The BOS-Lig Dataset: Accurate Ligand Charges from a Consensus Approach for 66,810 Experimentally Synthesized Ligands

The paper introduces the BOS-Lig dataset, which assigns accurate net charges and functional application areas to over 66,000 experimentally synthesized ligands derived from 126,985 transition metal complexes using a novel consensus-based charge-balancing workflow and topic modeling.

The Gist

Imagine you are trying to build a massive, complex Lego castle. You have a giant box of instructions and millions of individual Lego bricks, but there's a problem: the instructions are missing the most important part. They tell you what the bricks look like, but they don't tell you how much "weight" or "energy" each brick carries.

In the world of chemistry, these bricks are called ligands (the molecules that stick to a metal center), and the "weight" is their electric charge. Without knowing the charge, scientists can't accurately predict how the castle (a chemical complex) will behave, react, or function.

This paper is about a team of researchers at MIT who decided to fix this missing information for a huge collection of chemical structures. Here is the story of how they did it, explained simply:

1. The Problem: A Library with Missing Labels

The researchers started with the Cambridge Structural Database (CSD), which is like the world's biggest library of crystal structures. It contains over 126,000 different metal complexes (the "castles").

• The Issue: While the library has pictures of the structures, it often lacks clear labels for the charge of the ligands.
• The Consequence: If a computer tries to simulate these chemicals without knowing the charge, it's like trying to bake a cake without knowing if you have salt or sugar. The result is a mess. Previous methods to guess the charges were like using a "rule of thumb" that worked for simple cakes but failed for complex ones.

2. The Solution: The "Detective" Workflow

The team created a new dataset called BOS-Lig (Boston Open-Shell Ligand). To figure out the missing charges, they didn't just guess; they acted like detectives using a consensus approach.

Think of it like solving a mystery where you have 100 different witnesses (different scientific papers) describing the same suspect (a ligand).

• Step 1: The Easy Cases (Homoleptic Complexes). First, they looked at complexes where the metal is surrounded by only one type of ligand (like a castle made of only red bricks). If the whole castle has a known total charge, and there are 4 identical red bricks, they can easily divide the total charge by 4 to find the charge of one brick.
• Step 2: The Detective Work (Iterative Propagation). Once they knew the charge of the "red bricks," they moved to more complex castles made of mixed bricks (red, blue, and green). Since they now knew the charge of the red bricks, they could use math to figure out the charge of the blue and green ones.
• Step 3: The Voting System. Sometimes, different papers gave different answers for the same brick. The team didn't just pick one; they used a weighted voting system.
  • If a paper had a very clear, high-quality photo (low "noise"), their vote counted more.
  • If a paper had a blurry photo, their vote counted less.
  • They kept doing this until the votes settled on a single, most likely charge for each ligand.

3. The Results: A Massive New Map

By the end of their detective work, they successfully assigned charges to 66,810 unique ligands.

• The Scale: This is nearly 10 times larger than previous attempts. They covered a huge chunk of the chemical world that was previously a "dark forest."
• The Quality Check: They even created a "Purity Metric." Imagine a ligand that appears in 100 papers. If 99 papers say it's "negative" and 1 says it's "positive," the purity is high. If the papers are split 50/50, the purity is low, and the team flagged it as "uncertain" so scientists know to be careful.

4. Beyond Charges: The "Job Description" of a Ligand

The team didn't just stop at charges. They also wanted to know: What do these ligands actually do?

• They used Artificial Intelligence (AI) to read the abstracts of the scientific papers associated with each ligand.
• They sorted the ligands into "job categories":
  • The Reactors: Used for making new chemicals (Catalysis).
  • The Healers: Used in biology and medicine.
  • The Lighters: Used in glowing lights and screens (Photophysics).
  • The Spinners: Used in magnets and data storage.
• The Insight: They found that some ligands are "specialists" (only good for one job, like a surgeon), while others are "generalists" (good at many jobs, like a handyman). They even created a "Purity Score" to tell you how specialized a ligand is.

5. The Gift to the World: The BOS-Lig Browser

Finally, they didn't keep this data in a locked drawer. They built a free, interactive website (the BOS-Lig Browser).

• How it works: Imagine a search engine for chemistry. You can type in a chemical name or a code, and the website instantly tells you:
  • "This ligand has a charge of -1."
  • "It usually connects to metals with 2 or 3 arms."
  • "It is mostly used in making medicines."
• Why it matters: This allows scientists to design new drugs, better batteries, or more efficient catalysts much faster because they have a reliable map of the chemical landscape.

Summary

In short, this paper is about cleaning up a messy library. The researchers took thousands of confusing chemical structures, used a smart, step-by-step detective method to figure out the missing "electric weights" of the pieces, and organized them by what they are used for. They then gave this organized map to the entire scientific community for free, making it much easier to invent the next generation of medicines, materials, and technologies.

Technical Summary

1. Problem Statement

Transition metal complexes (TMCs) are vital for catalysis, photochemistry, magnetism, and bioinorganic chemistry. However, computational high-throughput screening (HTS) and machine learning (ML) efforts are hindered by the lack of consistent, experimentally grounded data regarding ligand properties.

• Data Gaps: Fundamental properties like net ligand charge, coordination modes, and functional application areas are often missing or inconsistently recorded in crystallographic databases like the Cambridge Structural Database (CSD).
• Limitations of Current Methods: Existing methods for inferring ligand charges rely on simplified heuristics (e.g., the octet rule) or specific tools (e.g., cell2mol) that fail on complex, disordered, or large unit cells. These approaches often generalize poorly across the diverse chemical space of TMCs, leading to systematic errors in Density Functional Theory (DFT) calculations and automated design workflows.
• Contextual Blindness: There is a lack of systematic linkage between specific ligand structures and their functional application domains (e.g., biological vs. photophysical), making it difficult to distinguish specialized ligands from general-purpose ones.

2. Methodology

The authors developed a comprehensive workflow to curate the Boston Open-Shell Ligand (BOS-Lig) dataset from the CSD (March 2024 update). The methodology consists of four main stages:

A. Dataset Curation and Pre-processing

• Source: Extracted 599,180 mononuclear TMC entries from the CSD.
• Filtering: Excluded polymeric networks, disordered fragments, and non-molecular species.
• Hydrogen Addition: Used the CCDC Python API to infer missing hydrogens. Structures with excessive or unlikely hydrogen additions were flagged as "non-trustworthy" and removed.
• Final Set: Resulted in 126,985 high-confidence mononuclear TMCs containing 29 transition metals.

B. Complex Charge Inference (Iterative Consensus)

• Decomposition: Unit cells were decomposed into molecular components (TMCs, solvents, counterions).
• Hashing: A Weisfeiler–Lehman (WL) graph hash was used to identify unique molecular species based on topology and atom identity.
• Iterative Solving:
  1. Seeding: Manually defined charges for 267 common non-metal species (e.g., counterions).
  2. Propagation: Enforced unit cell neutrality to infer charges of unknown components by subtracting known charges.
  3. Consensus: Applied a majority-vote consensus across multiple crystallographic observations.
  4. Purity Metric: A charge was only assigned if it had $\ge$3 independent observations and a "charge purity" (percentage of agreement) of $\ge$67%.
  5. Convergence: The process ran for 9 iterations until no new charges were assigned.

C. Ligand Charge Assignment

• Oxidation States: Metal oxidation states were obtained via text parsing of CSD metadata and the cell2mol tool.
• Iterative Ligand Solving:
  1. Homoleptic First: Charges were assigned to ligands in homoleptic complexes (all ligands identical) by dividing the complex charge by the ligand count.
  2. Heteroleptic Propagation: Charges were iteratively propagated to heteroleptic complexes where all but one ligand type had known charges.
• Weighted Voting: To resolve conflicts where the same ligand hash yielded different charges, a weighted scheme was applied based on:
  • Crystallographic resolution (lower R-factor = higher weight).
  • Simplicity of the complex (fewer unique ligand types = higher weight).
  • Integer nature of the derived charge.
• Outcome: Successfully assigned charges to 66,810 unique ligands (71% of the 94,581 unique ligands identified).

D. Coordination and Application Analysis

• Coordination: Extracted donor atom identities and counts to identify binding modes and hemilability (dynamic coordination changes).
• Topic Modeling: Linked CSD entries to journal abstracts. Used a BERTopic-based workflow to classify ligands into five application areas: Reactivity, Biological, Magnetism, Redox, and Photophysical.
• Purity Metric: Defined an "application purity" score ($p$) to quantify whether a ligand is specialized (high $p$) or broadly reused (low $p$) across different fields.

3. Key Contributions

1. BOS-Lig Dataset: The largest experimentally grounded dataset of ligand charges to date, covering 66,810 unique ligands from 126,985 TMCs.
2. Robust Charge Assignment Workflow: A novel iterative, consensus-based approach that outperforms simple heuristics (octet rule) and automated tools (cell2mol) in coverage and accuracy, particularly for complex or hypervalent systems.
3. Application Linkage: The first large-scale effort to map ligand structures to functional application domains using NLP, providing a "purity" metric to guide ligand selection.
4. Hemilability Classification: Systematic identification of hemilabile ligands (approx. 8% of the set) directly from crystallographic evidence.
5. Open Access Tools: Development of the BOS-Lig Browser, a web interface allowing users to query ligands by SMILES or graph hash to view charges, coordination modes, and application profiles.

4. Key Results

• Coverage: The dataset covers 94,581 unique ligand structures, with charges assigned to 66,810 (71%).
• Charge Distribution: Assigned charges range primarily from -4 to +2. Neutral ligands (39.7%) and -1 ligands (38.5%) are most common.
• Accuracy Validation:
  • vs. CSD Metadata: 91% agreement with CSD-reported charges where available.
  • vs. Octet Rule: 87% agreement, but the iterative workflow corrected errors in 28 out of 40 manually inspected disagreements (mostly involving $\pi$-acceptors and delocalized systems).
  • vs. cell2mol: The workflow successfully assigned charges to diverse structures where cell2mol failed to parse the unit cell.
• Consistency: Only ~4% of ligands showed significant charge inconsistency (inconsistency score > 0.25), indicating high reliability in the consensus approach.
• Application Insights:
  • Reactivity is the most populated domain, dominated by neutral, monodentate ligands.
  • Biological and Photophysical domains are more specialized (higher purity).
  • Redox and Magnetism show high overlap with other domains.
  • Specific high-purity examples were identified (e.g., phenylalaninato for bio, specific phosphines for photo).

5. Significance

The BOS-Lig dataset addresses a critical bottleneck in computational chemistry: the lack of reliable, large-scale ligand property data.

• For Computational Screening: Provides accurate initial charges for DFT calculations, reducing systematic errors in electronic structure predictions for transition metal complexes.
• For Machine Learning: Offers a high-quality training set for models predicting ligand properties, reactivity, and complex stability.
• For Ligand Design: The application purity metrics allow researchers to rationally select ligands based on their historical success in specific domains (e.g., choosing high-purity "bio" ligands for drug discovery).
• Scalability: The consensus-based workflow is robust enough to handle the noise and variability inherent in large crystallographic databases, setting a new standard for data curation in inorganic chemistry.

The dataset and tools are publicly available via the BOS-Lig Browser and Zenodo, facilitating data-driven discovery in transition metal chemistry.