ByteDance Spins Off Anew Labs for Independent Financing: A General AI Model Company's Full Play in Drug Discovery
TL;DR: In June 2026, ByteDance's AI drug discovery business line officially launched a spinoff and independent financing round. ByteDance retains a controlling stake, with the core team, algorithms, technology platform, and existing pipeline assets moving wholesale into the new entity. Behind this is a concentrated burst of a "three-layer model stack" since late 2025 — Protenix-v2 (protein structure prediction), Anew Omni (fully atom-based universal generative model), and AnewSampling (fully atom-based molecular dynamics). This is the first time a Chinese internet giant has upgraded AI drug discovery from a "business unit" to an "independent company."
Origin: When a General AI Model Company Decides to Make Drugs Itself
In June 2026, news spread through the AI drug discovery community: ByteDance's AI drug discovery line formally launched a spinoff and independent financing. ByteDance keeps a controlling stake in the new company.
This was not ByteDance's first move into the field — it began AI drug discovery recruiting in its AI Lab in late 2020, covering molecular simulation, computational biology, and systems biology. But the concentrated model breakthroughs since late 2025 let ByteDance jump from "foundational layout" to "pipeline validation." Notably, Tencent's model is "empowering drug R&D," while ByteDance's model is "making drugs itself" — two fundamentally different judgments about AI pharma's core competitiveness.
A research note (cited from a joint research lab monthly report, August 2026) states: "General AI model companies enter AI drug discovery differently from Biotechs: they use foundational model capabilities (protein structure prediction, molecular generation, scientific agents) as the fulcrum, extending toward drug discovery platforms and pipeline validation. ByteDance is the most systematic player and the first to industrialize in 2026."
The Three-Layer Model Stack: Protenix × Anew Omni × AnewSampling
Layer 1: Protenix-v2 Protein Structure Prediction
Released by ByteDance's Seed team in late 2025, Protenix claimed to surpass Google's AlphaFold3 on multiple tasks. The v2 iteration arrived in April 2026:
> Protenix-v1 became the first fully open-source model to match or exceed AF3 under strictly aligned data, parameter count, and inference budgets.
AlphaFold3 never released its training code; Protenix-v2 is the first open-source alternative filling that gap — reproducible by both academia and industry.
Layer 2: Anew Omni — Fully Atom-Based Universal Generative Model
In March 2026, ByteDance and Tsinghua University jointly released Anew Omni — described as the world's first omni-modal drug design large model spanning small molecules, peptides, and antibodies.
Core idea: unify at the atomic level. At the atomic scale, carbon, nitrogen, oxygen, and hydrogen are not fundamentally different between small molecules and peptides — since the underlying physics is shared, one model can learn uniformly and gain across modalities. The model has completed wet-lab validation on two classically "undruggable" targets:
- KRAS G12D
- PCSK9
- Start: high-throughput wet-lab screening
- Data: experiment-centric, model-assisted
- Business model: pipeline licensing + internal pipelines
- Risk: pipeline failure → valuation collapse
- Start: foundational model capabilities (protein prediction + molecular generation + scientific agents)
- Data: model-centric, experiment-assisted validation
- Business model: technology platform + pipeline licensing + model APIs
- Risk: model failure → platform valuation collapse
- Anew Labs' independent financing size (expected 2026 H2)
- IL-17 small-molecule IND filing progress (preclinical → IND typically 12–18 months)
- Additional hard-to-drug target disclosures after KRAS G12D + PCSK9
- Pipeline-count comparisons vs. Insilico and Recursion
This demonstrates that cross-modal knowledge transfer works in real tasks, not just on paper.
Layer 3: AnewSampling — Fully Atom-Based Molecular Dynamics
Traditional MD simulations rely on femtosecond-scale steps; observing millisecond conformational changes can take weeks of supercomputing. AnewSampling rebuilds the molecular dynamics equilibrium distribution with a generative model at the fully atomic level, reported to be 1,000x faster than traditional MD, precisely capturing protein flexibility and ligand binding modes.
IL-17 Pan-Inhibition: First AI-Designed Molecule Reaching Preclinical Stage
On April 25, 2026, Anew Labs gave an oral presentation at AAI 2026 (Immunology meeting, US), unveiling a preclinical-stage IL-17 small-molecule inhibitor:
> The first small-molecule-level pan-inhibition of the IL-17 family (AA/AF/FF). Previously, small molecules could only target IL-17AA and AF due to limitations of traditional structural design; ByteDance achieved inhibition of IL-17AA, IL-17AF, and the extremely hard-to-drug IL-17FF dimer.
IL-17 is a key pathway in autoimmune diseases such as psoriasis and ankylosing spondylitis. Anew Labs used AI-driven virtual screening plus molecular generation to block all three dimers simultaneously — a first. This marks ByteDance's transition from "basic research" into "specific target, specific molecule pipeline validation."
Seed STEM Scientist Program: 100 STEM PhDs + Dedicated Compute
On July 23, 2026, ByteDance launched the Seed STEM Scientist Program, recruiting 100 PhDs worldwide in STEM fields, offering dedicated compute clusters and cross-disciplinary team support. The initial phase focuses on AI applications in materials science and biomedicine. The logic: the next stage of foundational research is a "high-density PhD + strong compute" combination — a systems engineering effort rather than "a few algorithm engineers plus a lab team."
Key Milestones Timeline
| Date | Milestone | Key Figures | | --- | --- | --- | | Late 2020 | AI Lab starts AI drug discovery hiring | Molecular simulation + comp bio + systems bio | | 2023 | Seed founded | AI4S as frontier direction | | Late 2025 | Protenix released | Multi-task surpasses AF3 | | 2026.3 | Anew Omni + Tsinghua | KRAS G12D + PCSK9 wet-lab validation | | 2026.4 | Protenix-v2 | First strictly aligned open-source AF3-class model | | 2026.4.25 | AAI 2026 | IL-17 AA/AF/FF pan-inhibition | | 2026.6 | Anew Labs spinoff & financing | ByteDance retains control | | 2026.7.23 | Seed STEM Scientist Program | 100 STEM PhDs + dedicated compute | | 2026 H2 | Spinoff + financing complete | Accelerated industrialization |
General AI Model Company vs. Biotech: Two Fundamentally Different Paths
Biotech path (Recursion, Insilico Medicine):
General AI model company path (ByteDance Anew Labs):
ByteDance's core thesis is the "atomic weapon": at the atomic level, all molecules (small molecules, peptides, antibodies) share the same physical basis and can be learned by one unified model — fundamentally opposed to Biotech's per-modality modeling approach.
Peer Comparison
| Dimension | ByteDance Anew Labs | Google Isomorphic Labs | Insilico Medicine | | --- | --- | --- | --- | | Foundation models | Three-layer stack (Protenix + Omni + Sampling) | AlphaFold 3 | Chemistry42 | | Approach | Fully atomic unification | Structure prediction → docking | Generative + RL | | Strategy | Making drugs itself + spinoff financing | Partnerships with Novartis/Lilly | Internal pipelines + software licensing | | Disclosed pipeline | IL-17 small molecule (preclinical) | Not public | INS018_055 (Phase III) |
ByteDance's "making drugs itself" is closer to Insilico's model, but its underlying capabilities come from general large models — a hybrid strategy of "general AI model company + internal pipelines."
Concerns About the Spinoff: An Algorithm Giant Enters Deep Water
Strengths: compute infrastructure, foundational model R&D, internet-scale data, cross-disciplinary talent (100 PhDs).
Concerns: wet-lab capabilities (long accumulation required), regulatory experience (IND filings, FDA/NMPA), clinical trial execution (Phase I/II/III management), and failure tolerance (fast internet iteration vs. 5–10 year drug development cycles). Whether the spinoff succeeds will require 3–5 years of observation.
Author's Observations
The spinoff is a watershed. Before it, China's AI drug discovery track was driven mainly by Biotechs (Insilico, XtalPi, Molecule Hearts); after it, general AI model companies enter as independent drug makers. Success depends on three things:
1. Whether the three-layer stack forms a closed loop — static prediction (Protenix) → static design (Anew Omni) → dynamic validation (AnewSampling). 2. Whether wet-lab translation efficiency improves — ByteDance needs its own wet labs or deep CRO partnerships before 2027 H1. 3. Capital endurance after independent financing — AI drug discovery is a 5–10 year cycle.
What to watch in the next 6–12 months: