Overview
| Name | Dyno Psi-1 / Psi-Phi · aliases: Dyno Psi, DynoPsi, Psi-1, Psi-Phi |
|---|---|
| Full name | No separate expanded name |
| Ambiguous names (not auto-linked) | None |
| Family | Covers the Psi generator series and the companion filter/agent suite Psi-Phi; the wet-lab arm used the earlier Psi-0.5, which is separate from Psi-1
|
| Organization | Dyno Therapeutics |
| Organization type | AI-native biotech |
| Country | United States |
| Model class | Binder backbone generation |
| Domain | Proteins |
| First release | · code release Basis: GitHub v1.0.0 with the whitepaper and press release on the same day (conference launch) Note: Repository creation dates are not used as first release; the press-release date was not independently re-checked source |
| Other milestones | |
| Latest update | · new version Basis: GitHub release 2.0.0 Version 2.0.0: default weights switched to Psi-1.5 Cicada, with breaking API changes source The most recent sourced event, not a claim that nothing changed since. |
| Status | Active · New version and new weights within 12 months |
| Openness | Code: Open · Apache-2.0 Weights: Open · Apache-2.0 Weights and code Apache-2.0; the synthetic training dataset CC BY 4.0 |
| Benchmarks reported | 0 counted · none of the 44 |
| Wet-lab validation | 1 recorded campaign |
| Unverified fields | None |
| As of | 2026-11-16 |
What it does
An open-weight binder design model: a flow-matching backbone generator of about 200M parameters followed by a sequence-design tool, with Psi-Phi as the companion filter and agent suite.
Benchmarks reported
None of the 44 benchmarks. This model is on the zero-pairing list.
How it evaluates itself: In silico binder filter, 5 targets; IL-7RA yeast-display comparison; Press-release comparison
Other evaluations (outside the 44)
- Five-target in-silico filter pass rate (Own task set) Self-reported: Psi-1 computational filtering: 16,900 designs per model across 13 length bins; best on only 2 of 5 targets, i.e. parity not a win; this is not a wet-lab result
- Press-release comparison (Comparison against baselines only) Self-reported: The company states an advantage over several open methods without per-target numbers
Developers report: Developers report that this ~200M open-weight model, with test-time scaling against high-throughput experiments, gives a higher IL-7RA wet-lab success rate than the comparator
Independent evidence: No independent evidence found. No independent head-to-head found
Training data
Group in Figure 7: Public data + wet lab for validation only.
| Class | Value | Detail |
|---|---|---|
| a. Public databases | Yes | Experimental and predicted structure databases plus ~4M synthetic domain-interface dimers (32,253 high-confidence ones used for fine-tuning) |
| b. Private or licensed data | No | |
| c. Own wet-lab data used for training | Not disclosed | The extent of experimental feedback into training is not disclosed |
| d. Wet lab used only for validation | Yes | Yeast display and SPR used for validation only |
Wet-lab validation
Every rate carries its numerator and denominator. Claims that could not be verified are shown as unverified, without numbers; different paper versions are separate entries.
Denominator basis: Initial sampling budget matched (~12k), not design counts: the comparator's full pipeline yielded 125 designs from that budget while these designs were subsampled from the same budget; success was measured on the top 100 designs ranked by ipSAE, with no per-target numbers, so no rate is printed
The wet-lab arm is the earlier Psi-0.5, not Psi-1; developer-run, not an independent replication
Background
- Organization: Dyno Therapeutics is an AI-driven AAV capsid design company founded in Massachusetts in 2018
- Funding: Announced a $100M Series A in 2021-05 led by a16z
- Products and partnerships: Capsid deals with large pharma companies (one licence exercised in 2026-04); the whitepaper was written by the company's AI lab with engineering support from a GPU company
- Related models: La-Proteina, BoltzGen, RFdiffusion2 / RFdiffusion3
Reading notes
- The five-target result is Psi-1 computational filtering, not a five-target wet-lab campaign
- The wet-lab comparison matches sampling budget rather than design counts, and used Psi-0.5
- Success was measured only on the top 100 ranked designs, with no per-target numbers
Caveats
- No results reported on any of the 44 benchmarks
- The press release's advantage claim has no per-target data behind it
Sources
- GitHub repository dynotx/dynopsi (repository)
- GitHub repository dynotx/dynopsi (repository)
- Hugging Face page api/models (repository)
- Official page dynopsi.dynotx.com (company blog)
- Hugging Face page dynotx/dynopsi (repository)
Corrections
2 corrections
Part of the AI for Bio benchmarks survey · all 112 models · open in the explorer. Record as of 2026-11-16.