I created a simplified implementation of AlphaFold 3 (limited strictly to protein monomers) in PyTorch to see if it was possible to train on a single GPU. Rather than aiming to reproduce the accuracy of the original paper, the goal was only to generate plausible looking 3D protein structures. This turned out to be harder than expected – you can find a more in depth write up here – https://medium.com/@ogchen/the-worst-method-for-learning-ml-….
In summary, while most of the auxiliary losses did improve with training suggesting it was learning some structural information, the model’s key loss defined on 3D coordinates never seemed to trend downwards (at the time of writing, the model is still training).
Despite failing on that aspect, this project was initially a learning exercise (and one of my first machine learning related projects) and it was a success in that regard. While the outcome is perhaps not as exciting as I had initially hoped, I believe the project code stands on its own so decided to share it here!
Comments URL: https://news.ycombinator.com/item?id=40413372
Points: 1
# Comments: 0