AffinityFlow: Guided Flows for Antibody Affinity Maturation
Authors: Can Chen, Karla-Luise Herpoldt, Chenchao Zhao, Zichen Wang, Marcus D. Collins, Shang Shang, Ron Benson
ICML 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Our method, Affinity Flow, achieves state-of-the-art performance in proof-of-concept affinity maturation experiments. ... We conduct our experiments using a sd Ab subset of the SAb Dab dataset (Dunbar et al., 2014). ... We measure performance using three metrics: functionality, specificity, and rationality, following (Ye et al., 2024). ... In Table 1, we present the experimental results on four settings CDR-H1, CDR-H2, CDR-H3 and all design positions. ... Ablation Studies We use Affinity Flow as the baseline to evaluate the effect of removing specific modules, with results shown in Table 2. |
| Researcher Affiliation | Collaboration | 1Mila Quebec AI Institute (work done during an Amazon internship) 2Amazon. Correspondence to: Can (Sam) Chen <EMAIL or EMAIL>, Marcus Collins <EMAIL>. |
| Pseudocode | No | The paper describes its methods through illustrations (Figure 1 and Figure 2) and narrative text. It does not contain any clearly labeled pseudocode or algorithm blocks. |
| Open Source Code | No | The paper does not provide an explicit statement about releasing its source code, nor does it provide a link to a code repository for the methodology described. |
| Open Datasets | Yes | We conduct our experiments using a sd Ab subset of the SAb Dab dataset (Dunbar et al., 2014). |
| Dataset Splits | Yes | Our study focuses on sd Abs, selecting PDB files of 120 labeled sd Ab-antigen pairs to initially train our predictors using mean squared loss. From these files, we extract 77 sd Abs and 54 antigens, resulting in 4, 158 docked complex structures generated by Geo Dock. Rosetta is then used to calculate the G for these complexes. For maturation testing, we select 60 sd Ab-antigen PDB files, ensuring that each antigen is unique and these antigens and antibodies were not included in the training set. |
| Hardware Specification | Yes | All experiments are conducted on a g5.24xlarge server equipped with GPUs with 23GB of memory. |
| Software Dependencies | No | The paper mentions specific software tools like ESM2-8M, GVP model, and Adam optimizer (Kingma, 2014) but does not provide version numbers for these or other crucial software components like programming languages or libraries. |
| Experiment Setup | Yes | Our method utilizes the alternating optimization framework with three iterations, where each iteration introduces single-point, double-point, and triplepoint mutations. This allows for producing 1 to 9 mutations in total. We set the Alpha Flow sampling steps T to 3 per iteration with a schedule of [1.0, 0.6, 0.3, 0.0] and use a default scaling factor γ of 5. We employ ESM2-8M, followed by a hidden-dim-320 three-layer MLP, as the sequence-based predictor parameterized by α. For the structure-based predictor parameterized by β, we use a five-layer GVP model, which takes the structure and ESM2-8M residue embeddings as input. For the co-teaching module, we use a batch size of 256 and a learning rate of 1 10 4 with the Adam optimizer (Kingma, 2014). |