← Back to main page
Two Proteins Is All You Need INITIALIZING
Test-Time Training for Protein-Protein Interaction Prediction
Extending ProteinTTT to protein pairs for complex structure prediction. We adapt masked-language-model test-time training to jointly refine ESM2 representations of two interacting proteins, improving DockQ scores for predicted protein-protein complex structures.
Last updated: loading...
--
Experiments Run
--
Best DockQ
--
Success Rate
--
GPU Hours

Progress Timeline

Loading...
Fetching progress data...

All Experiments

No experiments recorded yet. Data will appear once experiments begin.

Results

Baselines vs TTT Strategies (Mean DockQ)

Awaiting experiment results...

Per-Dataset Breakdown

Awaiting experiment results...

DockQ Score Distributions

Awaiting experiment results...

Method Overview

ProteinTTT-PPI Pipeline

Input: Protein A (sequence) + Protein B (sequence) | v +------------------+ +-------------------+ | Concatenation | | Linker Strategy | | A + [linker] + B | <-- | poly-G / flexible | +------------------+ +-------------------+ | v +----------------------------+ | ESM-2 (650M) Embedding | | Token representations | +----------------------------+ | v +----------------------------+ | Test-Time Training (TTT) | | Masked LM self-supervision | | on the input pair | +----------------------------+ | v +----------------------------+ | ESMFold Structure Module | | Predict paired structure | +----------------------------+ | v +----------------------------+ | DockQ Evaluation | | Compare to native complex | +----------------------------+

Masking Strategies

Random Uniform
M_KL_AGGGR_DE_K
15% random masking across both chains uniformly
Interface-Biased
MKL___GGG___EKR
Higher masking rate at predicted interface residues
Chain-Alternating
______GGGRDEKAL
Mask one full chain, predict from the other
Span Masking
M___ALGGGRD___K
Contiguous spans of 3-7 residues masked

Key Hyperparameters

Parameter Value Description
modelESM-2 (650M)Base protein language model
ttt_steps100TTT gradient steps per sample
ttt_lr1e-5Learning rate for TTT updates
mask_ratio0.15Fraction of tokens masked
linkerpoly-G (25)Linker sequence between chains
max_length800Max total sequence length (A+linker+B)
num_recycles4ESMFold recycling iterations
batch_size1Per-sample TTT (no batching)

Experiment Logs

ttt-ppi-latest.log
[2026-03-25 12:00:00] INFO ProteinTTT-PPI initialized [2026-03-25 12:00:00] INFO Awaiting experiment logs... [2026-03-25 12:00:00] INFO Logs will auto-refresh every 5 minutes