After three beam-search runs with identical beam_width=8, n_branch=4, nsamples=4 and standard checkpoints, best i_pae stalled at ~0.158 and top candidates have low pLDDT (0.04–0.09). Hypothesis: expanding length coverage (increase nsamples) may find better-length binders and improve i_pae; consider increasing nsamples to 8 while keeping other params constant for the next run.
Increasing nsamples from 4 to 8 did not improve best i_pae (worsened from 0.1559 to 0.1714) and pLDDT remains very low (~0.04–0.06), suggesting structural confidence is poor. Hypothesis: the search is stuck in similar low-confidence modes; increasing beam width to 16 to increase sequence/structure diversity may find higher-affinity, higher-confidence binders. Monitor for improved i_pae and pLDDT; if no improvement, consider changing branching (n_branch) or switching search_algorithm.
Increasing beam_width from 8 to 16 improved best i_pae from 0.1448 to 0.1387, indicating broader beam search yields better interface predictions. nsamples increase to 8 caused CUDA OOM at current batch size/config, so we should avoid raising nsamples without reducing batch_size or memory usage. Next experimental variable to try: increase branching factor (n_branch) to 8 to increase diversity while keeping beam_width at 16 and nsamples at 4.
Fine-grained step_checkpoints ([0,65,130,200,270,340,400]) combined with sampling more binder lengths (nsamples=8) and beam-search (beam_width=8, n_branch=4) produced the best binders for 01_PD1 (best i_pae=0.133). Increasing beam_width or branching (beam_width=16 or n_branch=8) caused CUDA OOM on the available GPU. MCTS did not outperform beam-search. Recommendation: treat the beam-search, fine-checkpoint, nsamples=8 configuration as the baseline for further local refinement; consider increasing nsamples to 16 or exploring small algorithmic tweaks only if more improvement is required.
Best result so far: beam_width=16, n_branch=4, nsamples=4, nsteps=400, fine checkpoints produced best_i_pae=0.13441561 (run_20260406_034234_02_PDL1). However pLDDT remains low (~0.04–0.07). Increasing n_branch to 8 caused CUDA OOM. Budget has reached 10 runs; stop launching new experiments. Recommended next steps: (1) take top candidates from run_20260406_034234_02_PDL1 for structural refinement (AF refinement/relaxation) and scoring; (2) if budget is extended, test search_algorithm=fk-steering or mcts, or try beam_width=16 with n_branch=8 but smaller batch_size to fit memory, aiming to increase diversity and pLDDT.
Best run (run_20260406_034234_02_PDL1) achieved i_pae=0.1344 with beam-search (beam_width=16, n_branch=4, nsamples=4, fine checkpoints), but pLDDT remains very low (~0.04–0.07). Next recommended actions: (1) perform structural refinement (AF-relax/refinement) on top candidates to improve pLDDT and re-score; (2) if running another design, change search algorithm (try fk-steering or mcts) rather than repeating beam-search; avoid increasing n_branch beyond 4 due to OOM risk.
Increasing beam_width from 8->16 and n_branch from 4->8 improved best i_pae from 0.1448 to 0.13016. Attempting nsamples=8 previously caused CUDA OOM; keep nsamples=4 unless batch_size is reduced. Next planned change: use fine step_checkpoints [0,65,130,200,270,340,400] to give more branching decision points and refine interface sampling.
Recent attempts using fine step_checkpoints caused CUDA OOMs (batch_size=8 and batch_size=4 failed). Memory pressure likely from increased branching checkpoints. Next attempt: keep beam_width=16 and n_branch=8 but reduce batch_size to 2 to fit GPU memory. If this still OOMs, consider reducing beam_width or n_branch or switching to standard checkpoints.
For 05_CD45, using beam_width=16 can improve i_pae (best so far 0.173) but requires a small generation batch_size to avoid CUDA OOM. Increasing n_branch to 8 caused OOM with batch_size=8; similarly nsamples=8 with beam_width=16 and batch_size=8 also OOM'd. Recommendation: keep beam_width=16 but reduce batch_size (e.g., 2) or reduce n_branch to 2–3 to fit GPU memory. Also, fine checkpoints improved best scores in earlier runs.