Every stage of capsid engineering now has an AI that beats the human record
The AAV environment turns gene-therapy vector design into four verified, leakage-audited tasks built on real screening and structural data. Inside it, frontier systems clear the published state of the art at all four stages — viability, tropism, structure, and de-novo design. Apodex model-generated capsids are in preclinical laboratory testing; no human data.
The environment
Four tasks inside one executable sandbox
Not a prompt and not a static dataset. The solver acts through a fixed tool interface inside an isolated, stateful sandbox that answers back — data, code execution, folding engines, a metered scoring action. Ground truth stays inside; the full trajectory is recorded and graded twice, once on outcome and once on process (HDS6).
TASK 1binary classification
Viability
Will this engineered capsid assemble and package its genome at all? Trained on lightly mutated variants, tested on heavily mutated ones.
per-distance AUROC · chance 0.5
TASK 2regression / ranking
Tropism
Which tissue does a 7-mer insert reach — in mouse, across species to macaque, and in human cells?
rank correlation vs. measured enrichment
TASK 3agentic · structure
Structure
Build the capsid in 3-D from sequence alone: subunit surface loops, the full 60-mer shell, and the complex with an antibody or receptor.
positional accuracy · interface quality
TASK 4agentic · generative
Design
Generate 1,000 novel peptides for a target under a 100-query oracle budget. A design counts only if it is novel, manufacturable and on target.
hit rate over 1,000 designs
WHY A SCORE HERE MEANS SOMETHING
Real data, not simulation — deep-mutagenesis and pooled screens (Bryant 2021, Ogden 2019, Fit4Function 2024, LY6A/LY6C1 2023) plus experimental PDB structures.
Out-of-distribution by construction — splits across mutational load, sequence position, species, and structure deposition date (all targets deposited after 30 Sep 2021, past every folding engine’s training cut-off).
Labels are unreachable — ground truth is exposed only through a metered scoring action, never as readable labels; structure runs are network-isolated.
Anti-gaming in the metric — viability AUROC is computed within each exact mutation count, so “count the mutations” scores chance; the design oracle is capped far below the submission size, so it cannot be brute-forced.
Environment · at a glance
SCOPE
4 tasks · 18 episodes · 11 solvers evaluated
GRADING
hidden outcome verifier + blind HDS6 process judge (0–4)
LEAKAGE CONTROL
OOD splits on load, position, species, deposition date
FEEDBACK BUDGET
design oracle capped at 100 queries per episode
REFERENCE
per-task published SOTA = strongest single published method on that task
Why this task
Delivery is the bottleneck of gene therapy — and the capsid decides it
AAV is the leading vector for treating inherited retinal disease, hemophilia, muscular dystrophy and neurological conditions. Its protein shell — the capsid — decides the two things that gate every program: which organ the vector reaches, and whether it can be manufactured at all.
These pull against each other. The surface changes that improve delivery or dodge pre-existing immunity are the same changes that stop the virus assembling or packaging its DNA. And the search space is not searchable by hand: a single 7-amino-acid insertion in one surface loop already spans 20⁷ ≈ 1.28 billion variants, of which the heavily engineered ones survive at roughly 0.3%.
So this is not a toy protein benchmark. It is the full decision chain a real campaign runs on — manufacturable → on-target → novel → worth the bench time — and every one of those decisions is scored separately, against measurements someone actually made.
~0.3%
of heavily engineered variants survive assembly and packaging. Viability is the rate-limiting filter of any capsid campaign.
1.28 B
7-mers at the AAV9 588 loop — the space a useful model has to triage in silico before anyone pipettes.
4 approved
AAV gene therapies to date. Every additional one starts with a capsid that reaches the right tissue and can be made.
Results
The published state of the art fell at every stage — and hardest where it counts
Eleven frontier systems ran the full pipeline. apodex-1.1 — our in-house solver on top of the environment — exceeds the published state of the art on all four tasks, and the margin is widest on the generative stage, the one that actually produces candidates.
Ranked by mean result score over Tasks 1–3. Each cell shows the result score with its HDS6 process score (0–4) beside it. All solvers run inside the domain-specific environment except CC opus-4-8, which uses the generic Claude Code harness.
| # | Solver | T1 Viability | T2 Tropism | T3 Structure | T4 Design | Mean · T1–T3 |
|---|---|---|---|---|---|---|
| 1 | kimi-k3 | 0.9363.49 | 0.6493.61 | 0.6793.68 | 0.1943.59 | 0.755 |
| 2 | claude-opus-5 | 0.9363.46 | 0.6533.55 | 0.6683.61 | — † | 0.752 |
| 3 | glm-5.2 | 0.9323.27 | 0.6443.76 | 0.6653.39 | 0.1913.32 | 0.747 |
| 4 | claude-opus-4-8 | 0.9313.65 | 0.6353.30 | 0.6573.36 | — † | 0.741 |
| 5 | apodex-1.1 in-house | 0.9043.36 | 0.6353.37 | 0.6493.52 | 0.1803.23 | 0.729 |
| 6 | CC opus-4-8 w/o env | 0.921— | 0.633— | 0.595— | — † | 0.716 |
| 7 | deepseek-v4-pro | 0.9183.37 | 0.6132.97 | 0.6102.19 | 0.1292.64 | 0.714 |
| 8 | claude-sonnet-4-6 | 0.9223.45 | 0.6282.77 | 0.5752.73 | 0.0942.97 | 0.709 |
| 9 | gpt-5.5 | 0.9092.99 | 0.6003.57 | 0.5441.22 | 0.1422.24 | 0.684 |
| 10 | qwen3.5-397b | 0.8963.27 | 0.6052.93 | 0.5431.97 | 0.0852.21 | 0.682 |
| 11 | apodex-1.0 in-house | 0.8843.23 | 0.5182.90 | 0.5261.84 | 0.0722.34 | 0.643 |
| Reference — published state of the art | ||||||
| published SOTA | 0.878 | 0.622 | 0.605 | 0.116 | 0.702 | |
Apodex (in-house) score = task headline metric · small = HDS6 process score (0–4) † Opus-series models declined Task 4, consistently triggering biosafety refusal
The environment is worth more than a model upgrade. The same model — claude-opus-4-8 — scores 0.741 inside the domain-specific environment and 0.716 with the generic Claude Code harness. The gap is widest exactly where orchestration and self-validation matter most: structure, 0.657 vs. 0.595.
apodex-1.1 vs. the published state of the art, stage by stage. published SOTA is the strongest single published method on each task.
| Stage | published SOTA method | published SOTA | apodex-1.1 | Δ | Best in env |
|---|---|---|---|---|---|
| Viability | CAP-PLM (capsid language model) | 0.878 | 0.904 | +0.026 | 0.936kimi-k3 |
| Tropism | Fit4Function production model | 0.622 | 0.635 | +0.013 | 0.653opus-5 |
| Structure | AlphaFold 3 (+ symmetry expansion) | 0.605 | 0.649 | +0.044 | 0.679kimi-k3 |
| Design | AAVDiff (specialist generator) | 0.116 | 0.180 | +0.064 | 0.194kimi-k3 |
Design Δ is the largest in relative terms: +55% for apodex-1.1 and +67% for the best solver in the environment, over the published SOTA (AAVDiff).
Beaten on the filter
0.904 OOD viability AUROC vs. 0.878 for CAP-PLM — the model tells you which candidates deserve bench time before anything else is computed.
Beaten on the biology
Tissue ranking beats the Fit4Function production model, and 3-D reconstruction beats the best published reference at all three structural levels — AlphaFold 3 with symmetry expansion on the loop and the 60-mer shell, template docking on the complex.
Beaten on generation
Under a 100-query budget, 18% of 1,000 designs are novel, manufacturable and on target — over half again the rate of purpose-built generative models.
Process scores track outcomes rather than presentation. gpt-5.5 scores 0.544 on structure with HDS6 down at 1.38 and 0.94 — it skips the self-validation the task requires and submits the cheapest answer. kimi-k3, the strongest structure solver, earns 3.56 and 3.60 on the same instances.
From benchmark to bench
Top of the leaderboard? We’ll take your designs to the lab
A benchmark score is a proxy for one thing only: whether the candidates are worth making. For the models at the top of this leaderboard, we open the loop — we will reach out to run a joint lab collaboration, synthesize your model’s designs, assay them, and push the discovery task forward together.
This is already running. Apodex model-generated capsid candidates are in preclinical laboratory testing; no human data. It is the same design task, the same viability and selectivity criteria, but measured on real virus instead of a trained oracle. Every measurement that comes back becomes new ground truth inside the environment.
HOW THE COLLABORATION WORKS
You design in silico — your model runs the Task 4 episode: 1,000 novel peptides per target, under the same 100-query oracle budget and the same novelty, manufacturability and on-target gates.
We build and test — a selected panel goes to synthesis, packaging and assay: does it actually assemble, and does it actually reach the target tissue.
Measurements come back — results are returned to you and folded into the environment as held-out ground truth, so the next round is graded against reality rather than a surrogate.
Credit is shared — joint reporting of what worked at the bench, model attributed.
Research collaboration on capsid engineering for gene delivery. Not a clinical claim; no therapeutic use is implied. Designs are screened under the environment’s manufacturability and biosafety gates before anything is synthesized.
Collaboration · at a glance
WHO WE CONTACT
solvers at the top of the Task 4 design leaderboard
WHAT YOU PROVIDE
a model endpoint or an agent that can run the design episode
WHAT WE PROVIDE
synthesis, packaging, viability and tropism assays, and the readouts
CURRENT STATUS
Apodex model-generated candidates in testing now
CONTACT
sheng@apodex.com