Beam Search Configuration Sweep

Date: 2026-04-05  |  Configs: 5  |  Proteins: 10  |  Samples: 26,560 (1,600 finals)  |  Runtime: ~93 min on 8x H100  |  PRODUCTION (400 steps)

Key Findings

1. E_fine_selector dominates. Won 7/10 proteins by best single sample and 7/10 by mean reward. Its frequent pruning (6 checkpoints, 3 branches) consistently finds the best binders.
2. D_early_brancher drops to 4th with full data (-0.2792). Its earlier lead was biased by only completing easier proteins. Frequent pruning beats coarse exploration.
3. B_length_explorer is worst (-0.3501 mean reward). 16 lengths with 2 beams each lacks depth. Length diversity doesn't compensate for shallow search.
4. Best binding prediction: IFNAR2 with i_PAE = 0.130 (D_early_brancher). Hardest target: BetV1 at best i_PAE = 0.205.
5. Lookaheads beat finals in all configs - the search is not optimally selecting intermediate candidates. Room for improvement in pruning/scoring.

Config x Protein Heatmaps

Mean i_PAE across final samples (lower = better binding). Missing cells = experiment failed.
Mean i_PAE across final samples (lower = better binding). Missing cells = experiment failed.
Best single-sample i_PAE. Gold border = winner for that protein target.
Best single-sample i_PAE. Gold border = winner for that protein target.

Configuration Rankings

Left: mean reward. Center: wins by best sample. Right: wins by mean reward.
Left: mean reward. Center: wins by best sample. Right: wins by mean reward.
RankConfigStrategyMean RewardWins (best)Wins (mean)
1E_fine_selector6 checkpoints, 3 branches-0.247577
2A_balancedPaper default-0.257102
3C_deep_exploiter2 lengths, 16 beams-0.270011
4D_early_brancher2 checkpoints, 8 branches-0.279220
5B_length_explorer16 lengths, 2 beams-0.350100

Per-Protein Comparison

Best single-sample i_PAE by protein target, grouped by config.
Best single-sample i_PAE by protein target, grouped by config.

Protein Difficulty

Easiest to hardest protein targets. Green < 0.16 (easy), Orange 0.16-0.22 (medium), Red > 0.22 (hard).
Easiest to hardest protein targets. Green < 0.16 (easy), Orange 0.16-0.22 (medium), Red > 0.22 (hard).
DifficultyProteinBest i_PAEBest Config
EasyIFNAR20.130D_early_brancher
EasyPD10.138E_fine_selector
EasyPDL10.150C_deep_exploiter
EasyCrSAS60.156E_fine_selector
MediumCD450.159E_fine_selector
MediumSpCas90.168E_fine_selector
MediumDerF70.175E_fine_selector
MediumClaudin10.198C_deep_exploiter
HardBetV10.205E_fine_selector
HardHER2_AAV0.222E_fine_selector

Compute Budget

Total PDBs generated vs finals kept per experiment. E_fine_selector generates the most (608) but wins the most.
Total PDBs generated vs finals kept per experiment. E_fine_selector generates the most (608) but wins the most.

Next Steps

  1. Add refinement to E_fine_selector: sequence_hallucination is expected to dramatically improve pLDDT and i_PAE
  2. Fix D_early_brancher failures: Reduce n_branch from 8 to 6 to avoid OOM on larger targets
  3. Full pipeline on top candidates: filter → evaluate → analyze for designability (scRMSD, ProteinMPNN)
  4. Hybrid config: Combine E's frequent pruning with D's wide branching (n_branch=6, 4 checkpoints)
Appendix: Configuration Details
Confignsamplesbeam_widthn_branchCheckpointsFinalsTotal PDBs
A_balanced484[0,100,200,300,400]32544
B_length_explorer1624[0,100,200,300,400]32544
C_deep_exploiter2164[0,100,250,400]32416
D_early_brancher488[0,200,400]32544
E_fine_selector483[0,65,130,200,270,340,400]32608

All: 400 diffusion steps, beam-search, AF2 multimer reward (i_pae=-1.0), batch_size=8, seed=5

Appendix: Execution Summary
Appendix: Per-Config Per-Protein Mean Reward
ProteinA_balancedB_length_expC_deep_expD_early_brE_fine_sel
PD10.000-0.2380.000-0.161-0.156
PDL1-0.184-0.210-0.172-0.191-0.176
IFNAR2-0.148-0.154-0.150-0.148-0.146
CD45-0.244-0.579-0.274-0.361-0.365
Claudin1-0.225-0.252-0.226-0.224-0.212
CrSAS6-0.178-0.266-0.190-0.187-0.178
DerF7-0.200-0.262-0.202-0.201-0.192
BetV1-0.790-0.650-0.893-0.760-0.598
SpCas9-0.316-0.398-0.306-0.266-0.213
HER2_AAV-0.288-0.493-0.287-0.294-0.239

Generated by Analyst Team — 2026-04-05 — Full analysis (markdown)