Psi-1.5 Cicada
Overview
Psi is a family of generative protein design models for de novo binder design, developed by Dyno Therapeutics. For more information, see the Psi-1 white paper.
Cicada (v1.5-cicada, dynopsi-1.5-cicada.ckpt) is the default Psi model and the default checkpoint of the dynopsi package (pip install dynopsi). When checkpoint and --checkpoint are omitted, this checkpoint is loaded. Psi-1 (dynopsi-1.ckpt) is selected with checkpoint: v1 or --checkpoint v1.
Description
The Psi approach to binder design consists of a backbone generation model, a sequence design component, and in silico filters. Cicada and Psi-1 sample new protein backbones via a flow-matching-based denoising process. Sequences are designed post hoc using ProteinMPNN, and these candidates are filtered using a combination of physics-based and refolding confidence metrics.
Architecture
Psi-1 is a ~200M-parameter non-equivariant transformer neural network. The architecture adapts key elements from the Proteina model, including triangle updates, conditioning via adaptive layer norms, and pair-biased multi-head attention, to the binder design setting. Cicada is a ~120M-parameter neural network that is faster and more efficient than Psi-1, owing to its reduced parameter count, removal of expensive triangle updates, and cheaper self-conditioning. Cicada adds chain_is_loopy and has_homooligomer_partner conditioning inputs, which Psi-1 does not use. The recommended initial values are NON_LOOP for chain_is_loopy and NO_PARTNER for has_homooligomer_partner. Psi-1 declares neither feature, and design_conditioning is omitted for that checkpoint. A mismatch between checkpoint and conditioning inputs raises an error.
Training
Psi models were trained on multiple data sources: experimentally resolved structures from the PDB, structure predictions from the AFDB, protein–protein interactions from the PDB, and ~1 million cluster representatives of a custom curated set of synthetic domain-domain interaction pairs derived from AFDB monomers. Psi-1 was trained in multiple phases. Training of Cicada focused on binder design.
License
Psi weights in this repository are released under the Apache-2.0 license. The repository is public, and downloading a file does not require an access request.
Limitations
Generated binders contain backbone coordinates. Side-chain atoms are not generated for the binder; target side chains are retained from the input structure. Sequence design is performed separately.
Responsible use
Psi models are not validated therapeutics. Experimental testing is required before any biological use.