Скачать или смотреть VeGA: A Transformer-Based Tool for Generative Chemistry

Today, we’re going to have a quick look at VeGA (Versatile Generative Architecture), a Transformer-based generative model for molecular design. It enables efficient fine-tuning and de novo generation of bioactive compounds.

VeGA was developed by Pietro Delre and Antonio Lavecchia and published in J. Chem. Inf. Model., September 2025. The article is titled: "VeGA: A Versatile Generative Architecture for Bioactive Molecules across Multiple Therapeutic Targets."

🧪 In this video, we’ll walk through a practical example using Factor X, a key target in blood coagulation. I’ve skipped the installation process, as it’s clearly explained on the project’s GitHub page. Here’s the plan:

a) Fetch bioactive compounds from ChEMBL, easily done using DataWarrior (Check out my other tutorials on how to use it, here on the channel) but you can of course use a Python script for that. Then you may need a small data curation step, for a real project you need a real data curation step, it can be removing salt, standardize the molecules while here, as they all come from ChEMBL this is done already....

b) Fine-tune the VeGA foundation model using these compounds, like you select 100 or so that have good affinity to the target. Compounds need to prepare following the process/data preparation/curation reported in the publication to be consistent with the pretrained model.

c) Use the best performing fine-tuned model (thus most likely not the last one obtained after 50 epochs, the default value in the code, anyway, the package output both, a best model and the final one after the 50 epochs) to generate novel molecules targeting Factor X.

d) Dock the generated compounds using DataWarrior (I’ve got a simple tutorial on docking with DataWarrior as well) or with your favorite docking engine.

🔬 This workflow demonstrates how VeGA can be used effectively in computational drug discovery, showing promising results and highlighting how such tools can spark new ideas. Indeed several compounds have a much better binding score than an FDA approved drug. But of course, these are predicted scores, and in any case, affinity is not the only parameter that defines a good compound, you need the right ADMET properties, but the tool can really help generate nice ideas.

🔗 Try it yourself: https://github.com/piedelre93/VeGA-fo...

⚠️ Disclaimer: Please note that some of the visuals are AI-generated and simplified for illustrative purposes only. They may not accurately reflect the true scientific details or complexities of the subject. It should be clear from the video that some images are cartoon-like and not scientifically accurate, but I wanted to point this out explicitly to avoid any potential confusion.

🤖🎙️🌀 Narrated with Chatterbox TTS, an AI voice model built in Python and custom-tuned on my own voice, slightly robotic, occasionally confused, but always delivering real science stories❗

#DrugDiscovery #GenerativeChemistry #AI