Description
ORBIT: Oracle-Reseeded Binder Design with Interface Targeting
Team: Aryan Chandak, Surabhi Rathore, Kamalnayan Pathak — Mandrake Bio
DESIGN APPROACH We developed ORBIT, a multi-stage pipeline combining two complementary strategies for de novo binder design.
Strategy 1: Differentiable Binder Hallucination with Spatially-Biased Interface Optimization We optimize sequence probability distributions (PSSMs) via backpropagation through Protenix, a structure prediction model. Starting from a uniform distribution over 19 amino acids (excluding cysteine), we minimize a composite loss combining ipTM, pLDDT, PAE, weighted contact, and inverse folding terms. Optimization uses simplex-projected accelerated gradient descent in three phases: 100 steps of soft optimization, 50 steps of logarithmic sharpening, and 15 steps of final refinement. Target features include the 2LGV NMR template with explicit zinc ions.
For epitope targeting, we introduce per-residue contact weights on the target surface: weight 2 for E2/GLMN-binding RING residues, weight 1 for hinge and free-surface residues, and weight 0 for IDR, Cullin-face, and zinc-buried residues. We explored fixed weighting, heavy weighting (4x), and cosine-annealed scheduling where the weight scale decays from 2x to 1x over the trajectory, preventing gradient instability during sequence sharpening.
Strategy 2: Structure-Informed Reseeding Single-round optimization often converges to local minima. We introduce an oracle-in-the-loop cycling strategy: (1) run 20 steps of differentiable hallucination to get a predicted complex structure, (2) pass the binder chain to SolubleMPNN to extract sequence log-probabilities, (3) temperature-scale these into a soft PSSM, and (4) restart optimization from this structure-informed prior. The temperature controls exploration-exploitation: T=0.01 (70% of designs), T=0.1 (15%), T=0.5 (10%), T=1.0 (5%). This distribution was calibrated to maximize hit rate. Reseeded designs consistently outperformed cold-start baselines.
Trajectory-Based Early Termination We trained a logistic regression classifier on optimization trajectory features (ipTM, pLDDT, PAE, contact metrics) at step 60 of 165 total steps. The classifier identifies designs that will ultimately fail with 97.7% accuracy and zero false negatives, enabling early termination of unpromising trajectories and saving ~40% of compute.
FILTERING PIPELINE Candidates pass through five independent quality gates:
- Protenix ipSAE (computed from PAE matrices, noise-robust interface confidence)
- Boltz-2 cross-validation (independent structure prediction confirms binding)
- Rosetta energetics (dG, shape complementarity, BSA, hydrogen bonds)
- Monomer fold stability (binder must fold independently with pLDDT > 0.85 and maintain structural consistency upon binding with RMSD < 4 Å)
- Expressibility filtering (candidates with >55% lysine+glutamate content in helical regions are deprioritized due to known translation failure modes)
We decompose interface PAE into IDR PAE (residues 1-36, always high, noise) versus RING PAE (residues 37+, the meaningful binding signal). This is critical for zinc-finger targets where large disordered regions inflate standard ipTM scores.
RESULTS From 74 designed candidates across warm-start, baseline, and weighted epitope configurations, progressive filtering yielded 18 final candidates. All are 100-residue binders. Mean ipSAE is 0.804 with RING PAE averaging 3.7 Å. 13 of 18 have Boltz-2 ipTM > 0.5 confirming cross-model agreement. Mean monomer pLDDT is 0.958. K+E content ranges from 20-49%, with two candidates in the expressibility-safe zone (<25%).
The structure-informed reseeding strategy produced 10 of 18 final candidates, confirming its superiority over cold-start baselines (4/18) and weighted epitope variants (4/18). Candidates were ranked using a composite of structural confidence (ipSAE), biophysical quality (dG, BSA, SC, H-bonds), epitope targeting (E2 face fraction), and expressibility (K+E content).