One-line summary: In June 2026, ByteDance officially launched a spin-off and independent funding round for its AI drug discovery business line. ByteDance will retain a controlling stake, with the technology stack and existing pipeline assets moving wholesale into the new entity, Anew Labs. Behind this is a "three-layer model stack" that exploded onto the scene from late 2025: Protenix-v2 (protein structure prediction — the first fully open-source model to surpass AlphaFold3 under strictly aligned AF3 data, parameter count, and inference budget), Anew Omni (all-atom general generative model, wet-lab validated on the hard-to-drug targets KRAS G12D and PCSK9), and AnewSampling (all-atom molecular dynamics, 1000x faster than traditional MD). On April 25 at AAI 2026, the company disclosed an IL-17 small-molecule inhibitor (first pan-inhibition of the IL-17 family AA/AF/FF dimers at the small-molecule level). On July 23, the Seed STEM Scientist Program launched to recruit 100 STEM PhDs. This is the first time a Chinese internet giant has upgraded "AI drug discovery" from a "business unit" to an "independent company."
The Origin: When a General Foundation Model Company Decides to Build Drugs Itself
In June 2026, news spread through the AI pharma circle: ByteDance's AI drug discovery business line officially launched its spin-off and independent fundraising. After the split, ByteDance remains the controlling shareholder; the core team, algorithms, technology platform, and pipeline assets move into the new entity as a whole.
This is not ByteDance's first cross-industry move — it began recruiting for AI drug discovery within its AI Lab at the end of 2020, covering molecular simulation, computational biology, and systems biology. But the concentrated surge in model capability since late 2025 moved ByteDance from "foundational positioning" to "pipeline validation." This is the first time a Chinese internet giant has upgraded AI drug discovery from a business unit to an independent company.
Tencent's model is "empowering drug R&D"; ByteDance's model is "building drugs hands-on." The two strategies reflect fundamentally different judgments about AI pharma's core competitiveness.
From a Guotai Haitong joint research lab monthly report (August 2026): "General foundation model companies enter AI drug discovery differently from Biotechs: they leverage foundation model capabilities (protein structure prediction, molecular generation, scientific agents) as a fulcrum, extending toward drug discovery platforms and pipeline validation. ByteDance is the most systematic among them and the first to move toward industrialization in 2026."
The Three-Layer Model Stack: Protenix x Anew Omni x AnewSampling
ByteDance's AI pharma unit combines "foundation model + molecular generation + molecular dynamics" into a complete three-layer stack. Each layer exists not as an isolated breakthrough but for one purpose: enabling AI to handle small molecules, peptides, and antibodies uniformly at the atomic level.
Layer 1: Protenix-v2 Protein Structure Prediction
In late 2025, ByteDance's Seed team released Protenix, claiming multi-task performance beyond Google's AlphaFold3. The v2 iteration came in April 2026:
> "Protenix-v1 became the first fully open-source model to match or exceed AF3 under strict alignment of AF3 data, parameter count, and inference budget. v2 iterated in April 2026."
Fully open-source + strict AF3 alignment + superior performance means ByteDance produced a reproducible "open-source AF3 replacement" for both academia and industry. AlphaFold3 never open-sourced its training code; Protenix-v2 is the first work to fill that gap.
Layer 2: Anew Omni All-Atom General Generative Model
In March 2026, ByteDance jointly released Anew Omni with Tsinghua University — "the world's first omni-modal drug design foundation model spanning small molecules, peptides, and antibodies."
Core idea: unify at the atomic level. At the atomic level, carbon, nitrogen, oxygen, and hydrogen are fundamentally the same between small molecules and peptides — since the underlying physics is shared, one model can learn it all, with cross-modal mutual benefits.
Anew Omni enables de novo design of cross-class molecules within a unified generative framework, and has completed wet-lab validation on two hard-to-drug targets:
- KRAS G12D — a classic "undruggable" target
- PCSK9 — a key cardiovascular target
- Starting point: high-throughput wet-lab screening
- Data accumulation: experiment-centric, models assist
- Business model: pipeline licensing + proprietary pipelines
- Risk: pipeline failure → valuation collapse
- Starting point: foundation model capabilities (protein prediction + molecular generation + scientific agents)
- Data accumulation: model-centric, experimental validation assists
- Business model: technology platform + pipeline licensing + model APIs
- Risk: model failure → platform valuation collapse
- Compute infrastructure (dedicated clusters)
- Foundation model R&D capability (Protenix / Anew Omni / AnewSampling)
- Data scale (internet-scale training data)
- Interdisciplinary synergy (100 Seed STEM PhDs)
- Wet-lab capability (biology requires long-term accumulation)
- Regulatory experience (IND filings / FDA / NMPA)
- Clinical advancement (Phase I/II/III management)
- Failure tolerance (internet-style fast iteration vs 5-10 year drug development cycles)
- Anew Labs' independent funding round size (expected 2026 H2)
- IL-17 small molecule IND filing progress (preclinical → IND typically takes 12-18 months)
- Follow-on pipelines after KRAS G12D + PCSK9 (more hard-to-drug targets?)
- "Proprietary pipeline count" comparisons: ByteDance vs Insilico vs Recursion
This means Anew Omni's cross-modal knowledge transfer is actually feasible in real tasks, not just on paper.
> Tip: Traditional AI drug design treats different molecule types (small molecules / peptides / antibodies) as separate domains with dedicated generative models, data, representations, and architectures. Anew Omni breaks apart this "domain fragmentation" — a disruptive narrative in which molecular design may evolve from "isolated point technologies" toward "a universal, programmable platform capable of deep molecular reasoning."
Layer 3: AnewSampling All-Atom Molecular Dynamics
If Anew Omni handles "from nothing to something," AnewSampling handles "from something to motion."
Traditional MD simulations rely on femtosecond-scale steps; visualizing millisecond-scale conformational changes can require weeks on supercomputers. AnewSampling uses a generative model that reconstructs the equilibrium distribution of molecular dynamics at the all-atom level, 1000x faster than traditional MD, precisely capturing protein flexibility and ligand binding modes.
From an analysis: "AnewSampling attempts to solve a hard problem: protein-ligand binding is inherently a dynamic process — conformations constantly change, energy states exchange, key interactions form and vanish across time scales. Understanding the dynamic thermodynamic landscape has always been a core challenge in drug discovery."
IL-17 Pan-Inhibition: The First AI-Designed Molecule Reaching Preclinical Stage
On April 25, 2026, ByteDance's AI pharma subsidiary Anew Labs gave an oral presentation at the AAI 2026 annual meeting, publicly disclosing for the first time a preclinical IL-17 small-molecule inhibitor.
IL-17 is a key pathway in psoriasis, ankylosing spondylitis, and other autoimmune diseases. The significance:
> "For the first time, pan-inhibition of the IL-17 family (AA/AF/FF) was achieved at the small-molecule level. Previously, due to limitations of conventional structure-based design, small-molecule drugs could only target IL-17AA and AF; ByteDance achieved inhibition of IL-17AA, IL-17AF, and the extremely hard-to-drug IL-17FF dimer."
IL-17FF is a widely recognized "undruggable" target — no small molecule had ever successfully inhibited it. Using AI-driven virtual screening and molecular generation algorithms, Anew Labs achieved simultaneous blockade of all three dimers at the small-molecule level.
This marks ByteDance's AI pharma capability crossing from "basic research" into "concrete target, concrete molecule pipeline validation," forming a complete technical chain.
Seed STEM Scientist Program: 100 STEM PhDs + Dedicated Compute
On July 23, 2026, ByteDance launched the Seed STEM Scientist Program — a global recruitment of 100 STEM PhDs, providing dedicated compute clusters and interdisciplinary team support. The initial six-month phase focuses on AI applications in materials science and biomedicine.
The logic: the next stage of basic research is the "high-density PhDs + strong compute" combination. AI drug discovery is no longer "a few algorithm engineers plus a lab team" but a systems engineering effort of "100 STEM PhDs + dedicated compute clusters + interdisciplinary teams."
Key Milestones Timeline
| Time | Milestone | Key Numbers | | --- | --- | --- | | Late 2020 | AI Lab begins AI pharma recruiting | Molecular simulation + computational biology + systems biology | | 2023 | Seed founded | AI4S becomes a frontier direction | | Late 2025 | Protenix released | Multi-task performance beyond AF3 | | 2026.3 | Anew Omni with Tsinghua | KRAS G12D + PCSK9 wet-lab validation | | 2026.4 | Protenix-v2 + IL-17 disclosure | First strictly aligned fully open-source AF3-class model | | 2026.4.25 | AAI 2026 | Pan-inhibition of IL-17 AA/AF/FF dimers | | 2026.6 | Anew Labs spin-off / independent funding | ByteDance retains control | | 2026.7.23 | Seed STEM Scientist Program | 100 STEM PhDs + dedicated compute | | 2026 H2 | Split completion + funding in place | Accelerated industrialization |
"General Foundation Model Company" vs "Biotech": Two Fundamentally Different Paths
Biotech path (Recursion, Insilico Medicine, etc.):
General foundation model path (ByteDance Anew Labs, Alibaba's Tongyi Dianshi, etc.):
ByteDance's core thesis: "atomic weaponry" — return to the atomic level, where all molecules (small molecules, peptides, antibodies) share common physics, learned by one unified model. This is fundamentally opposed to the Biotech approach of "separate models per molecule type."
Comparison: ByteDance vs Google / Meta / Insilico
| Dimension | ByteDance Anew Labs | Google Isomorphic Labs | Meta FAIR Chemistry | Insilico Medicine | | --- | --- | --- | --- | --- | | Foundation models | Three-layer stack (Protenix + Omni + Sampling) | AlphaFold 3 | OMat24 / Open Molecules 2024 | Chemistry42 | | Core approach | All-atom unification | Structure prediction → docking | Equiformer 2 + OMat24 | Generative + RL | | Business strategy | Building drugs directly + independent spin-off | Partnerships with Novartis / Lilly | Open-source platform | Proprietary pipelines + software licensing | | Disclosed pipeline | Preclinical IL-17 small molecule | Not disclosed | None | INS018_055 Phase III | | Funding intensity | ByteDance ~35B RMB/year R&D | Google + Novartis + Lilly | Meta FAIR | Hundreds of millions USD raised |
ByteDance's "building drugs directly" resembles Insilico's "proprietary pipelines," but its underlying capability comes from general foundation models rather than traditional Biotech — a new hybrid paradigm of "general foundation model company + proprietary pipelines."
The Risks of the Split: An Algorithm Giant's "Deep Water"
Strengths:
Concerns:
One analysis concludes: "AI is changing this landscape. Molecule-heart-level players have achieved preliminary experimental results on industrially valuable, high-difficulty modalities such as nanobodies and cyclic peptides." Whether ByteDance succeeds post-split will require 3-5 years to observe.
My Observations
ByteDance's spin-off is a watershed event. Before it, China's AI pharma track was mainly Biotech-driven (Insilico, XtalPi, Molecule Heart); after it, general foundation model companies begin entering drug discovery as independent companies.
Success depends on three things:
1. Whether the three-layer stack forms a closed loop — Protenix (static structure) → Anew Omni (static design) → AnewSampling (dynamic validation). If the "static prediction → static design → dynamic validation" cycle works, model capability compounds. 2. Whether wet-lab conversion efficiency improves — ByteDance needs its own wet labs or deep CRO partnerships by 2027 H1. 3. Capital endurance after independent funding — AI pharma is a 5-10 year game; how much the spin-off raises determines how far ByteDance can go in the "no-man's land."
Worth tracking over the next 6-12 months: