K-Fold: The Foundation Model for Korean AI Drug Development Has Arrived
K-Fold, a foundation model for drug development artificial intelligence (AI) built by Korean researchers, has been released under an open-source license (Apache 2.0) that anyone can freely use. It is noteworthy because AI can shorten the drug development process, which typically takes years and enormous costs just to identify candidate drug compounds. CEO Woo-yeon Kim of Hitz (professor of chemistry at KAIST and former head of the AI Drug Development Support Center at the Korea Pharmaceutical and Bio-Pharma Manufacturers Association), who led the development, built the model with 30 researchers and explained the significance of its development and its technical distinctions at a meeting with the KPBA reporting team.
Regarding the motivation behind the development, Kim said, "We wanted to turn technology monopolized by overseas big tech into domestic proprietary technology." He explained that proprietary technology creates a supply chain anyone can use, which can then be advanced and competed upon. "K-Fold is the starting point in the bio field," he said.
K-Fold's key technical distinction is that it does not use multiple sequence alignment (MSA), which has been employed by models from AlphaFold 1 to RosettaFold. MSA is a highly time-consuming process—if a full prediction takes 26 to 27 minutes, MSA accounts for about 25 minutes—so eliminating it speeds things up. K-Fold aimed to achieve equivalent performance without using MSA.
The second goal was to predict changes in structure rather than the structure itself. Since a protein's structure changes and its function is altered when a drug binds to it, the judgment was that the actual problem to be addressed in drug development is structural change. K-Fold predicts in two stages the structures of proteins A, B, and C both when separate and when bound, simultaneously capturing the changes before and after binding. As a result, it showed performance comparable to or better than existing models in co-folding model benchmarks.
It demonstrated high predictive accuracy in the antibody field, PROTACs, kinase prediction, and GPCRs, which account for one-third of FDA-approved drug mechanisms. Kim explained that GPCRs change structure differently depending on which drug binds to them, and K-Fold accurately captured this.
"From the beginning, the idea was not to make an AlphaFold version 3.1, but to make our own AlphaFold 4," Kim said. "We wanted to create something so different that you would think, if DeepMind had made AlphaFold 4, this is what it might have looked like." In this context, K-Fold stands apart from overseas models released after AlphaFold 3. AlphaFold 3 only published its paper without releasing its code, meaning developers have had to reproduce it through inference, making matching performance difficult.
There are also distinctions in how it was released. While some AI models claim to open-source their code, in many cases only the trained code is actually released. Critics point out that training cannot easily be restarted with trained code, and released data often requires extensive processing. "The reason we released all of the code needed for training this time is exactly that," Kim said. "The data used is built into K-Fold, and the training code is available, so the next steps can be taken."
He compared this to a camera. AlphaFold only revealed the camera's position, and models that released only trained code showed the camera but made it impossible to replicate how it was made. K-Fold, by analogy, is a better camera designed and built entirely from scratch.
