scaffolds for future development. Molecular complexity and accessible surface area emerged as key binding determinants; yet complexity alone proved insufficient, indicating additional structural features (hydrogen bonding, aromatic positioning) drive superior binding (Figure 5). Excellent predictive accuracy (R² 0.78-0.96) shows potential for this machine learning for prospective drug design (Figure 6). AI-generated molecules clustered around 250–350 Da while FDA-approved drugs showed broader molecular weight diversity (Figure 7).

Dual-target inhibitors binding both 7P9V and 1SUV offer redundant pathway blocking, reducing required doses and off-target effects. This integrated framework identified 147 strong binders from FDA libraries and AI generated molecules with optimal drug-like properties, significantly accelerating the drug discovery timeline while reducing costs. FDA-approved drug candidates can advance directly toward clinical trials, while AI-generated molecules serve as starting points for next-generation neuroprotective drug design.

Limitations in this work include static docking predictions lacking dynamic protein flexibility and computational ADMET predictions do not represent human pharmacokinetics. Most critically, results require experimental validation through in vitro and in vivo studies. Limited training data may bias the AI model toward known chemical classes.

This work establishes a reproducible computational framework applicable to other therapeutic targets and disease contexts. The integrated strategy of AI-driven generation combined with rigorous docking validation demonstrates how computational approaches can provide practical pathways for accelerated pharmaceutical development, addressing urgent unmet medical needs in neurodegeneration.

ACKNOWLEDGMENTS.

The author thanks The Benjamin School for providing computational resources and research support.

REFERENCES.

1. K.-J. Lin et al., Iron brain menace: The involvement of ferroptosis in Parkinson disease. Cells 11, 3829 (2022).
2. J. L. Parker et al., Molecular basis for redox control by the human cystine/glutamate antiporter system xc⁻. Nat. Commun. 12, 7147 (2021).
3. A. Mahammedi, M. A. Baraibar, M. Rederstorff, Ferroptosis in Parkinson’s disease: Glia–neuron crosstalk. Trends Mol. Med. 28, 258–269 (2022).
4. B. Do Van et al., Ferroptosis, a newly characterized form of cell death in Parkinson’s disease that is regulated by PKC. Neurobiol. Dis. 94, 169–178 (2016).
5. Z. Xu et al., A generative AI-discovered TNIK inhibitor for idiopathic pulmonary fibrosis: a randomized phase 2a trial. Nat. Med. 31, 2602–2610 (2025).
6. M. H. S. Segler, T. Kogej, C. Tyrchan, M. P. Waller, Generating focused molecule libraries for drug discovery with recurrent neural networks. arXiv 1701.01329 (2017).
7. O. Trott, A. J. Olson, AutoDock Vina: Improving the speed and accuracy of docking with a new scoring function, efficient optimization, and multithreading. J. Comput. Chem. 31, 455–461 (2010).

---

---

Posted by buchanle on Monday, June 1, 2026 in May 2026.

Tags: Drug Discovery, Ferroptosis, Machine learning, Molecular Docking, Parkinson’s disease