📄 medicine

Generative AI–enabled discovery of a CDK2-biased inhibitor for colorectal cancer

This study presents a generative AI-driven framework that successfully discovered and validated CDKCRC-EMBS, a potent and selective CDK2-biased inhibitor that induces cell-cycle arrest and apoptosis in colorectal cancer cells while sparing normal tissue.

Original authors: Anuraj Nayarisseri, Khushboo Sharma, Radhika Swami, Sajal Suhane, Smita Suhane, Swarna Dabral, Srinivas Bandaru, Rinku Chaudhary, V. Natchimuthu, Siddhi Shiva Kumar Thirumanwar, Leena Prajapati, Swati

Published 2026-07-23

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Anuraj Nayarisseri, Khushboo Sharma, Radhika Swami, Sajal Suhane, Smita Suhane, Swarna Dabral, Srinivas Bandaru, Rinku Chaudhary, V. Natchimuthu, Siddhi Shiva Kumar Thirumanwar, Leena Prajapati, Swati Smita Dhan, Taniya Dey, Habeebunnisa Dudekula, Prasannakumar Mohan, Kalaivani Selvam, Tajamul Hussain, Francisco Jaime Bezerra Mendonça-Junior, Marcus T Scotti, Luciana Scotti

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

Imagine your body as a bustling, high-tech city where trillions of cells are constantly being built, repaired, and retired. To keep this city running, there's a strict traffic control system that tells cells when to divide and when to stop. One of the most important traffic cops in this system is a protein called CDK2. Think of CDK2 as a master switch that flips the "Go" signal for cell division. In a healthy city, this switch is flipped only when needed. But in colorectal cancer, the switch gets stuck in the "On" position, causing cells to multiply uncontrollably and build chaotic, dangerous neighborhoods (tumors).

For decades, scientists have tried to build a "brake" to stop this runaway switch. The problem is that CDK2 looks almost identical to its neighbors in the protein family. It's like trying to handcuff one specific person in a crowd of twins wearing the exact same outfit; if you grab the wrong one, you might stop the whole city's traffic, causing massive side effects. Traditional methods of designing these brakes often rely on taking a static snapshot of the protein, which is like trying to catch a spinning top by looking at a single photograph. It misses the movement and the wobble that happen in real life. This is where a new kind of tool comes in: Artificial Intelligence. Instead of just looking at a photo, AI can learn the "dance moves" of the protein, predicting how it wiggles and shifts, allowing scientists to design a brake that fits perfectly into the moving target.

This is exactly what a team of researchers did in a new study to fight colorectal cancer. They combined the power of generative AI with real-world lab experiments to design a brand-new molecule from scratch. Instead of just searching through a library of existing chemicals, they taught an AI to imagine millions of new possibilities. The AI used a clever combination of three digital tools: a Variational Autoencoder (VAE) to explore new chemical shapes, a Generative Adversarial Network (GAN) to make sure those shapes looked like real, buildable molecules, and Reinforcement Learning (RL) to act as a coach, constantly telling the AI, "No, that one won't work, try this one that fits better."

The AI generated over 2.3 million unique molecular candidates and simulated how they would interact with the moving CDK2 protein. After filtering out the ones that didn't fit or were too unstable, the AI picked a champion: a molecule named CDKCRC-EMBS. The researchers then took this digital design and built it in their actual lab, synthesizing it through a five-step chemical process. When they tested it, the results were promising. The molecule acted like a precise key, locking into the CDK2 switch and stopping it from working. In lab tests, it inhibited the CDK2/cyclin E complex with an IC₅₀ of 1.92 ± 0.12 µM, meaning it took that specific amount of the drug to cut the enzyme's activity in half. It also showed a binding strength (Kᴅ) of 1.35 µM.

Crucially, the study suggests that this new molecule is "biased" toward CDK2. In the crowded protein family, it didn't get confused by the look-alike twins; it ignored them and focused only on the target. When tested on colorectal cancer cells, CDKCRC-EMBS successfully stopped the cells from dividing, forced them to pause their growth cycle, and even triggered them to self-destruct (a process called apoptosis). Perhaps most importantly, it did this without harming normal, healthy colon cells, suggesting it could be a safer option than older drugs that hit everything in sight.

The researchers didn't just stop at the computer screen. They validated their findings with rigorous experiments. They confirmed the molecule's structure using advanced tools like NMR and mass spectrometry. They ran 100-nanosecond-long computer simulations to watch how the molecule and protein danced together over time, confirming that the bond was stable. They also checked the molecule's "resume" (its ADMET profile) to ensure it could be absorbed by the body and wouldn't be toxic. While the paper confirms the molecule works in cells and test tubes, it stops short of claiming it is a cure for humans yet; it is a highly promising lead that has passed the first major hurdles of discovery. By listing the new compound in the public PubChem database (CID 9829487), the team has opened the door for other scientists to build upon this work, proving that when AI learns the rhythm of biology, it can help us compose a new song for fighting cancer.

Technical Summary: Generative AI–Enabled Discovery of a CDK2-Biased Inhibitor for Colorectal Cancer

Problem Statement
Colorectal cancer (CRC) remains a leading cause of global cancer mortality, with advanced cases often resistant to current therapies targeting EGFR and VEGF pathways. Cyclin-dependent kinase 2 (CDK2) is a critical regulator of the G1/S cell-cycle transition and a promising therapeutic target; however, developing selective CDK2 inhibitors is challenging due to the high structural homology of ATP-binding pockets across the CDK family. Traditional small-molecule inhibitors often exhibit broad kinase activity, leading to off-target toxicity. Furthermore, conventional computational drug discovery methods frequently rely on static structural models, failing to capture the conformational flexibility of drug-like molecules and the dynamic nature of protein–ligand interactions, which contributes to the high failure rate of computationally designed molecules during experimental validation.

Methodology
The authors present an integrated "AI-to-experiment" framework designed to discover de novo CDK2-biased inhibitors. The workflow combines generative artificial intelligence with dynamic structural screening and experimental validation:

Generative AI Framework:
- Architecture: A hybrid Variational Autoencoder–Generative Adversarial Network–Reinforcement Learning (VAE–GAN–RL) system was developed.
- Training Data: The models were trained on a curated set of 18 known CDK2 inhibitors (including 7-hydroxystaurosporine, dinaciclib, and roscovitine) and their conformational ensembles.
- Conformer Generation: A conditional GAN (ConfGAN) learned distance–energy relationships to generate energetically consistent 3D conformers, preserving chirality and pharmacophoric constraints.
- Molecule Generation: The VAE mapped molecular graphs into a latent space, while the GAN ensured chemical validity and novelty. A Reinforcement Learning (RL) module optimized the generated molecules using a multi-objective reward function ( $R = 0.4 \times QED + 0.3 \times SAnorm + 0.3 \times Dockproxy$ ), balancing drug-likeness, synthetic accessibility, and predicted binding affinity.
- Scale: The pipeline generated over 2.3 million unique, chemically diverse molecules.
Computational Screening & Triaging:
- Deep Learning Docking: Top candidates were screened using Gnina, a CNN-based docking tool, followed by energy minimization with CHARMM force fields.
- Dynamic Validation: Selected compounds underwent 100 ns molecular dynamics (MD) simulations using GROMACS to assess complex stability, RMSD, and hydrogen-bond persistence.
- Selectivity Profiling: An in-silico kinome-wide docking study was performed against 520 human kinases to evaluate selectivity.
Experimental Validation:
- Synthesis: The lead compound, CDKCRC-EMBS, was synthesized via a five-step route and characterized by NMR, high-resolution mass spectrometry, and HPLC.
- Biochemical Assays: Inhibition of CDK2/cyclin E was measured via ADP-Glo™ kinase assays and Surface Plasmon Resonance (SPR).
- Cellular Assays: The compound's efficacy was tested in colorectal cancer cell lines and normal colonic epithelial cells to assess cell-cycle arrest, Rb phosphorylation suppression, apoptosis induction, and cytotoxicity.

Key Results

Lead Identification: The AI framework identified CDKCRC-EMBS (PubChem CID: 9829487) as the top candidate. It demonstrated a predicted binding energy of –11.95 kcal mol⁻¹ and a Quantitative Estimate of Drug-likeness (QED) score of 0.83, outperforming the reference scaffold 7-hydroxystaurosporine.
Binding Mode: Docking and MD simulations revealed that CDKCRC-EMBS binds stably within the CDK2 ATP-pocket, forming canonical hinge-region hydrogen bonds with Glu81 and Leu83, and additional stabilizing interactions with Lys33 and Asp145. The quinazolinone core engages in $\pi$ – $\pi$ stacking with Phe80.
Selectivity: Kinome-wide profiling indicated strong bias toward CDK2. While closely related kinases (CDK1, CDK4, CDK7, CDK9) showed reduced binding, clinically relevant off-targets (e.g., EGFR, PI3K, AKT) exhibited significantly weaker predicted interactions.
Biochemical Potency: Experimental assays confirmed potent inhibition of CDK2/cyclin E with an IC₅₀ of 1.92 ± 0.12 µM and a dissociation constant (Kᴅ) of 1.35 µM.
Cellular Efficacy: In CRC cell models, CDKCRC-EMBS induced cell-cycle arrest, suppressed Rb phosphorylation, and promoted apoptosis. Crucially, the compound showed minimal toxicity toward normal colonic epithelial cells.
Ablation Studies: Comparative analysis demonstrated that the full VAE–GAN–RL pipeline, particularly the inclusion of conformational learning (ConfGAN) and RL optimization, yielded the highest enrichment of top-ranked, stable candidates compared to ablated variants.

Significance and Claims
The paper claims to establish a scalable, AI-driven discovery paradigm that successfully bridges the gap between in silico design and experimental reality by integrating conformational learning with dynamic screening. The authors assert that this approach overcomes the limitations of static structure-based screening, enabling the identification of novel, synthetically accessible, and biologically validated inhibitors. The study identifies CDKCRC-EMBS as a promising, CDK2-biased therapeutic candidate for colorectal cancer, demonstrating that generative AI can effectively navigate complex chemical spaces to produce molecules with favorable dynamic stability and selectivity profiles. The work highlights the potential of combining deep learning with molecular dynamics to prioritize compounds that are more likely to succeed in experimental validation.

Technical Summary: Generative AI–Enabled Discovery of a CDK2-Biased Inhibitor for Colorectal Cancer

More like this