FrameDiPT extends the FrameDiff framework for protein structure inpainting tasks via SE(3) diffusion. It captures the dynamic between protein-protein interactions, especially in the CDR3 loop design applications. Discover more in github.com/instadeepai/Frame…! (3/4)

LightMHC is a graph neural network based peptide-MHC (pMHC) protein structure prediction model. It predicts pMHC structure with a higher accuracy than #alphafold while being five times faster! Dive into the code: github.com/instadeepai/Light…! (2/4)

We tested the recycling training strategy across four medical imaging data sets (muscle ultrasound, abdominal CT, prostate MR, and brain MR). Our approach significantly outperformed the standard diffusion and self-conditioning training, for both DDPM and DDIM inference. 📊(3/5)

We proposed a novel approach: recycling the model's prediction at the fully-noised timestep for denoising training instead of using ground truth. Particularly crucial for image segmentation labels as the shape/morphological information is often preserved despite noising. 🔄(2/5)

Proud of my latest research on laparoscopic image segmentation. We showed that independent of an increase in unlabelled data, the semi-supervised training strategy improves the segmentation accuracy. Accepted for the @miccai2019 MIL3ID workshop, arxiv.org/abs/1908.08035.