Variational AI & Life Chemicals join forces to discover selective dual EGFR/FGFR1 inhibitors using generative AI

The Resistance Challenge in Kinase-Driven Cancers

Therapeutic resistance remains a major limitation in kinase inhibitor therapy. In non-small cell lung cancer (NSCLC)¹, head and neck squamous cell carcinoma (HNSCC)², and cholangiocarcinoma³, EGFR (epidermal growth factor receptor) is frequently overactivated and has been successfully targeted clinically with multiple generations of moderately selective small-molecule Tyrosine Kinase Inhibitors (TKIs)⁴ (e.g., erlotinib, gefitinib, afatinib, osimertinib). However, compensatory upregulation of FGFR1 (fibroblast growth factor receptor 1) and the development of tyrosine kinase mutations are well-documented resistance mechanisms in both preclinical and clinical settings^5,6. 

Studies have demonstrated that dual blockade of EGFR and FGFR1 can overcome acquired resistance and yield synergistic anticancer effects. However, despite this evidence, structurally validated dual EGFR/FGFR1 inhibitors are rarely reported in the literature and often carry additional kinase inhibitory activity. The chemical space for dual kinase inhibitors is not well-explored, and existing examples often involve non-selective or pan-kinase scaffolds.

A Shared Problem Sparks a Collaborative Approach

Following their successful 2022 collaboration, which targeted the SARS-CoV-2 main protease (Mpro), Variational AI (VAI) and Life Chemicals (LC) have partnered for a second time, this time focusing on kinase selectivity.  Exploratory profiling of LC’s kinase-directed library against EGFR and aligned scientific interests led to a new joint program to discover de novo generated selective dual EGFR/FGFR1 inhibitors.

The strong biological and clinical rationale was faced with a technical challenge: public datasets contain an abundance of individual EGFR and FGFR1 inhibitor potency data.  However, there were only 90 dual-active molecules (IC₅₀ < 300 nM) encompassing 7 hinge binding scaffolds. Many of these dual EGFR/FGFR1 active compounds had higher potency for other kinases and these activities were picked up in selectivity screens. To our knowledge, no compounds with selective dual EGFR and FGFR1 inhibition are under development. This dataset of dual active compounds represents a very small amount of data to train on for our desired activity profile. Additionally, dual kinase targeting poses significant design challenges given the structural similarity overlaps across the kinase family and limited chemical diversity. For example, EGFR and FGFR1 share only 44% sequence identity within their kinase domains⁷, which limits the extent of pharmacophoric overlap.

Despite these constraints, the two teams were excited to take on this challenge.

Tackling Multi-Target Design with Generative AI

To address the low-data challenge, the project deployed Enki™, VAI’s generative AI proprietary platform, to identify novel compounds. Enki™ is a fully-trained foundation model built around a variational autoencoder, a type of generative model that learns a smooth chemical latent space in which structurally novel molecules can be optimized for multiple properties simultaneously, such as potency, selectivity, ADMET, synthesizability, and more…⁸ It generates structures inherently optimized to meet a defined target product profile.

The project goal was to design, synthesize, and test approximately 20 compounds for the desired profile. The Target Product Profile used for Enki™ was optimized for:

• Potency for both  EGFR and FGFR1⁹ targets (predicted IC₅₀ < 100 nM)
• Selectivity versus mitotic kinases: CDK1, AURKA (two kinases chosen to show initial representative selectivity)
• Novelty in chemical structure (no known analogs or Bemis-Murcko scaffolds¹⁰)
• Synthetic chemistry feasibility

A traditional method for generating hits would involve a high-throughput screen (HTS). However, using HTS to find dual active compounds is not typically done because the hit rates per individual target are already low and would be compounded when searching for dual activity without pan-kinase activity.  Furthermore, HTS doesn’t give novel starting points unless one has access to a novel chemical library.  As shown in this case study below, a single run of Enki for lead generation is approximately 5-10 times less expensive than a HTS on approximately 100,000 compounds and leads to novel lead generation for dual kinase active compounds with initial selectivity. 

Output Shows Dual Potency and Novelty

From a single run, Enki™ was able to generate a collection of structures, in which:

• 40 diverse molecules that fit the Target Product Profile (TPP) were designed
• 21 diverse compounds were selected by Life Chemicals for route development
• 19 (90%) were successfully synthesized within the timeframe of the project.

Of the 19 compounds synthesized, the majority exhibited meaningful biochemical activity:

• Active at either EGFR or FGFR1 kinase (IC₅₀ < 5 µM) in a biochemical assay¹¹: 15/19 compounds
• Dual EGFR + FGFR1 activity: 5/19 compounds

The Lead compound:

• EGFR IC₅₀ = 29 nM
• FGFR1 IC₅₀ = 175 nM

Selectivity:

• 100-fold over CDK1/cyclin B1
• ~10-fold over AURKA

Figure 1: The generic scaffold core
and R-group vectors

Docking-based interpretation of selectivity:

• FGFR1 & EGFR: Our top hit engages favorably with the hinge region in both kinases and fits well into the back hydrophobic pocket, supporting the strong observed activity.
• Aurora Kinase A: While our lead compound maintains hydrogen bonding with the hinge, steric constraints in the binding site appear to reduce the compound’s ability to access the back hydrophobic interactions. The interaction pattern is shifted toward more solvent-exposed residues, likely contributing to reduced potency and explaining the ~10-fold selectivity.
• CDK1: The compound fails to dock productively at the active site. The poor fit suggests a lack of stable binding interactions, consistent with the observed >100-fold selectivity.

Structural novelty:

• All molecules were absent from SciFinder-n (April 2025)
• Each featured a unique Bemis–Murcko scaffold
• Tanimoto similarity for each compound was ≤ 0.5 vs. known EGFR and FGFR1 inhibitors
• No prior dual EGFR/FGFR1 inhibitors shared the designed hinge-binding scaffold

Figure 2: The docking overlays of the blinded ligand in yellow to the EGFR and FGFR1 kinase binding site.

Conclusions and Path Forward

This collaborative effort demonstrates that generative AI can be used to design novel, synthesizable dual-active kinase inhibitors, even from a sparse training set. In just one generative cycle, multiple first-in-class scaffolds with dual EGFR/FGFR1 activity were identified, synthesized, and validated.

The current phase of the ongoing collaboration between Variational AI and Life Chemicals has focused on hit generation and chemical novelty.  In the future, there is an opportunity to continue the development and validation of the most promising compounds using Enki™’s lead optimization capabilities. 

The project remains active, and results continue to evolve, so stay tuned for further updates.

Interested parties are encouraged to reach out for discussion or potential collaboration.

Research conducted by the Variational AI team:  Marawan Ahmad, John Boylan, Marshall Drew-Brook, Peter Guzzo, Ahmad Issa, Mehran Khodabandeh, Sara Omar, Jason Rolfe, and Ali Saberali.

And by the Life Chemicals team: Olga Balabon, Vasily Pinchuk, Alexander Shivanyuk, Taras Tymoshenko and Iryna Vashchenko. 

We thank Amazon Web Services (AWS) for providing the cloud computing resources that made this research possible.