README.md

---
license: cc-by-4.0
pretty_name: SynthonBench
tags:
  - chemistry
  - cheminformatics
  - drug-discovery
  - virtual-screening
  - combinatorial-chemical-space
  - synthon
  - benchmark
  - docking
size_categories:
  - 100M<n<1B
task_categories:
  - tabular-regression
configs:
  - config_name: glide_1M
    default: true
    data_files: scores/glide_1M.parquet
  - config_name: surrogate_1M
    data_files: scores/surrogate_1M.parquet
  - config_name: surrogate_10M
    data_files: scores/surrogate_10M.parquet
  - config_name: surrogate_100M
    data_files: scores/surrogate_100M.parquet
  - config_name: seeds
    data_files: seeds/*.parquet
---

# SynthonBench — data

Frozen tasks, score tables, synthon spaces and shared warm-start seeds for
**SynthonBench**: a benchmark for *budgeted search over synthon combinatorial
(make-on-demand) chemical spaces*.

> Search methods over synthon spaces never enumerate the full product library;
> they propose **reaction-component tuples** `(reaction_id, synthon_ids)` and pay
> a fixed oracle-call budget. SynthonBench scores such methods *fairly and
> reproducibly* against frozen score tables and the **top-k DCRF** efficiency
> metric (random search ≈ 1.0).

Companion code: [`synthonbench`](https://github.com/mireklzicar/synthonbench)
(`pip install synthonbench`). The package downloads this dataset for you:

```bash
pip install "synthonbench[hf]"
synthonbench download --scale 1M            # spaces + glide_1M + surrogate_1M + seeds
synthonbench download --scale all           # everything, including the 100M table (~5 GB)
```

## Targets & scales

Three Glide docking targets — **KIF11**, **PYRD**, **TGFR1** — over the same 42
canonical reactions, at three scales of the synthon product space:

| scale | products | unique synthons |
|-------|----------|-----------------|
| 1M    | 990,610  | 6,270  |
| 10M   | 10,022,100 | 13,880 |
| 100M  | 99,851,700 | 31,993 |

## Files

```
spaces/
  synthon_space_{1M,10M,100M}.synthons.tsv    # smiles  synthon_id  position  reaction_id
  synthon_space_{1M,10M,100M}.properties.csv  # synthon_id + synthetic per-synthon attributes
  reactions.tsv                               # 42 canonical reactions (SMIRKS) + reagent patterns
scores/
  glide_1M.parquet          # REAL Glide @1M, headline docking oracle (LE/medchem-filtered), all 3 targets
  glide_1M.parquet.sidx/    # mmap product_id -> score index (bounded-memory oracle lookup)
  glide_1M.parquet.topk/    # top-100k product cache per score column (exact top-k for DCRF)
  surrogate_1M.parquet      # LightGBM docking surrogate, all 3 targets
  surrogate_1M.parquet.sidx/    surrogate_1M.parquet.topk/
  surrogate_10M.parquet     (+ .sidx/ .topk/)
  surrogate_100M.parquet    # ~4.6 GB  (+ .sidx/ ~3.8 GB, .topk/)
seeds/
  seeds_<target>_s{0,1,2}.parquet   # 3 shared 100k warm-start pools per target (one per protocol seed)
```

### Sidecars (`*.parquet.sidx`, `*.parquet.topk`)

Large score tables are **not** loaded into a Python `{product_id: score}` dict.
Each table ships two bounded-memory sidecars that `synthonbench` reads directly:

- **`.sidx`** — a memory-mapped `product_id → score` index. `synthonbench.load_oracle`
  returns an `IndexedScoreTableOracle` backed by it, so oracle lookups never
  materialise the table.
- **`.topk`** — the best 100k products per `(score column, direction)`, so top-k
  recall / DCRF and the reported docking-score means need no full-table sort.

`synthonbench download` fetches these alongside each parquet automatically. If you
stage a parquet by hand, rebuild them with
`synthonbench build-score-index` and `synthonbench build-topk-cache`
(the glide table needs `_glide_le`@maximize + `_glide_raw`@minimize; the surrogate
tables need `_predicted_dock_score`@minimize, for each of the three targets).

### `scores/glide_1M.parquet` (headline real-Glide oracle)

One row per product (990,610). Keyed by `product_id` (md5 of the assembled
product) which also keys every surrogate table.

| column | meaning |
|--------|---------|
| `product_id`, `reaction_id` | product id and its reaction |
| `smiles`, `canonical_smiles` | assembled product |
| `<target>_glide_raw` | raw Glide docking score |
| `<target>_glide_le` | **benchmark score**: medchem-filtered, lipophilicity-penalized (−Glide−cLogP) ligand-efficiency score; `0.0` if the product fails the medchem filter or docking |
| `<target>_le_ratio`, `<target>_valid` | ligand efficiency and validity flag |
| `passes_exp27_exact_filter`, `exp27_filter_reason` | medchem filter outcome |
| `mol_wt`, `mol_logp`, `hba`, `hbd`, `rotatable_bonds`, `fraction_csp3`, `tpsa`, `aromatic_rings`, `fluorine_count`, `amide_count` | RDKit physchem descriptors |

`<target>` ∈ {`kif11`, `pyrd`, `tgfr1`}.

### `scores/surrogate_{1M,10M,100M}.parquet`

| column | meaning |
|--------|---------|
| `product_id`, `reaction_id`, `smiles` | product identity |
| `synthon1_id`, `synthon2_id`, `synthon3_id` | the synthon tuple (`synthon3` null for 2-component reactions) |
| `kif11_predicted_dock_score`, `pyrd_predicted_dock_score`, `tgfr1_predicted_dock_score` | LightGBM surrogate scores |

### `seeds/seeds_<target>_s{0,1,2}.parquet` (shared warm-start pools)

Three independent uniform-random 100k samples of the 1M product space per
target — one per protocol seed `0,1,2`. Every method may **initialize from the
same pool** so warm-start does not confound the comparison. Columns: `target`,
`seed`, `product_id`, `reaction_id`, `synthon1_id`, `synthon2_id`,
`synthon3_id`, `smiles`, `surrogate_score`, `glide_le_score`. (The maximum
benchmark budget is 100k oracle calls, so a full seed pool is itself within
budget.)

## Quick load

```python
from datasets import load_dataset
glide = load_dataset("mireklzicar/synthonbench", "glide_1M", split="train")
seeds = load_dataset("mireklzicar/synthonbench", "seeds", split="train")
```

or directly with pandas / pyarrow:

```python
import pandas as pd
df = pd.read_parquet("hf://datasets/mireklzicar/synthonbench/scores/glide_1M.parquet")
```

## Provenance & licensing

Synthon spaces are derived from publicly described make-on-demand reaction
schemes; real-Glide scores were produced with Schrödinger Glide and the
LightGBM surrogate was trained on them. The **scores and synthon tables**
released here are redistributable under CC-BY-4.0. Docking/space construction
*software* (Glide, BiosolveIT FTrees/CoLibri, SpaceHASTEN) is **not** included
and requires the respective vendor licenses; see the code repository for how
external methods are wired in optionally.

## Citation

See [`CITATION.cff`](https://github.com/mireklzicar/synthonbench/blob/main/CITATION.cff).