Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
GLID$^2$E: A Gradient-Free Lightweight Fine-tune Approach for Discrete Biological Sequence Design
Authors: Hanqun Cao, Haosen Shi, Chenyu Wang, Sinno Jialin Pan, Pheng-Ann Heng
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We validate GLID2E on DNA and protein sequence design benchmarks, demonstrating comparable or superior performance to state-of-the-art methods while achieving significantly faster training and inference with reduced computational requirements. Ablation studies confirm that both the clipped likelihood constraint and reward shaping contribute substantially to stability and performance. |
| Researcher Affiliation | Academia | Hanqun Cao1 , Haosen Shi1 , Chenyu Wang2, Sinno Jialin Pan1, Pheng-Ann Heng1 1The Chinese University of Hong Kong 2Massachusetts Institute of Technology |
| Pseudocode | Yes | Algorithm 1 GLID2E Training Algorithm based on PPO algorithm |
| Open Source Code | Yes | The code is available at: https://github.com/chq1155/GLID2E. |
| Open Datasets | Yes | DNA Design. We use a comprehensive enhancer dataset containing approximately 700,000 DNA sequences of 200 base pairs [44], characterized for activity in human cells via massively parallel reporter assays. Protein Design. The discrete diffusion model is pretrained on 19,700 high-resolution single-chain structures from PDB, following Protein MPNN [2]. |
| Dataset Splits | No | Following DRAKES [14], we use established benchmarks for both tasks. |
| Hardware Specification | Yes | All experiments are conducted on a single NVIDIA A40 GPU with 20GB memory. |
| Software Dependencies | No | The paper does not explicitly provide specific version numbers for software dependencies like Python or PyTorch. It mentions using frameworks like PPO and GAE, but these are algorithms, not software with version numbers. |
| Experiment Setup | Yes | A.1 Implementation details Common Hyperparameters: In our implementation, we utilized several key hyperparameters that were common across experiments. These parameters control various aspects of our reinforcement learning approach, particularly focusing on reward handling and likelihood clipping mechanisms. Table 6: Common hyperparameter configuration utilized across experiments. Table 7: Hyperparameter configuration utilized in our DNA experiments. Table 8: Hyperparameter configuration utilized in our protein experiments. |