dynotx/dynopsi

Dyno Psi Overview

Dyno Psi-1 is a generative protein design model for de novo binder design. For more information, please see our white paper.

Description

The Dyno Psi approach to binder design consists of a backbone generation model (Dyno Psi-1), a sequence design component, and in silico filters. Dyno Psi-1 samples new protein backbones via a flow-matching-based denoising process. Sequences are designed post hoc using ProteinMPNN, a state-of-the-art inverse folding model, and these candidates are filtered using a combination of physics-based and refolding confidence metrics.

This Hugging Face repository and corresponding GitHub repository enable the use of the Dyno Psi-1 backbone generation model.

Architecture & Training

Dyno 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.

Dyno Psi-1 was trained in multiple stages to improve its ability to generalize. Training began on a large-scale dataset of experimentally-resolved structures from the PDB and high-quality structure predictions from the AlphaFold Database (AFDB). The second stage added protein–protein interactions from the PDB as well as ~1 million cluster representatives of a custom curated set of synthetic domain-domain interaction pairs derived from AFDB monomers. This mixture of experimental and synthetic interaction data enables the model to learn diverse geometric and interface features relevant to de novo binder generation.

Training Phases

Phase	Steps	GPUs	Max Tokens	Data
Monomer pretraining	232k	16× H200	256	PDB monomers, AFDB monomers
Interface-aware pretraining	143k	96× H200	512	PDB monomers, AFDB monomers, PDB multimers, ~ 4M synthetic intra-domain interaction pairs (~1M clusters)
Binder design fine-tuning	154k	96× H200	512	PDB complexes, AF2-filtered synthetic interaction pairs