How a Decade-Old Graph Embedding Paper Became the Backbone of AI‑Driven Drug Design

On July 1 2026, the International World Wide Web Conference honored the 2015 paper LINE: Large‑scale Information Network Embedding with the WWW Test of Time Award. Authored by Dr Tang Jian and collaborators, the paper has been cited over 7,300 times and is credited with enabling large‑scale graph representation learning that powered search engines, PageRank, knowledge graphs, and recommendation systems.

Beyond the web, the same graph‑embedding mathematics proved adaptable to the emerging field of geometric deep learning. By treating molecules, proteins, and cellular interaction networks as graphs, the authors extended LINE’s ideas to bio‑AI. In 2023 Dr Tang founded the AI‑native biotech company Hundred‑Ao Geometry , whose core technology follows the LINE lineage.

The company released a series of models named GeoFlow . GeoFlow V1 (June 2024) performed full‑atom modeling of protein‑protein complexes and achieved structure‑prediction accuracy comparable to DeepMind’s AlphaFold 3, a benchmark that had previously defined the state of the art.

GeoFlow V2 (April 2025) introduced a unified architecture that combined structure prediction and de‑novo design in a single model using a “pseudo‑protein sequence” trick. This allowed the model to predict structures when given a complete sequence and to generate novel sequences when parts of the input were masked. The model reached a Top‑1 success rate 45 % higher than V1 on protein‑protein complex prediction and achieved an average hit rate of 18.7 % on ten independent antibody design tasks—approximately a hundred‑fold improvement over prior methods.

GeoFlow V3 (October 2025) added multi‑step reasoning, mimicking the natural affinity‑maturation process of antibodies. The model iteratively generates, evaluates, and refines designs, enabling “self‑assessment and autonomous evolution.” In internal benchmarks on seven clinical targets and ten nanobody design tasks, V3’s hit rate remained near two‑tenths, and its Top‑1 success rate on complex prediction improved another 45 % over V2.

These technical advances are situated within a broader industry shift: after AlphaFold’s Nobel‑winning breakthrough, major AI labs (Anthropic, OpenAI, DeepMind/Isomorphic Labs) have poured resources into life‑science AI. Nonetheless, many AI‑driven drug projects have struggled to reach the clinic, highlighting the significance of Hundred‑Ao Geometry’s tangible deliverables—high‑affinity antibodies, stabilized dengue‑virus envelope protein dimers, and AI‑designed enzymes with >99.9 % chiral purity and up to 80 % cost reduction.

Funding milestones include a multi‑hundred‑million‑yuan strategic round led by Shanghai Biomedical Innovation Fund and others. The company’s contributions also extend to open‑source efforts such as NVIDIA’s La‑Proteina and the in‑house virtual‑cell model PerturbDiff , underscoring its role as both a research leader and a commercial pioneer in AI‑for‑science.

Overall, the article argues that the ten‑year trajectory from LINE’s web‑graph embeddings to GeoFlow’s atomic‑level protein design exemplifies a rare “test‑of‑time” scientific lineage, where a foundational algorithmic insight continues to generate transformative value across disparate domains.