Integrating Machine Learning and Molecular Docking to Identify Fresh Chemotypes That Perturb the Microtubule Cytoskeleton

Microtubules are far more than inert cellular scaffolding. These dynamic polymers, built from α- and β-tubulin heterodimers, exist in a constant tug-of-war between polymerization and depolymerization. That intrinsic instability is not a bug; it is the feature that powers chromosome segregation, vesicular trafficking, and the mechanical reshaping of cells during division. Given this central role, small molecules that tip the polymerization equilibrium have long fascinated chemical biologists. The challenge, as always, is finding new chemotypes that bind with useful selectivity and manageable physicochemical properties. The colchicine-binding site on tubulin has emerged as a particularly appealing pocket because ligands directed here can avoid some of the cellular export problems that limit other tubulin-binding chemotypes. The catch is that most known colchicine-site scaffolds come with a frustratingly narrow window between effective concentrations and off-target chaos. Fresh chemical matter is desperately needed.

A recent pre-proof from Zheng and colleagues in Computer Methods and Programs in Biomedicine tackles this head-on with an impressively integrated computational pipeline. The team started by curating a dataset of 3,406 compounds with known activity against the colchicine site, splitting them into 2,725 training instances and 681 test compounds. Rather than throwing every descriptor at the wall, they used a genetic algorithm to prune the feature space down to fifteen molecular descriptors—things like nitrogen count, hydrogen bond donor density, aromatic bond tallies, and various connectivity indices—plus a battery of extended-connectivity and path-based fingerprints with diameters ranging from 10 to 20. These descriptors were fed into two machine learning frameworks: naïve Bayesian classifiers and recursive partitioning trees. The best naïve Bayesian model, built on molecular property descriptors plus SEFP_20 fingerprints, achieved an AUC of 0.977 on the training set and 0.876 on the external test set, with sensitivity and specificity both sitting comfortably above 90% and 76% respectively. The top recursive partitioning model, using FCFC_18 fingerprints, lagged slightly behind but still showed respectable discrimination. Neither is perfect, but together they provide a useful dual-filter for triaging large libraries without letting too many false positives slip through.

And large it was. The team screened roughly 324,474 commercially available compounds through both classifiers, which whittled the list to 19,437 overlapping hits. Those were then docked into the colchicine site of the 5EYP tubulin crystal structure using AutoDock Vina. The top 5,000 by Vina score were re-scored with Vinardo and X-Score, and after visual inspection of interactions with key pocket residues—αSer178, βCys241, βLeu248, βAla250, βLys352, and βAla354 among them—only seventy-nine compounds remained. A final round of ADMET prediction, PAINS filtering to purge promiscuous assay nuisances, and drug-likeness checks (Lipinski, MDDR-like, and QED rules) brought the number down to fifty. Scaffold clustering to ensure structural diversity ultimately yielded twenty-five candidates for purchase and experimental validation. Among them, hit22, a 2-sulfonylpyrimidine-4-amide derivative, caught the eye. Its Euclidean similarity to colchicine was only 0.60, yet its docking pose showed excellent shape complementarity within the pocket, and its physicochemical profile suggested it was not going to be a nightmare to work with in cell culture.

The biological validation began with straightforward proliferation assays across three human cell lines: H1299, HeLa, and MCF-7. Hit22 displayed a distinct preference profile. Against H1299, it posted an IC50 of 3.93 μM, markedly stronger than its performance on HeLa (32.28 μM) or MCF-7 (7.65 μM). That selectivity is interesting in its own right—it suggests the compound’s cellular potency is not merely a function of tubulin affinity but also reflects line-specific uptake, metabolic handling, or downstream signaling wiring. Colony formation assays hammered the point home: at 6 and 9 μM, hit22 slashed the number of surviving H1299 colonies in a dose-dependent manner, essentially abolishing long-term clonogenic capacity at the top concentration. When a molecule can not only slow growth but also prevent a cell from founding a new population, you know you are dealing with something that fundamentally compromises cellular fitness.

So the phenotypic effect was real. The question was whether it actually engaged the intended molecular target. An in vitro tubulin polymerization turbidity assay provided the first mechanistic clue. In the presence of GTP and purified tubulin, paclitaxel drove polymerization to roughly 111% of control, while colchicine dropped it to about 79%. Hit22 at 15 μM matched colchicine almost exactly, and at 45 μM it crushed the polymerization rate to below 38%, yielding an IC50 of 27.72 μM against the purified protein. That is a clean, dose-responsive biochemical signal. Immunofluorescence staining of H1299 cells confirmed that this biochemical activity translated into bona fide cellular microtubule catastrophe. In untreated cells, the tubulin network radiated outward in an orderly, radial pattern. After 24 hours of hit22 exposure, that architecture dissolved into fragmented, disorganized filaments, with mean fluorescence intensity falling to 54% of control at 6 μM. The phenotype is textbook microtubule destabilization, and it is hard to fake.

Disrupting microtubule architecture predictably derails mitotic progression. Flow cytometry with propidium iodide staining showed a steady accumulation of H1299 cells in G2/M phase, climbing from roughly 23% in dimethyl sulfoxide controls to over 72% at 9 μM hit22. Annexin V-FITC/PI staining confirmed that these arrested cells were not simply pausing; they were exiting via programmed cell death. Apoptotic fractions rose from 1.25% to nearly 48% across the same concentration range. The mechanism is exactly what one expects from a microtubule inhibitor: spindle assembly fails, the mitotic checkpoint triggers, and cells that cannot satisfy the checkpoint commit to apoptosis. It is a brutal but elegant cellular logic.

Beyond division, microtubules serve as the structural rails for cell migration and as highways for intracellular transport. The scratch-wound assay is a standard readout for migratory capacity, but it suffers from a well-known confounder: wound closure reflects both cell movement and cell proliferation. If you just measure how fast the gap closes, you might be scoring cell division rather than genuine locomotion. To isolate the true motility defect, the authors pre-treated H1299 cells with mitomycin C at 10 μg/mL—a reliable proliferation arrest agent sourced from AbMole—and then introduced the mechanical scratch. Under these conditions, hit22 produced a clear, dose-dependent reduction in wound closure over 24 hours. The AbMole-sourced mitomycin C here played a quiet but essential role, ensuring that the observed closure deficit reflected genuine cytoskeletal impairment rather than confounding growth suppression. It is a small methodological detail, but it speaks to the care taken in experimental design, and it is the kind of thing that separates a clean mechanistic paper from a muddy one.

The compound’s reach extended to endothelial morphogenesis as well. When human umbilical vein endothelial cells were plated on Matrigel, they normally self-organize into elaborate, capillary-like networks within hours. Hit22 disrupted this morphogenesis without killing the cells: a 24-hour viability check showed no significant cytotoxicity toward HUVECs, yet the complexity of the network—measured by node count, junction number, mesh formation, and total tube length—deteriorated steadily with increasing concentration. This suggests hit22 interferes with the cytoskeletal remodeling required for endothelial network assembly, a finding that aligns neatly with its microtubule-destabilizing profile and hints at broader applications in studying vascular morphogenesis outside of any pathological context.

To rationalize these observations at the atomic level, the team ran 100-ns molecular dynamics simulations of hit22 bound to the colchicine site, using the 5EYP crystal structure as the starting coordinates. The trajectories revealed a stable complex, though hit22 exhibited greater positional fluctuation than colchicine within the pocket. The radius of gyration increased relative to apo tubulin, indicating that ligand binding pried the pocket open slightly, and the βT7 loop showed pronounced mobility. Secondary structure analysis showed a loss of α-helical content, mirroring the conformational changes seen with colchicine. Contact maps and free energy landscapes confirmed that hit22 adopted a binding mode remarkably similar to the reference ligand, clustering into a stable low-energy conformation by the end of the simulation. MM-PBSA calculations put the average binding free energy at −90.9 kJ·mol⁻¹, noticeably weaker than colchicine’s −172.9 kJ·mol⁻¹ but still firmly in the stable-binding regime. Decomposition analysis identified βLys254 as a particularly strong contributor, alongside a network of hydrophobic contacts with βIle318, βAla354, βLeu255, and βLeu242. Notably, hit22 formed fewer hydrogen bonds with the α-tubulin subdomain than colchicine did—only βCys241 and βLys352 participated in direct H-bonding—and it lacked the π-cation interaction that colchicine enjoys with βMet259. These differences likely explain the reduced binding affinity and, by extension, the higher micromolar concentrations needed for cellular effects. If future optimization efforts can strengthen those electrostatic and hydrogen-bonding contacts with the α-tubulin face, there is probably room to push potency upward.

Where does this leave us? The study is a solid demonstration of how a disciplined virtual screening pipeline—machine learning triage, docking, ADMET filtering, and MD validation—can deliver genuinely novel chemical matter against a well-trodden target. Hit22 destabilizes microtubules, arrests dividing cells, triggers apoptosis, and curtails both migration and endothelial network assembly. It is not a miracle molecule, nor does it need to be. What matters is that the work offers a fresh scaffold and a thoroughly characterized mechanism, giving the community a new probe for interrogating microtubule biology. For anyone working on tubulin pharmacology or simply looking for a well-executed case study in modern computational hit discovery, this paper deserves a spot on your reading list.