🧬 biology

Integrated Machine Learning and Structure-Based Computational Prioritization of KIF11 Inhibitor Candidates for CNS-Oriented Drug Discovery

This study presents an integrated machine learning and structure-based computational workflow that successfully prioritized five structurally diverse KIF11 inhibitor candidates with high predicted potency and blood–brain barrier permeability for use in CNS-oriented oncology drug discovery.

Original authors: Sourajyoti Goswami, Puspita Roy, Shouvik Kumar Nandy

Published 2026-09-29

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Sourajyoti Goswami, Puspita Roy, Shouvik Kumar Nandy

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

Cancer is a disease of uncontrolled growth, where cells divide without stopping. To halt this runaway process, scientists often look for ways to disrupt the machinery that allows cells to split in two. One critical piece of this machinery is a protein called KIF11, which acts like a molecular motor to organize the internal scaffolding needed for cell division. When this protein is blocked, the cell cannot divide properly and eventually dies. This makes KIF11 a promising target for new cancer drugs. However, treating cancers that originate in the brain, such as glioblastoma, presents a unique challenge. The brain is protected by a biological barrier that keeps most substances out, meaning a drug must be specifically designed to cross this wall to reach the tumor. Finding a molecule that is both potent enough to stop the cancer and small enough to enter the brain is a difficult balancing act for researchers.

In a recent study, a team of scientists from India developed a new computer-based strategy to find potential drugs that could solve this problem. Instead of testing thousands of chemicals in a lab one by one, which is slow and expensive, they used advanced computer programs to simulate how different molecules might interact with the KIF11 protein. The researchers started by gathering a massive collection of existing data on how various chemicals affect KIF11, cleaning the information to ensure it was accurate and consistent. They then taught a computer system to recognize patterns in the shapes and chemical properties of these molecules, allowing the system to predict which new, untested chemicals might be effective inhibitors. To ensure their predictions were reliable, they used a rigorous method where the computer learned from one group of chemicals and was tested on a completely separate group it had never seen before, preventing the system from simply memorizing the answers.

The computer model successfully identified several promising candidates, but the researchers did not stop there. They wanted to know if these molecules could actually fit into the specific pocket on the KIF11 protein where they need to bind to do their work. Using a technique called molecular docking, they simulated the physical interaction between the top candidates and the protein structure. This process is like trying to fit a key into a lock to see if the shape matches, but done entirely within a computer. The simulations showed that the selected molecules could indeed settle into the correct spot on the protein. The team also ran checks to see if these molecules would be able to cross the blood-brain barrier and reach the brain, as well as whether they would be safe for the body to handle. They found that the top candidates had a very high probability of entering the brain, which is a crucial requirement for treating brain tumors.

Among the molecules identified, one stood out as the most promising. This compound, known by the code CHEMBL392148, was predicted to be highly effective at stopping the KIF11 protein and had an excellent chance of crossing into the brain. The researchers also looked closely at the chemical structure of this molecule and others to ensure they did not contain parts known to be toxic or unstable. While one of the candidates showed a small warning sign regarding a specific chemical feature, the top choice appeared clean and safe based on the computer models. The study also highlighted a potential safety concern: all five of the best candidates showed a signal that they might affect the heart's electrical system, a risk that would need to be tested in a real lab before any human use.

The researchers emphasize that these findings are a starting point, not a finished product. The molecules they identified are still just computer predictions. The next step would be to synthesize these chemicals in a lab and test them on actual proteins and cells to see if they behave as the computer predicted. The study provides a clear, rational path forward, offering a shortlist of specific chemical structures that are worth investigating further. By combining machine learning with detailed structural analysis, the team has narrowed down the search for a new brain-cancer drug to a handful of high-quality candidates, saving time and resources while focusing efforts on the most likely solutions.

Technical Summary: Integrated Machine Learning and Structure-Based Computational Prioritization of KIF11 Inhibitor Candidates for CNS-Oriented Drug Discovery

Problem Statement
Glioblastoma (GBM) remains a significant challenge in oncology due to its aggressive nature, therapeutic resistance, and limited treatment efficacy. While Kinesin Family Member 11 (KIF11/Eg5) is a validated target for antimitotic therapy by disrupting bipolar spindle assembly, identifying new inhibitors capable of crossing the blood-brain barrier (BBB) for CNS-oriented drug discovery requires rigorous computational strategies. The study addresses the need to reduce the chemical space for experimental testing by integrating machine learning (ML) with structure-based methods to prioritize KIF11 inhibitors with predicted CNS permeability.

Methodology
The study employed a systematic, multi-stage computational workflow:

Data Curation and Partitioning:
- A dataset of 1,817 KIF11 activity records was retrieved from ChEMBL (target CHEMBL4581).
- Strict curation criteria were applied: retention of exact IC50 measurements, removal of "outside typical range" warnings, exclusion of mutant-specific assays, and removal of cell-based assays to ensure a target-level biochemical endpoint.
- The final curated dataset comprised 831 unique compounds (pIC50 range: 4.046–9.854).
- The dataset was partitioned into a training set (654 compounds) and a locked external test set (177 compounds) before model development. Crucially, the partition was scaffold-disjoint (Bemis–Murcko scaffolds), ensuring no shared scaffolds between training and testing sets to prevent chemical-series leakage.
QSAR Modeling and Validation:
- Descriptors: Eleven physicochemical descriptors and 2048-bit ECFP4 (Extended Connectivity Fingerprint 4) molecular fingerprints were generated.
- Algorithms: Random Forest (RF), Extra Trees (ET), and Histogram-based Gradient Boosting (HGB) were evaluated using hybrid, descriptor-only, and fingerprint-only representations.
- Validation Strategy:
  - Scaffold-aware cross-validation: Ensured compounds sharing a scaffold remained within the same fold.
  - Nested scaffold cross-validation: Separated model selection from performance estimation to provide a conservative internal estimate.
  - Y-randomization: Performed 10 iterations to rule out chance correlations.
  - Applicability Domain (AD): Defined using Morgan ECFP4 Tanimoto similarity (threshold: 0.5701) to assess chemical-space coverage.
  - Locked External Validation: The final frozen model was applied to the 177 independent compounds without further optimization.
Virtual Screening and Structure-Based Analysis:
- The QSAR model predicted pIC50 values for the external set.
- A diversity-based selection (greedy algorithm, max Tanimoto similarity < 0.70) identified 20 high-potential candidates.
- Molecular Docking: Performed using AutoDock Vina 1.2.7 against human KIF11 (PDB ID: 2WOG). The protocol was validated by redocking the co-crystallized ligand STLC. While exact crystallographic pose reproduction (<2 Å RMSD) was not achieved due to ligand symmetry and sampling limits, the docking successfully localized ligands within the experimentally defined binding pocket.
- Interaction Profiling: Residues within 5 Å of docked poses were analyzed to identify common interaction patterns.
ADMET and Drug-Likeness Profiling:
- Five top candidates underwent comprehensive ADMET prediction using ADMET-AI, focusing on BBB penetration, bioavailability, P-glycoprotein activity, CYP450 interactions, and toxicity endpoints (hERG, AMES, DILI, etc.).
- Structural alert screening (PAINS, BRENK, NIH) and drug-likeness metrics (QED, Lipinski rules) were applied.
- A preliminary 1 ns solvent-relaxation MD simulation was conducted for a representative complex (KIF11–CHEMBL195232) to assess structural stability, reported separately in Supplementary Information.

Key Results

Model Performance:
- Hybrid models (combining descriptors and ECFP4) outperformed descriptor-only models. The hybrid Extra Trees model achieved the best baseline performance ( $R^2$ = 0.5985).
- Nested Scaffold Cross-Validation: The rigorous internal validation yielded an $R^2$ of 0.5471, RMSE of 0.9021, and MAE of 0.7040.
- Locked External Validation: The independent test set yielded an $R^2$ of 0.2671, RMSE of 0.8754, and MAE of 0.7384. The lower external $R^2$ reflects the challenge of predicting activity for chemically distinct scaffolds and the fact that only 48% of external compounds fell within the predefined applicability domain.
- Y-randomization: Randomized models produced negative $R^2$ values (mean -0.3593), confirming the model captures genuine structure-activity relationships rather than chance correlations.
Prioritized Candidates:
Five structurally diverse compounds were prioritized based on an integrated score combining QSAR predictions, docking scores, AD similarity, and ADMET profiles:
1. CHEMBL392148: Highest predicted potency (pIC50 = 7.75; IC50 ≈ 17.77 nM), highest BBB probability (0.98), and no structural alerts.
2. CHEMBL1086410: Most favorable docking score (-9.882 kcal/mol) and high BBB probability (0.92).
3. CHEMBL383402: Highest AD similarity (0.80) but contained one BRENK alert (aliphatic long-chain).
4. CHEMBL400042: Highest QED (0.87) and balanced physicochemical profile.
5. CHEMBL1085666: High bioavailability prediction and favorable drug-likeness.
Docking Interactions:
All prioritized compounds occupied the STLC-binding pocket. Recurrent interactions were observed with residues Arg119, Glu116, Tyr211, Leu214, Ala218, and Arg221.
ADMET Findings:
All five candidates showed high predicted BBB penetration probabilities (0.91–0.98) and oral bioavailability. However, a consistent computational signal was observed for hERG liability across all five candidates (probabilities 0.71–0.96), indicating a potential cardiac safety risk requiring experimental verification.

Significance and Claims
The paper claims to provide a rational computational starting point for the discovery of KIF11 inhibitors specifically tailored for CNS oncology. The study emphasizes that the identified compounds are computationally prioritized candidates, not experimentally confirmed inhibitors.

Key contributions include:

The development of a robust, scaffold-aware QSAR workflow that minimizes chemical-series leakage and provides a stringent external validation framework.
The integration of ligand-based (QSAR) and structure-based (docking) approaches with ADMET profiling to filter for CNS permeability.
The identification of five specific chemical entities (led by CHEMBL392148) that satisfy multiple computational criteria for KIF11 inhibition and BBB penetration.

The authors explicitly state that the study does not establish experimental inhibitory activity or safety. The predicted hERG liability and the limitations of the docking protocol (inability to reproduce exact crystallographic poses) are highlighted as critical areas requiring experimental validation. The work serves as a hypothesis-generating framework for subsequent biochemical inhibition assays, cellular evaluation in GBM models, and pharmacokinetic studies.

More like this