PEACCEL at NeurIPS 2025: Epistasis-Aware AI for Drug Discovery
23 December 2025
NeurIPS, San Diego, December 2025. PEACCEL presented new research on hypercube- constrained graph learning for protein fitness prediction, emphasizing constraint handling and epistasis modeling, two levers with direct impact on discovery timelines and hit quality.
See NeurIPS 2025 (San Diego) and the Constrained Optimization for ML (COML) Workshop: https://neurips.cc/virtual/2025/loc/san-diego/123294
What we showcased
- Hypercube-Constrained Graph Learning to enforce smoothness only where biology supports it (i.e., constraint-aware learning over mutational neighborhoods).
- Wavelet-denoised supervision to counter measurement noise and sparsity in sequence–fitness labels, boosting label-efficiency for low-N regimes common in early discovery.
- Epistasis-aware prediction on discrete mutational spaces (Hamming hypercubes), improving generalization to unseen variants while respecting feasibility constraints relevant to downstream design.
Why it matters for drug discovery and design
- Higher hit rates with fewer assays. Constraint-aware training and denoising reduce experimental burden while preserving accuracy on rugged, epistatic landscapes.
- Faster iteration cycles. Semi-supervised learning leverages unlabeled variants accelerating prioritization before scale-up.
- Better transfer to developability filters. Graph-structured outputs and calibrated constraints integrate cleanly with ADMET and safety heuristics.
Collaboration footprint
This work was delivered with academic partners across Australia and the United States, reflecting a globally distributed R&D model aligned with pharma expectations for translational validation.
Engage with us
If you are exploring AI-driven drug discovery or protein design, we welcome a technical discussion.
Event links: NeurIPS 2025 (San Diego, Dec 2–7, 2025): neurips.cc
For more information:
PEACCEL
Making the world disease free
Contact: AI-team@peaccel.com
http://www.peaccel.com/