Disentangled Wasserstein Autoencoder for T-Cell Receptor Engineering
Authors: Tianxiao Li, Hongyu Guo, Filippo Grazioli, Mark Gerstein, Martin Renqiang Min
NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | To demonstrate its effectiveness, we apply it to T-cell receptors (TCRs), a well-studied structure-function case. We show that our method can be used to alter the function of TCRs without changing the structural backbone, outperforming several competing methods in generation quality and efficiency, and requiring only 10% of the running time needed by baseline models. |
| Researcher Affiliation | Collaboration | 1 NEC Laboratories America, 2 Yale University 3 National Research Council Canada, 4 NEC Laboratories Europe |
| Pseudocode | No | The paper describes the model architecture and training process in text and equations but does not include any explicit pseudocode or algorithm blocks. |
| Open Source Code | No | The paper does not contain an explicit statement or link indicating that the source code for the described methodology is publicly available. It references an external tool's GitHub repository, but not its own code. |
| Open Datasets | Yes | Interacting TCR-peptide pairs are obtained from VDJDB [36], merged with experimentallyvalidated negative pairs from Net TCR [26]. We also performed the same experiments on the Mc PAS-TCR dataset [37] (Appendix A.1). |
| Dataset Splits | Yes | We then split the TCRs into train/test/validation sets with a ratio of 8:1:1, and put all pairings of each TCR to the respective subset, to ensure all TCRs in the test and validation sets are not seen in the training. |
| Hardware Specification | Yes | The model is trained on 2 rtx3090 GPUs with a batch size of 256 (128 per GPU). |
| Software Dependencies | No | The paper describes model components like Transformer encoders [38] and LSTM [39], and optimizers like Adam [61], but does not provide specific version numbers for software libraries or dependencies (e.g., PyTorch 1.9, Python 3.8). |
| Experiment Setup | Yes | The dimension of zf is fixed to 8 and zs to 32. We train each model with 200 epochs and a learning rate of 1e 4, and evaluate the checkpoint of every 50 epochs... The hyperparameter settings of the models for comparison and visualization are: [β1 = 1.0, β2 = 0.1, epoch = 200] where β s are weights of the losses: L = Lrecon + β1Lf_cls + β2LWass. ... We use the scheduled sampling technique [62] for the LSTM decoder during training, where for each position in the input sequence, there is a 0.5 probability of using the previous predicted token, instead of the original token, to calculate the hidden state for the next position. |