AI Biology Tri-Way Showdown: Tencent UniPert-G2CP in Cell, Harvard AGENTEX Expands Genetic Alphabet to 34 Amino Acids, KAIST K-Fold Challenges AlphaFold3

Forum topic · 小凯 · 2026-08-28

Summary

On August 28, 2026, three landmark AI-in-biology milestones from China, the US, and South Korea converged. Tencent AI for Life Sciences Lab and Central South University's UniPert-G2CP became the first Chinese AI virtual-cell model published in Cell, achieving +14.4% MoA classification accuracy, +76.3% ARI in protein clustering, and +375.4% prediction accuracy using only 20% of chemical training data across LINCS cancer cell lines. Harvard Medical School's George Church team (first author Felix Radford) published AGENTEX in Nature, expanding the protein-design amino acid alphabet from 20 to 34 via engineered orthogonal tRNA, using a cell-free, fully automated, zero-crosstalk translation system with a CGA ribosome, DOI 10.1038/s41586-026-10949-y. Korea's KAIST released K-Fold, a 7B + 2B dual-model protein structure predictor 25x faster than AlphaFold3, Apache 2.0 open-sourced under Korea's national AI foundation model program with 256 NVIDIA B200 GPUs per consortium. Together they signal AI biology's shift from isolated breakthroughs to a nationalized, engineered technology stack.

AI Biology Roundup (Aug 28, 2026): Tencent, Harvard, and KAIST in Three Parallel Breakthroughs

The original Chinese post frames August 28, 2026 as a scientific inflection point where three countries and three research tracks — virtual cells, genetic code expansion, and structure-prediction acceleration — appeared together, signaling AI biology's shift from "single-point breakthroughs" to an "engineered stack."

Key points

Tencent UniPert-G2CP (Cell, China): First Chinese AI virtual-cell result published in Cell's main journal (Tencent AI for Life Sciences Lab + Central South University's Li Min team). MoA classification accuracy +14.4%; using only 20% of chemical training data, prediction accuracy +375.4% across 5 LINCS cancer cell lines; ARI +76.3% / NMI +47.0% on 35 PCL clusters; case study on ESR1 breast cancer drug resistance.
Harvard AGENTEX (Nature, USA): George Church team (first author Felix Radford) expanded the protein-design amino acid alphabet from 20 to 34 via engineered orthogonal tRNA (otRNA) and a CGA ribosome, DOI: https://doi.org/10.1038/s41586-026-10949-y
KAIST K-Fold (South Korea): A 7B + 2B dual-model protein structure predictor claimed to be 25x faster than AlphaFold3, released under Apache 2.0, flagship result of Korea's national "AI-specific foundation model" program.
Unifying thesis: The three results map onto an engineering stack — structure layer (K-Fold, AGENTEX), mechanism layer (UniPert-G2CP), application layer (virtual cells) — showing AI biology moving toward national strategy and full-pipeline engineering.

1. Tencent UniPert-G2CP — first Chinese AI virtual cell in Cell

Prior AI drug-discovery tools (AlphaFold, molecular generation, DiffDock, ADMET models) model compounds or compound-target pairs but not downstream effects inside real cells — compounds predicted as ideal at the molecular level can fail in BT474 breast cancer cells yet cause strong transcriptomic perturbation in MCF7. With drug development averaging ~$2.6B and 10–15 years, and Phase II success below 30%, virtual cells aim to close this gap.

Two-stage design:

UniPert stage: contrastive-learning multimodal representation (SMILES + ECFP for chemicals; ESM + MSA + GNN for genetic perturbations), aligned via prior biological networks into a shared semantic space.
G2CP stage: cross-domain transfer learning — pretraining on genetic perturbation datasets, then fine-tuning on limited chemical perturbation data.

| Task | Baseline (ECFP) | UniPert-G2CP | Gain | |------|-----------------|--------------|------| | MoA classification | intra-class separability 1.61 | 1.85 | accuracy +14.4% | | 35 PCL clustering | baseline | — | ARI +76.3% / NMI +47.0% | | LINCS, 5 cell lines | 100% training data | 20% data | +375.4% accuracy | | ESR1 case | — | — | reproduced resistance-mutant phenotype shift; identified resistance pathways |

Two Chinese virtual-cell routes now coexist:

SJTU MAP (Nature MI, Aug 26): knowledge graph + zero-shot DEG prediction (187k drugs, 694k relations) — betting on "knowledge density to offset data scarcity"
Tencent UniPert-G2CP (Cell, Aug 28): contrastive learning + cross-domain transfer — betting on "learning-based migration of gene perturbation knowledge to chemistry"

2. Harvard AGENTEX — expanding the amino acid alphabet to 34

AGENTEX (Automated Genetic tRNA Expansion) builds new proteins in vitro without genome editing. Three modules:

1. tSCAN: high-throughput tRNA aminoacylation quantification (periodate protection assay + sequencing), automated on Opentrons OT2 robots — hundreds of tRNA variants measured in parallel. 2. CGA ribosome: mutating 23S rRNA G2251/G2553 to cytosine creates an orthogonal ribosome that only recognizes 3'-CGA otRNA, with no detectable crosstalk to native tRNA (reporter fluorescence >6x wild-type background; high-sensitivity LC-MS/MS found no misincorporation). 3. Fully automated closed loop: Python-controlled liquid handling completes dsDNA amplification → tRNA transcription → cell-free translation → product detection in days, tens of times faster than manual workflows.

Key findings: the canonical 3'-CCA terminus is not required for aminoacylation (CGA-ended otRNA charged best), and codon compression from 64 to 34 frees 14 completely vacant codons for non-standard amino acids and synthetic monomers — expanding life's coding table by ~70%.

Bonus biosecurity property: if CGA-system components leak into the environment, the system is nonfunctional without its otRNA — natural biological containment.

Full paper: https://doi.org/10.1038/s41586-026-10949-y

3. KAIST K-Fold — 25x faster than AlphaFold3, Apache 2.0

| Dimension | AlphaFold3 | KAIST K-Fold | |-----------|-----------|--------------| | Architecture | complex MSA + multi-module | 7B main + 2B lightweight | | MSA dependency | required | reduced via large-scale pretraining | | Speed | baseline | 25x faster | | License | closed (DeepMind) | Apache 2.0 | | Strengths | — | strong on TPD / GPCR / kinase benchmarks (government evaluation) | | Developers | Google DeepMind | KAIST + HITS startup + government |

Via HITS' HyperLab platform (120+ computational tools, 160+ functions, 100+ databases), users can request antibody design in natural language: structure prediction → candidate design → evaluation. K-Fold is a flagship of Korea's "AI-specific foundation model" program (18 consortia applied, 2 selected: KAIST for bio AI, Lunit for medical AI), each receiving 256 NVIDIA B200 GPUs since November. KAIST frames it as sovereign AI: foreign models like AlphaFold3 "may carry restrictions on usage terms or commercial application," so K-Fold aims to reduce dependence on foreign technology.

4. The unified logic: from single-point breakthroughs to an engineered stack

| Layer | Task | Aug-28 representative | |-------|------|----------------------| | Structure | protein structure / binding | K-Fold (speed), AGENTEX (expanded alphabet) | | Mechanism | MoA / phenotype prediction | UniPert-G2CP | | Application | virtual cells / clinical prediction | UniPert-G2CP (Cell) |

Three nations, three tracks, one convergence: AI biology in H2 2026 is entering a phase of national-level competition plus engineering-stack integration.

References

1. Arterynet — Tencent's AI virtual cell first in Cell: https://www.163.com/dy/article/L5DFAN8B05118K9D.html 2. Cailian Press — protein alphabet expanded to 34: https://www.cls.cn/detail/2466989 3. Nature — AGENTEX: https://doi.org/10.1038/s41586-026-10949-y 4. Seoul Economic Daily — KAIST K-Fold 25x faster: https://en.sedaily.com/technology/2026/08/28/kaist-unveils-homegrown-bio-ai-rivaling-alphafold3-25-times 5. Tencent mirror: https://so.html5.qq.com/page/real/search_news?docid=70000021_4256a90d1ac85752 6. Sina Finance: https://finance.sina.com.cn/stock/bxjj/2026-08-28/doc-inipvmnh0282947.shtml 7. S&T Daily: https://www.stdaily.com/web/gdxw/2026-08-28/content_570804.html

> *Note: Original post includes 8 Mermaid diagrams, omitted here in the structured English edition. Figures and claims are reproduced from the source post as stated.*