From Peak to Overshadowed: How AlphaFold3 Was Surpassed in Just 21 Months
*English summary of a Chinese tech forum post analyzing the extreme pace of technical iteration in AI-powered biomolecular computation.*
TL;DR
The pace of iteration in AI bio-computation far exceeds other industries. Core drivers include:
- Technical democratization via open-source ecosystems
- Standardized benchmarks lowering the barrier to comparison
- Preprint culture accelerating knowledge diffusion
- The 2024–2025 investment boom concentrating resources
- Major leap from single-protein structure prediction to all-atom biomolecular complex prediction
- Improved Pairformer module (evolved from Evoformer) combined with a diffusion network for iterative refinement
- ~50% higher accuracy than traditional methods on the PoseBusters benchmark
- Adopted by 3M+ researchers across 190+ countries; core contribution recognized with a Nobel Prize
- Structural vulnerability: source code was not released until February 2025, under a restrictive license — leaving room for open-source alternatives
- SeedFold (Jan 17, 2026): protein monomer lDDT 0.8889 vs AF3's 0.8880; antibody–antigen interface DockQ 53.21% vs 47.90%
- Protenix-v1 (Feb 5, 2026): open-sourced inference-time scaling; antibody–antigen prediction improved from 36.01% to 47.68%
- IsoDDE (Feb 10, 2026): 2.3x AF3's success rate on the hardest cases; 19.8x Boltz-2 on high-accuracy antibody–antigen prediction; binding-affinity prediction surpassing traditional FEP methods
- Hardware: $10K–$500K
- Software deployment: $10K–$100K
- Personnel training: $50K–$200K per person
- Workflow adaptation: $100K–$1M
AlphaFold3 was still considered the "global leader" in December 2024. Yet within roughly 14 months (early 2025 to early 2026), it was surpassed across core tasks by new models including Protenix, SeedFold, IsoDDE, and D-I-TASSER — with performance gaps of 2–3x in some scenarios. For end users, this creates severe sunk-cost risk on technology investments, model selection difficulties, and broken result comparability. The post recommends strategies spanning technical architecture (MaaS-first, multi-model ensembles), team capability (T-shaped structures), decision-making (dynamic model evaluation), and ecosystem participation (benchmarking communities).
Key points
AlphaFold3's peak (May 2024 – end of 2024)
The competitive wave (early 2025 – early 2026)
| Time | Model | Organization | Key advance | |---|---|---|---| | Jan 2025 | Protenix | ByteDance AI4Science | First open-source model to match/exceed AF3 under strict alignment conditions | | May 2025 | D-I-TASSER | Nankai University / NUS | CASP15 single-domain and multi-domain champion; physics + deep learning hybrid | | Oct 2025 | OpenFold3 | OpenFold consortium | Fully open-source; best RNA structure prediction | | Jan 2026 | SeedFold | ByteDance Seed team | Width scaling + linear attention + 26.5M distilled training samples; surpassed AF3 on core tasks | | Feb 2026 | IsoDDE | Isomorphic Labs | "AlphaFold4-level" performance, fully closed-source and commercialized |
The concentrated burst of early 2026
Within just 1–2 months:
A cautionary end-user case ("the friend from Qingdao")
A user who completed a local AlphaFold3 deployment at the end of December 2024 incurred substantial costs:
Recommended coping strategies
1. Technical architecture: prefer Model-as-a-Service over local deployment; build multi-model ensembles 2. Capability building: adopt T-shaped team structures 3. Decision mechanisms: continuously and dynamically re-evaluate model choices 4. Ecosystem participation: engage with benchmarking communities to track the state of the art
Conclusion
From AlphaFold3's release (May 2024) to its comprehensive supersession (February 2026) took only 21 months — an innovation density rarely seen in the history of science and technology, reflecting the field's rapid shift from "technology scarcity" to "choice overload."