AbDiffuser: full-atom generation of in-vitro functioning antibodies
Authors: Karolis Martinkus, Jan Ludwiczak, WEI-CHING LIANG, Julien Lafrance-Vanasse, Isidro Hotzel, Arvind Rajpal, Yan Wu, Kyunghyun Cho, Richard Bonneau, Vladimir Gligorijevic, Andreas Loukas
NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We validate Ab Diffuser in silico and in vitro. Numerical experiments showcase the ability of Ab Diffuser to generate antibodies that closely track the sequence and structural properties of a reference set. Laboratory experiments confirm that all 16 HER2 antibodies discovered were expressed at high levels and that 57.1% of the selected designs were tight binders. |
| Researcher Affiliation | Collaboration | Karolis Martinkus1, Jan Ludwiczak1, Kyunghyun Cho1,4, Wei-Ching Liang2, Julien Lafrance-Vanasse2, Isidro Hotzel2, Arvind Rajpal2, Yan Wu2, Richard Bonneau1, Vladimir Gligorijevic1, Andreas Loukas1 1Prescient Design, Genentech, Roche 2Antibody Engineering, Genentech, Roche 4NYU |
| Pseudocode | Yes | For convenience, in Algorithms 1 and 2 we describe the full training and sampling procedures. |
| Open Source Code | No | The paper does not contain an explicit statement or link indicating that the source code for their methodology is publicly available. |
| Open Datasets | Yes | We evaluate Ab Diffuser on the generation of antibodies from paired Observable Antibody Space (p OAS) [61], modeling of HER2 binders and antigen-conditioned CDR redesign. |
| Dataset Splits | Yes | The Paired OAS dataset [61] was split into 1000 test samples and 1000 validation samples. Similarly, the HER2 binder dataset [55] was split to have 1000 test samples and 100 validation samples. |
| Hardware Specification | Yes | In Table 2 we show that APMixer is able to use an order of magnitude more parameters with a smaller memory footprint during training and offers more efficient sample generation, compared to the baseline architectures, on Nvidia A100 80GB GPU. |
| Software Dependencies | No | The paper mentions "PyTorch [65]" but does not provide a specific version number for PyTorch or any other software libraries used, which is required for reproducibility. |
| Experiment Setup | Yes | We use Adam W for training [52] with a weight decay of 0.01 and with a learning rate of 2 10 4 for all structure models, while the transformer used a learning rate of 1 10 4. We experimented with weight decay of 1e 12 and learning rates in the range of 1 10 3 to 1 10 4 for all models, to determine the chosen values. We also normalize the gradient norm to unit length during training. All models use a batch size of 4. This is the maximum batch size that allows training the baseline models on our 80GB GPU. The APMixer due to the better memory complexity allows for batch sizes up to 32, but to keep the setup similar, we also used a batch size of 4 for it. For paired OAS we train the models for 50 epochs, while for HER2 binder generation we train for 600 epochs, for around 1.25M training steps in both cases. |