SAE-Steered Protein Binder Design with Topological Pre-Filtering

We present an interpretability-grounded approach to de novo binder design using sparse autoencoder (SAE) feature decomposition of ESM-2 combined with topological data analysis (TDA) for pre-filtering.

Pipeline: (1) Constrained sequence generation targeting RBX-1 RING domain (res 40-108, E2-binding face), (2) SAE-steered refinement using InterPLM SAE on ESM2-650M layer 18 with helix feature f5922 clamped at 10.0, (3) persistent homology entropy pre-filter (threshold 1.1841, empirically calibrated), (4) Boltz-2 structure prediction with iPTM >= 0.70 cutoff, (5) expression filtering (net charge, hydrophobic stretches, unpaired Cys).

RBX-1 results: 200 parents -> 1000 steered variants -> 359 passed topo filter (35.9%) -> 57 Boltz predictions -> 54 hits (94.7% hit rate) -> 35 passed expression filter. Best iPTM: 0.949. Mean iPTM (final 35): 0.834.

Prior validation on KRAS G12D: 84.3% hit rate (43/51), best consensus iPTM 0.960 +/- 0.012 vs 0.838 +/- 0.016 unsteered baseline. SAE steering produced higher-quality, more reproducible candidates.

Novel contributions: (1) First use of SAE feature clamping as a protein generation steering mechanism, (2) persistent homology entropy as topological pre-filter validated against 4 alternative TDA methods (Cohen's d=0.487), (3) multi-seed consensus ranking. To our knowledge, this is the first competition entry using mechanistic interpretability of a protein language model as the primary design mechanism.

Method PDF: https://drive.google.com/file/d/1RxmzSYQ4GhEuQQVV5pg--INPEgS1ysAR/view?usp=sharing