Uni-Bond: How a Two-Week Summer School Project Evolved into an ICML 2026 Workshop Paper
Uni-Bond is a molecular bond graph prediction model that outperforms the previous SOTA (YuelBond) by approximately 20 times in exact match error rate. Developed during a summer school project at AIRI, it was later refined and accepted at the Graph Foundation Models workshop at ICML 2026. The model uses a frozen Uni-Mol encoder with a lightweight MLP head on pairwise atomic representations.
The project started at the AIRI summer school at Tomsk State University in 2025. The team (Stepan Pavlenko, Pavel Maslov, Ivan Burov) initially aimed to generate SMILES strings from 3D atomic coordinates using LLMs with RL finetuning, but pivoted to bond graph prediction after three days. Using a pre-trained geometric encoder Uni-Mol (~47M parameters) and a lightweight MLP head (~74K trainable parameters), they achieved macro F1 of 0.9997 and exact match of 99.53% on the GEOM benchmark (scaffold split), compared to YuelBond's 0.975 F1 and 83.2% exact match — a ~20x reduction in error rate. The model also avoids false bonds in dimers unlike YuelBond. A parallel track used LoRA-finetuned Galactica-1.3B with coordinate tokenization (like BindGPT) for direct SMILES generation, achieving >99.7% validity on indistribution data but near-zero exact match on out-of-domain peptides. The work was accepted at the Graph Foundation Models workshop at ICML 2026 in Seoul.
- Сокращения
- LLM = Large Language Model — Большая языковая модель
- SOTA = State Of The Art — Современное состояние дел
- GNN = Graph Neural Network — Графовая нейронная сеть
- MLP = Multilayer Perceptron — Многослойный перцептрон
- F1 = F1 score (harmonic mean of precision and recall) — F1-мера
- RL = Reinforcement Learning — Обучение с подкреплением
- LoRA = Low-Rank Adaptation — Низкоранговая адаптация
- ICML = International Conference on Machine Learning — Международная конференция по машинному обучению
Source: Habr — хаб ИИ —
original