PD-L1 Binder Demo Analysis

Date: 2026-04-03  |  Target: PD-L1 (02_PDL1)  |  Stage: Generate only  |  DEMO RUN

Key Findings

1. Pipeline works end-to-end. Generated 2 binder candidates for PD-L1 in 82.92 seconds on 1x H100. AF2 reward scoring completed successfully.
2. Both samples fail production thresholds — expected at 100 steps with no refinement. i_PAE × 31 = 25.76 and 27.32 (need ≤ 7.0). pLDDT = 0.212 and 0.239 (need ≥ 0.9).
3. Sample 0 (n=262) is the better candidate: lower i_PAE (0.831 vs 0.881), higher i_pTM (0.147 vs 0.068), much lower RMSD (11.25Å vs 45.21Å).
4. Reward = -i_PAE only. All other reward weights (con, plddt, dgram_cce, etc.) are 0.0. Consider multi-objective reward for production runs.

Metrics Comparison

Key AF2 metrics for both generated binder candidates
Key AF2 metrics for both generated binder candidates

Quality Assessment

Demo values (blue/orange) vs production pass-rate thresholds (red dashed)
Demo values (blue/orange) vs production pass-rate thresholds (red dashed)
Criterion Sample 0 (n=262) Sample 1 (n=234) Threshold Status
i_PAE × 31 ≤ 7.0 (binding) 25.76 27.32 ≤ 7.0 FAIL
pLDDT ≥ 0.9 (structure) 0.212 0.239 ≥ 0.9 FAIL
scRMSD < 1.5Å (designability) N/A (evaluate stage not run) < 1.5Å —

Sample Profile

Normalized radar chart comparing structural quality metrics across both samples
Normalized radar chart comparing structural quality metrics across both samples

Timing & Configuration

Left: generation timing. Right: demo config vs recommended production settings
Left: generation timing. Right: demo config vs recommended production settings
ParameterDemoProductionImpact
Diffusion steps1004004× better convergence
Samples232–64Cover binder length space
Replicas14–16More candidates per length
Searchbest-of-nbeam-searchSmarter exploration
RefinementNonesequence_hallucinationKey to SOTA quality
GPUs1× H1004–8× H100Parallel generation

Next Steps

  1. Production run: 400 steps, 32 samples, beam-search, sequence_hallucination refinement on PD-L1
  2. Full pipeline: Add filter → evaluate → analyze for designability and pass-rate metrics
  3. Multi-target: Run on PD-1, IFNAR2, CD45 to compare difficulty across targets
  4. Search algorithm sweep: Compare best-of-n vs beam-search vs MCTS
Appendix: Full Per-Sample Metrics (49 columns)
MetricSample 0 (n=262)Sample 1 (n=234)
total_reward-0.831-0.881
af2folding_i_pae0.8310.881
af2folding_plddt0.2120.239
af2folding_i_ptm_log0.1470.068
af2folding_ptm_log0.4550.476
af2folding_con2.2611.715
af2folding_i_con5.1955.086
af2folding_rmsd11.24745.209
af2folding_pae0.5590.523
af2folding_min_ipae0.5680.727
af2folding_fape214.727193.455
af2folding_dgram_cce287.986511.127
af2folding_exp_res0.0630.009
af2folding_seq_ent0.00.0
af2folding_recycles_log3.03.0
binder_length262234
sample_typefinalfinal
total_time82.92s (2 samples)
Appendix: Generation Config 
Target: PD-L1 (02_PDL1)
Target PDB: assets/target_data/bindcraft_targets/PD-L1.pdb
Target chain: A1-115
Hotspots: A37, A39, A49, A98
Binder length range: 64–155

Model: complexa.ckpt (v1 architecture)
Autoencoder: complexa_ae.ckpt
nsteps: 100
self_cond: True
guidance_w: 1.0
seed: 5
batch_size: 16
search: best-of-n (1 replica)
refinement: None

Reward: AF2 multimer
  i_pae weight: -1.0
  all others: 0.0
  num_recycles: 3
  use_initial_guess: True

Generated by Analyst Team — 2026-04-03 — Full analysis (markdown)