Metanova/SAVI-2020

Synthetically Accessible Virtual Inventory (SAVI) 2020

Dataset Description

Dataset Summary

The Synthetically Accessible Virtual Inventory (SAVI) 2020 dataset is a comprehensive collection of over 1.5 billion computationally generated organic compounds designed to be easily and practically synthesizable. Created through expert-system type rules derived from established organic chemistry knowledge, SAVI represents one of the largest publicly available databases of synthetically accessible virtual compounds for drug discovery and chemical research.

The database was generated using 53 carefully selected chemical transformation rules encoded in the CHMTRN/PATRAN programming languages, originally developed for the LHASA (Logic and Heuristics Applied to Synthetic Analysis) retrosynthetic analysis system. These transforms were applied to approximately 152,532 commercially available building blocks from Enamine to create molecules through single-step, two-reactant synthesis pathways.

SAVI compounds are particularly valuable for virtual screening, drug discovery, and computational chemistry applications because each molecule comes with:

• A proposed single-step synthetic route
• Predicted synthetic accessibility scores
• Comprehensive molecular property annotations
• Quality assessments based on drug-likeness criteria

Supported Tasks

• Virtual Screening: High-throughput computational screening of large chemical libraries for drug discovery
• Chemical Space Exploration: Systematic exploration of synthetically accessible chemical space
• Lead Optimization: Structure-activity relationship (SAR) studies using built-in compound series
• Synthetic Route Planning: Each compound includes reactants and transformation information for synthesis planning
• Property Prediction: Training machine learning models for molecular property prediction using the extensive annotations

Getting Started

Quick Start

from datasets import load_dataset

# Load the dataset
dataset = load_dataset("Metanova/SAVI-2020")

# View basic information
print(f"Dataset size: {len(dataset['train'])}")
print(f"Features: {dataset['train'].features}")

# Examine first few examples
print(dataset["train"][:5])

Dataset Structure

Data Instances

Each instance represents a computationally generated compound with comprehensive annotations:

{
  "product_name": "61A77BD3826E0AD0_554BFAE008D564C8_6031_UN",
  "product_hashisy": "61A77BD3826E0AD0",
  "product_smiles": "O(C(=O)C1=C(OC=N1)C2=CC=CS2)CCC3OCCC3C",
  "formula": "C15H17NO4S",
  "weight": 307.3636,
  "rotbonds": 6,
  "h_donors": 0,
  "h_acceptors": 6,
  "e_tpsa": 89.8,
  "fsp3": 0.4667,
  "complexity": 367.5828,
  "product_qed_score": 0.7358,
  "pains_filter_match_count": 0,
  "bw_demerit_score": 35,
  "transform_name": "Mitsunobu Reaction",
  "transform_id": 6031,
  "r1_smiles": "OC(=O)C1=C(OC=N1)C2=CC=CS2",
  "r1_ident": "EN300-72222",
  "r2_smiles": "CC1CCOC1CCO",
  "r2_ident": "EN300-6499501",
  "score_class": "plus",
  "stdinchi": "InChI=1S/C15H17NO4S/c1-10-4-6-18-11(10)5-7-19-15(17)13-14(20-9-16-13)12-3-2-8-21-12/h2-3,8-11H,4-7H2,1H3",
  "stdinchikey": "InChIKey=FPQADJBFVMQUTQ-UHFFFAOYSA-N"
}

Data Fields

Field Name	Type	Description
Product Information
product_name	string	Unique SAVI identifier with hash and transform information
product_hashisy	string	Hash identifier for the product structure
product_smiles	string	SMILES representation of the generated product
formula	string	Molecular formula (e.g., "C15H17NO4S")
weight	float64	Molecular weight in Daltons
stdinchi	string	Standard International Chemical Identifier
stdinchikey	string	Standard InChI Key for database searching
product_stereo_tauto_hash	string	Hash considering stereochemistry and tautomers
Molecular Properties
rotbonds	int64	Number of rotatable bonds (flexibility measure)
h_donors	int64	Number of hydrogen bond donors
h_acceptors	int64	Number of hydrogen bond acceptors
e_tpsa	float64	Topological Polar Surface Area in Ų
fsp3	float64	Fraction of sp³ hybridized carbons (3D character)
complexity	float64	Bertz/Hendrickson molecular complexity score
heavy_atoms	int64	Total number of non-hydrogen atoms
hydrogen_bond_center_count	int64	Count of hydrogen bonding centers
Drug-likeness & Quality
product_qed_score	float64	Quantitative Estimate of Drug-likeness (0-1, higher better)
ro5_violations	float64	Number of Lipinski's Rule of Five violations
ro3_violations	float64	Number of Rule of Three violations
pains_filter_match_count	int64	Number of PAINS (Pan Assay Interference) alerts
pains_filter_match_name	string	Names of matching PAINS patterns
bw_demerit_score	int64	Bruns & Watson demerits (lower better)
bw_demerit_components	string	Components contributing to B&W demerits
genotoxic_alerts	string	Genotoxic structural alerts
Lipophilicity
xlogp2	float64	Predicted partition coefficient (lipophilicity)
xlogp	float64	Alternative XlogP calculation
Ring Analysis
ring_system_count	int64	Number of ring systems
product_aroring_count	int64	Number of aromatic rings
product_aliring_count	int64	Number of aliphatic rings
e_nesssr	int64	Number of rings in smallest set of smallest rings
e_nsssr	int64	Number of smallest set of smallest rings
benzenoid_index	float64	Measure of benzenoid character
Stereochemistry
stereo_count	string	Stereochemistry counts (format: "chiral_centers undefined_stereo defined_stereo etc.")
charged_group_counts	string	Count of charged groups
Reaction Information
reaction_hashisy	string	Hash identifier for the reaction
reaction_smiles	string	Reaction SMILES showing transformation
transform_name	string	Name of the chemical transformation (e.g., "Mitsunobu Reaction")
transform_id	int64	Numerical identifier for the transformation
lhasa_score	int64	LHASA expert system score for reaction feasibility
lhasa_scoring_history	string	Detailed scoring breakdown
lhasa_reaction_conditions	string	Recommended reaction conditions
lhasa_reaction_actual_conditions	string	Actual conditions used in computation
lhasa_reaction_warnings	string	Any warnings about the reaction
min_delta	int64	Minimum score delta
score_class	string	Reaction success class ("plus", "neg0", "neg10", "neg20", "neg30")
Ring Formation Analysis
reaction_delta_ring_count	int64	Change in total ring count
reaction_delta_aroring_count	int64	Change in aromatic ring count
reaction_delta_aliring_count	int64	Change in aliphatic ring count
Reaction Identifiers
rinchi	string	Reaction InChI identifier
rinchikey_long	string	Long format Reaction InChI Key
rinchikey_short	string	Short format Reaction InChI Key
rinchikey_web	string	Web format Reaction InChI Key
Reactant 1 (R1) Information
r1_hashisy	string	Hash identifier for first reactant
r1_smiles	string	SMILES of first building block
r1_bbsource	string	Source of building block (e.g., "ENAMINE2019Q4")
r1_ident	string	Building block catalog identifier
r1_inchi	string	InChI of first reactant
r1_inchikey	string	InChI Key of first reactant
r1_protection_required	int64	Whether protecting groups were required (0/1)
r1_protected	int64	Whether reactant was protected (0/1)
r1_url	string	URL to purchase building block
Reactant 2 (R2) Information
r2_hashisy	string	Hash identifier for second reactant
r2_smiles	string	SMILES of second building block
r2_bbsource	string	Source of building block
r2_ident	string	Building block catalog identifier
r2_inchi	string	InChI of second reactant
r2_inchikey	string	InChI Key of second reactant
r2_protection_required	int64	Whether protecting groups were required (0/1)
r2_protected	int64	Whether reactant was protected (0/1)
r2_url	string	URL to purchase building block
Technical Fields
fmt_chunkid	int64	Data chunk identifier for processing
record_number	int64	Record number within dataset

Data Splits

The dataset contains approximately 1.53 billion unique compounds in a single training split.

Applications and Use Cases

Drug Discovery

• Virtual Screening: Large-scale computational screening for hit identification
• Lead Optimization: SAR exploration using built-in compound series
• Chemical Space Navigation: Systematic exploration of accessible chemical space

Computational Chemistry

• Property Prediction: Training datasets for QSAR and machine learning models
• Synthetic Route Planning: Integration with retrosynthetic analysis tools
• Chemical Space Analysis: Diversity and novelty assessment studies

Academic Research

• Method Development: Benchmarking virtual screening and ML approaches
• Chemical Education: Examples of structure-property relationships
• Open Science: Publicly available alternative to proprietary chemical libraries

Original Paper:

Patel, H., Ihlenfeldt, W.D., Judson, P.N. et al. 
SAVI, in silico generation of billions of easily synthesizable compounds through expert-system type rules. 
Sci Data 7, 384 (2020). 
https://doi.org/10.1038/s41597-020-00727-4