IgGM: A Generative Model for Functional Antibody and Nanobody Design

Authors: Rubo Wang, Fandi Wu, Xingyu Gao, Jiaxiang Wu, Peilin Zhao, Jianhua Yao

ICLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental 4 EXPERIMENTS We constructed our training, validation, and test sets from the SAb Dab database... Table 1: Complex structure prediction... Table 2: Results of the novel antibody design on SAb-2023H2-Ab... Ablation studies (Appendix E)
Researcher Affiliation Collaboration Rubo Wang1, 2, 3 , Fandi Wu3 , Xingyu Gao1, 2 , Jiaxiang Wu3, Peilin Zhao3, Jianhua Yao3 1Institute of Microelectronics, Chinese Academy of Sciences, Beijing, China 2University of Chinese Academy of Sciences, Beijing, China 3Tencent AI Lab, Shenzhen, China
Pseudocode Yes Algorithm 1 Ig GM Sampling Input: Model fθ( , ), sequence of time points τ1 > τ2 > > τN 1, initial noise (ˆs T , ˆx T ), antigen s A, x A) (s, x) fθ((ˆs T , ˆx T ), T, (s A, x A)) for n = 1 to N 1 do Sample Qz = Q1Q2...QT q(xt = j|xt 1 = i), xz (N(0, I), Uniform(SO(3))) ˆsτn s Qz, ˆxτn x + p τ 2n ϵ2xz x fθ((ˆsτn, ˆxτn), τn, (s A, x A)) end for Output: (s, x) =0
Open Source Code Yes 1Code is available at: https://github.com/Tencent AI4S/Ig GM
Open Datasets Yes We constructed our training, validation, and test sets from the SAb Dab database, employing the widely used method of dividing the dataset based on time, as previously established in other works (Jumper et al., 2021; Ruffolo et al., 2023; Wu et al., 2024; Abramson et al., 2024).
Dataset Splits Yes We constructed our training, validation, and test sets from the SAb Dab database... This process resulted in 101 validation samples and 60 test samples, both of which were completely unrelated to the training set.
Hardware Specification Yes This process lasted for 5 days on 8 A100 GPUs.
Software Dependencies No The paper does not provide specific software dependencies with version numbers, such as programming languages or libraries, that would be needed to replicate the experimental environment.
Experiment Setup Yes we use the Adam (Loshchilov & Hutter, 2017) optimizer and set the batch size of the training process to 32. We also maintain an EMA (Exponential Moving Average) decay of 0.999 for the model parameters... In the model training process, we assigned probabilities of 4 : 2 : 2 : 2 for the model to design CDR H3, CDR H, all CDRs, and to refrain from sequence design, respectively.