EVA: Geometric Inverse Design for Fast Protein Motif-Scaffolding with Coupled Flow
Authors: Yufei Huang, Yunshu Liu, Lirong Wu, Haitao Lin, Cheng Tan, Odin Zhang, Zhangyang Gao, Siyuan Li, Zicheng Liu, Yunfan Liu, Tailin Wu, Stan Z Li
ICLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We conducted experiments on the RFDiffusion dataset (Watson et al., 2023) and benchmarked EVA with various methods... We also designed a new vaccine design benchmark that reflects real-world scenarios with our curated datasets. Furthermore, we test EVA in cases requiring to find optimal motif indexes or multi-motif scaffolding... We compare EVA with state-of-the-art methods in RFDiffusion benchmark in Table 1. EVA is 70 faster than current state-of-the-art RFDiffusion, which supports our analysis about the trade-off and demonstrates the effectiveness of the spatial context. As a training-free approach, we achieve comparable performance with current SOTA training-based methods, RFDiffusion and superior performance on targets with total length <100. It should be noted that RFDiffusion is based on a much more complex protein structure prediction model, which is better at generating larger proteins and requires time-consuming structure prediction pretraining and generative fine-tuning. We also achieve leading performance across training-free methods, including SOTA particle-based SMC methods, TDS, which enjoys good performance in the cost of intensive guidance computation. ... Tables 1-12 and Figures 1-9 display empirical results and comparisons. |
| Researcher Affiliation | Academia | 1 Zhejiang University, Hangzhou 2 AI Lab, Research Center for Industries of the Future, Westlake University EMAIL; EMAIL |
| Pseudocode | Yes | Algorithm 1 A training-free Coupled Flow-based Framework for fast Motif-Scaffolding (EVA) Algorithm 2 A training-free Coupled Flow-based Framework for fast Motif-Scaffolding (EVA) |
| Open Source Code | No | The paper states, "We implement all the inverse solver baselines according to their official Git Hub repositories and change the data structure from images to protein." This refers to the baselines' code, not the authors' own implementation of EVA. There is no explicit statement or link provided for the code of the methodology described in this paper. |
| Open Datasets | Yes | We conduct sampling for motif-scaffolding on the RFDiffusion Benchmark (Watson et al., 2023) of 24 targets following previous works (Trippe et al., 2023; Wu et al., 2023; Zheng et al., 2024). ... For vaccine design, we established an in silico benchmark test comprising 10 vaccine design targets addressed in recent publications, including epitopes from the respiratory syncytial virus (RSV) fusion protein (RSVF) that can produce neutralizing effect (Castro et al., 2024). More details can be referred in the Appendix B. ... We download the RFDiffusion benchmark sets via PDB. |
| Dataset Splits | No | The paper does not provide specific train/test/validation splits for any dataset it uses for training or evaluating its model. It mentions generating "100 scaffolds per target" for evaluation, and uses "pretrained flow models", but does not specify how the underlying datasets for these pretrained models were split, nor does it define new splits for its own experiments beyond the number of generated samples. |
| Hardware Specification | Yes | All baselines and our approach are implemented using the Py Torch 1.6.0 library with Intel(R) Xeon(R)Gold6240R@2.40GHz CPU and NVIDIA A100 GPU. |
| Software Dependencies | Yes | All baselines and our approach are implemented using the Py Torch 1.6.0 library with Intel(R) Xeon(R)Gold6240R@2.40GHz CPU and NVIDIA A100 GPU. |
| Experiment Setup | Yes | We generate 100 scaffolds per target with 100 time-steps using Euler integrator... All sampling-based methods use a batch of 25 for every case, generating 100 scaffolds in total in a single A100 GPU. Because of the memory burden of running RFDiffusion, we instead use a batch of 2 in a single A100 GPU. ... Table 8: More ablation study of EVA. coupling strength gamma = 0.8 / 1.2, starting time t0=0.15 / 0.05. EVA (original, gamma=1.0, t0=0.1) |