soldatmat/CZAI_Summer_School-MLDE_landscapes

CZAI Summer School 2026 — MLDE / BOES landscape data

Precomputed data backing the BOES (Bayesian Optimization in Embedding Space) directed-evolution demo notebook for the CZAI Summer School 2026 lecture "AI pro enzymy". See the notebook `CZAI_Summer_School-MLDE_BOES.ipynb` for how this data is used.

This repo hosts, for three combinatorial 4-site protein fitness landscapes (GB1, PhoQ, TrpB):

• —*_fitness_full.csv — the full variant,fitness table, always exactly 160,000 rows — the complete 20^4 combinatorial grid over the 4 mutated positions (all combinations of the 20 standard amino acids). Variants that were directly, experimentally measured in the original study keep their real fitness value; every variant that was never directly measured is assigned a fitness of 0.0 (the assumption being: if it wasn't detected in the selection experiment, it didn't survive). See "Source data" below for how many variants are measured vs. zero-filled per landscape, and note the deliberate simplification for GB1.
• —*_superset_fitness.csv / *_superset_variant_order.json / *_superset_embeddings.npy — a fixed, seeded random subsample of 5000 variants drawn from the full 160,000-row table above (always including the wild type and the landscape's true global fitness maximum) together with their precomputed ESM-2 (facebook/esm2_t12_35M_UR50D) mean-pooled sequence embeddings (float32, shape (5000, 480), row-aligned with the variant order file). The notebook subsamples further from this superset at run time (seeded, so it's reproducible) to build the small demo pool used live in the lecture.
• —*_full_embeddings.npy / *_full_variant_order.json — precomputed ESM-2 (facebook/esm2_t12_35M_UR50D) mean-pooled sequence embeddings for every variant in the full 160,000-row combinatorial domain (float32, shape (160000, 480), row-aligned with the variant order file; same checkpoint and mean-pooling convention as the superset embeddings above, verified to match on overlapping variants). These back the notebook's optional USE_FULL_DOMAIN=True mode, which runs all three demo methods (random, zero-shot, BOES) against the full landscape instead of the 3,000-item demo pool, for a slower but bigger live BOES run. Available for all three landscapes (GB1, PhoQ, TrpB); computing these offline took ~5 min (GB1, 56-residue wild type) up to ~60-75 min (TrpB and PhoQ, 397- and 486-residue wild types respectively) on a CPU/MPS Mac, since attention cost scales worse than linearly with sequence length.
• —landscape_meta.json — wild-type full sequence, 1-indexed mutated positions, wild-type variant code, and the true global-maximum variant and fitness for each landscape (computed from the full 160,000-row table, before any subsampling).
• —{GB1,PhoQ,TrpB}_CITE_AS.txt — citation for each original fitness landscape dataset (see below).

Embeddings were computed once, offline, so the live lecture notebook only downloads ready-made files instead of recomputing them during the session.

Source data

For all three landscapes, "measured" means directly, experimentally measured in the cited study; every other one of the 160,000 possible 4-letter combinations is zero-filled (see above).

• —GB1: 149,361 variants directly measured (Fitness column of elife-16965-supp1.xlsx, Wu et al. 2016) keep their real fitness; the remaining 10,639 variants are zero-filled. Note: the original paper also publishes a second file, elife-16965-supp2.xlsx, with model-imputed fitness values for exactly those 10,639 variants (too low read-count to measure directly) — we deliberately do not use those imputed values here, to keep a single, simple, consistent "not measured → presumed dead" rule across all three landscapes rather than relying on a third-party imputation model for only one of them. Zero-fraction of the full table: ~25.1%.
• —PhoQ: 140,517 variants directly measured (PhoQ.xlsx of Podgornaia & Laub 2015) keep their real fitness (many of which are themselves exactly 0.0 — a real measured "dead" phenotype, not zero-fill); the remaining 19,483 never-measured variants are zero-filled. Zero-fraction of the full table: ~68.1% (real measured zeros + zero-filled unmeasured combined).
• —TrpB: fitness.csv from Johnston et al. 2024 is already combinatorially complete — all 160,000 combinations were directly measured (its own paper's title: "A combinatorially complete epistatic fitness landscape...") — verified by checking that its 160,000 variant codes exactly match the full 20^4 grid, with no duplicates or gaps. No zero-filling was needed or applied; ~22.6% of its (genuinely measured) values happen to be exactly 0.0.

All three source files, plus the reference Julia implementation of BOES, come from the public research repository soldatmat/PELLM (read-only reference for this demo; not modified).

Citations

BOES method:

@article{soldat2025boes,
  title={Directed Evolution of Proteins via Bayesian Optimization in Embedding Space},
  author={Sold{\'a}t, Matou{\v{s}} and Kl{\'e}ma, Ji{\v{r}}{\'i}},
  journal={arXiv preprint arXiv:2509.04998},
  year={2025}
}

GB1:

@article{wu2016adaptation,
  title={Adaptation in protein fitness landscapes is facilitated by indirect paths},
  author={Wu, Nicholas C and Dai, Lei and Olson, C Anders and Lloyd-Smith, James O and Sun, Ren},
  journal={Elife},
  volume={5},
  pages={e16965},
  year={2016},
  publisher={eLife Sciences Publications, Ltd}
}

PhoQ:

@article{podgornaia2015pervasive,
  title={Pervasive degeneracy and epistasis in a protein-protein interface},
  author={Podgornaia, Anna I and Laub, Michael T},
  journal={Science},
  volume={347},
  number={6222},
  pages={673--677},
  year={2015},
  publisher={American Association for the Advancement of Science}
}

TrpB:

@article{johnston2024combinatorially,
  title={A combinatorially complete epistatic fitness landscape in an enzyme active site},
  author={Johnston, Kadina E and Almhjell, Patrick J and Watkins-Dulaney, Ella J and Liu, Grace and Porter, Nicholas J and Yang, Jason and Arnold, Frances H},
  journal={Proceedings of the National Academy of Sciences},
  volume={121},
  number={32},
  pages={e2400439121},
  year={2024},
  publisher={National Academy of Sciences}
}