---
title: RXNFP inputs and parameters
description: 'Reaction SMILES format and parameters for fingerprinting and similarity search.'
tags:
  - rxnfp
  - inputs
---

You provide one or more reaction SMILES in `reactants>>products` form. Two parameters shape the result: `fingerprint_type`, which decides what a fingerprint *is* and therefore what similarity even means, and `min_similarity`, which decides how much of the tail you see. Raise `min_similarity` for close matches only, lower it to widen the set.

## Reaction SMILES format

Input reactions in the standard SMILES format:

```
reactants>>products
```

Multiple reactants are separated by `.` (dot).

### Examples

| Reaction | SMILES |
|----------|--------|
| Esterification | `CC(=O)O.OCC>>CC(=O)OCC` |
| Amide formation | `CC(=O)Cl.c1ccccc1N>>CC(=O)Nc1ccccc1` |
| Suzuki coupling | `c1ccc(B(O)O)cc1.c1ccc(Br)cc1>>c1ccc(-c2ccccc2)cc1` |

## Fingerprinting parameters (`rxnfp` / `rxnfp_fingerprint`)

| Parameter | Type | Required | Description |
|-----------|------|----------|-------------|
| `reaction_smiles` | string or string[] | Yes | One reaction SMILES, or an array for batch fingerprinting. |
| `fingerprint_type` | `rxnfp` \| `drfp` | No | Which fingerprint to compute. Defaults to `rxnfp`. |

## Choosing a fingerprint type

The two types are different objects, not two settings of one knob, and this is the choice that decides what everything downstream can do.

| | `rxnfp` (default) | `drfp` |
|---|---|---|
| What it is | A 256-dimensional learned RXNBERT embedding | A 2048-bit differential reaction fingerprint, radius 3 |
| Values | Floats | Binary 0/1 |
| Needs a model | Yes | No |

⛔ **Scores from the two types are not comparable.** `rxnfp` is scored by cosine distance on the embeddings; `drfp` by Tanimoto on the bit vector. A 0.8 under one does not mean what a 0.8 under the other means, so do not mix them in one ranking or compare a stored score against a run that used the other type.

For nearest-neighbour class retrieval the source reports `rxnfp` as the stronger of the two — USPTO 1k TPL 5-NN accuracy 0.989 against 0.917 for `drfp`. Reach for `drfp` when you want a model-free, reproducible bit vector rather than the best retrieval.

## Similarity search parameters (`rxnfp_similarity`)

| Parameter | Type | Required | Default | Description |
|-----------|------|----------|---------|-------------|
| `reaction_smiles` | string | Yes | -- | A single query reaction SMILES. |
| `top_k` | integer | No | `10` | Number of similar reactions to return (1–100). |
| `min_similarity` | number | No | `0.7` | Minimum similarity to include a match (0.0–1.0). |
| `fingerprint_type` | `rxnfp` \| `drfp` | No | `rxnfp` | Which fingerprint and metric to search with — cosine on embeddings for `rxnfp`, Tanimoto on bits for `drfp`. |
| `reference_reactions` | string[] | No | -- | Reaction SMILES to search **instead of** the server-side reference index. Up to 1,000. |

The query itself is a single reaction — `reaction_smiles` here is one string, not an array, unlike the fingerprinting input.

By default the query is looked up against a pre-built reference index drawn from ORDerly reaction data. Pass `reference_reactions` to search your own set instead: that is how you compare a query against a specific series — a project's own reaction history, or the steps of one retrosynthetic route — rather than against the whole index. It replaces the index for that call; it is not searched in addition to it.

## Batch fingerprinting

```json
{
  "reaction_smiles": [
    "CC(=O)O.OCC>>CC(=O)OCC",
    "CC(=O)Cl.c1ccccc1N>>CC(=O)Nc1ccccc1",
    "c1ccc(B(O)O)cc1.c1ccc(Br)cc1>>c1ccc(-c2ccccc2)cc1"
  ]
}
```

## Input validation rules

- Reaction SMILES must contain the `>>` separator.
- Both reactant and product sides must be valid SMILES.
- Very long reaction SMILES are truncated at the tokenizer's limit.
- A malformed reaction in a batch is reported per-item (`is_valid: false`, with an `error` message) — it does not fail the whole batch.

## Read next