Python library and command-line tool for extracting compounds from scientific literature. Written in Python.
Switch branches/tags
Nothing to show
Clone or download
Latest commit 86192a7 May 10, 2017
Failed to load latest commit information.
docs updated apidocs Mar 22, 2017
molminer Fixed finding path to libmagic file magic.mgc Apr 19, 2017
recipe added libmagic to deps Apr 23, 2017
scripts initial commit Mar 17, 2017
.gitignore updated Mar 18, 2017
LICENSE Initial commit Mar 16, 2017 Update May 10, 2017 modified Mar 20, 2017


MolMiner is a library and command-line interface for extracting compounds (called "chemical entities") from scientific literature. It extracts chemical entities both from text (Chemical Named Entity Recognition) and 2D structures (Optical Chemical Structure Recognition). It's written in Python (currently supporting only Python 3). It should work on all platforms, but problem is that some dependencies are very hard to compile on Windows. Actually it's a wrapper around several open-source tools for chemical information retrieval, namely ChemSpot, OSRA and OPSIN, using their command-line interface and adding some extended functionality.


MolMiner is able to extract chemical entities from scientific literature in various formats including PDF and scanned images. It extracts entities both from text and 2D structures. Text is normalized using part of code from ChemDataExtractor. Text entities are assigned by ChemSpot to one of classes: "SYSTEMATIC", "IDENTIFIER", "FORMULA", "TRIVIAL", "ABBREVIATION", "FAMILY", "MULTIPLE". IUPAC names are converted to computer-readable format like SMILES or InChI with OPSIN. 2D stuctures are recognised in document and converted to computer-readable format with OSRA. Entities successfully converted to computer-readable format are standardized using MolVS library. Entities are also annotated in PubChem and ChemSpider databases using PubChemPy and ChemSpiPy. For processing of PDF files is used GraphicsMagick and for OCR Tesseract.


MolMiner self is written in Python, but it uses several binaries and some of them have complicated compilation dependencies. So the easiest way is to install MolMiner including dependencies as a conda package hosted on Anaconda Cloud.

To install MolMiner without dependencies just download this repository and run $ python install. MolMiner will be then available from shell as molminer and also as a Python library.

Conda package (currently only for linux64)

Conda is a package, dependency and environment management for any language including Python. MolMiner package includes precompiled dependencies and data files. It also manages all the needed envinronment variables and enables bash auto-completion.

  1. Download and install conda.

  2. Add channels:

    $ conda config --add channels rdkit; conda config --add channels bioconda; conda config --add channels jirinovo; conda config --add channels conda-forge

  3. Create new virtual environment and install MolMiner:

    $ conda create -n my_new_env molminer

  4. Activate environment:

    $ source activate my_new_env

  5. Use MolMiner:

    $ molminer --help

Note that you must always activate virtual environment before using MolMiner. That's because the activation script is also modifying the environmental variables storing the paths to MolMiner data files.

From source (linux)


You need all these binaries for MolMiner. They should be installed so path to them is in PATH environmental variable (like /usr/local/bin). I haven't tried to compile these dependencies on Windows, but that doesn't mean it's impossible.

  • OSRA. This is probably the most complicated binary. Official information is here and here. My installation notes are here.
    • Compile GraphicsMagick with as many supported image formats as possible (dependencies). It's also used for converting PDF to images and for image editing/transformation.
    • Use Tesseract version 4 and up.
    • Patched version of OpenBabel is needed.
    • Put OSRA data files (spelling.txt, superatom.txt) to some directory and add this directory to OSRA_DATA_PATH environmental variable.
  • ChemSpot. Just download it and:
    • Put ChemSpot JAR file to directory accesible from PATH and rename it to chemspot.jar.
    • Also put there this bash script. It's used for running ChemSpot. Its first argument is maximum amount of memory for ChemSpot process. Subsequent arguments are forwarded to ChemSpot CLI.
    • Put ChemSpot data files (,, multiclass.bin) to some directory and add this directory to CHEMSPOT_DATA_PATH environmental variable.
  • OPSIN. Just download it and:
    • Put OPSIN JAR file to directory accesible from PATH and rename it to opsin.jar.
    • Also put there this bash script. It's used for running OPSIN. All arguments are forwarded to OPSIN CLI.
  • GraphicsMagick. OSRA needs it for compilation, but its binary is also directly used by MolMiner. Compile it with as many supported image formats as possible (dependencies).
  • Tesseract. OSRA needs it for compilation, but its binary is also directly used by MolMiner. Use version 4 and up.
    • Tesseract needs language data files. Download them here, put them to some directory and add this directory to TESSDATA_PREFIX environmental variable.
  • poppler-utils. Utils for PDF files built on top of Poppler library.
    • Ubuntu (or any OS with apt packaging): $ sudo apt-get install poppler-utils
  • libmagic. Reads the magic bytes of file and determine its MIME type.
    • Ubuntu (or any OS with apt packaging): $ sudo apt-get install libmagic1 libmagic-dev
  • OpenJDK. Java runtime environment. Installation.

Paths to data files can be also specified in both MolMiner CLI and library, but defining them in the environmental variables is the easiest way.

Python dependencies

Dependencies listed in will be installed automatically when you run $ python install. Unfortunately, there is a complicated dependency RDKit. It's best to install it as a conda package.


  • Basic syntax is: $ molminer COMMAND [OPTIONS] [ARGS]

  • MolMiner has four commands (you can view them with $ molminer --help):

    • ocsr: Extract 2D structures with OSRA. OCSR stands for Optical Chemical Structure Recognition.
    • ner: Extract textual chemical entities with ChemSpot. NER stands for Named Entity Recognition.
    • convert: Convert IUPAC names to computer-readable format with OPSIN.
    • extract: Combine all the previous commands.
  • To each command you can view its options with $ molminer COMMAND --help

  • Bash auto-completion is automatically available when MolMiner is installed through conda and virtual environment is activated. Then you can double-press TAB key to show MolMiner commands and options: $ molminer <TAB><TAB> to see commands and $ molminer ocsr --<TAB><TAB> to see options.

    • To manually activate bash auto-completion: $ eval "$(_MOLMINER_COMPLETE=source molminer)" You can put this to your .bashrc file at your home directory. Internally, it's a feature of click library documented here.


  • Input can be single PDF, image or text file. Type of input file will be automatically determined, but you can specify it with -i [pdf|pdf_scan|image|text] option (text value is of course not supported by OSRA, resp. ocsr command). Only PDF containing scanned papers cannot be identified so you must pass -i pdf_scan option.
  • Input from stdin is also supported. You can use it together with ner and convert command. For convert a list of IUPAC names is expected, each name on single line.
  • If you know that your text is paged, i.e. contains page separators -- ASCII control character 12 (Form Feed, '\f'), you can pass --paged-text flag and to each entity will be assigned page. This is automatically done when input is PDF file.


  • Result is a CSV file. Defaultly, MolMiner will write result to stdout. If you want to write result directly to file, use -o <file> option. To change CSV file delimiter use -d <delimiter> option.
  • Chemical entities, which were successfully converted to computer-readable format, can be also written to SDF file by specifying --sdf-output <file> option. If you don't want to create new SDF file and just append to it, use --sdf-append flag.
  • When using extract command, you can also output CSV files separately from OSRA, ChemSpot and OPSIN by using the --separated-output flag.

Defaultly enabled features

By default, these features are enabled:

  • Conversion of PDF files to temporary PNG images using GraphicsMagick (GM). OSRA itself can handle PDF files, but using this is more reliable, because OSRA (v2.1.0) is showing wrong information when converting directly from PDF (namely: coordinates, bond length and possibly more ones) and also there are sometimes incorrectly recognised structures. Also it seems that this is sometimes a little bit faster (internally each temporary image is processed in parallel and results are then joined). Use --no-use-gm flag to disable it.
  • Standardization of chemical entities converted to computer-readable format. See MolVS documentation for explanation. Use --no-standardization flag to disable it.
  • Annotation of chemical entities in PubChem and ChemSpider. This will try to assign compound IDs by searching separately with different identifiers (entity name, SMILES etc.). If single result is found by searching with entity name, missing indentifiers are added. InChI-key is preffered in searching. To annotate using ChemSpider you need ChemSpider API token. You can get it by signing up on their website. Then provide this token with --chemspider-token <token> option.
  • Normalization of text. This is strongly recommended to keep as is, because sometimes is ChemSpot producing weird and unparsable results. Use --no-normalize-text flag to disable it.
  • Parallel processing will use all available cores. Use -j <#cores> option to change it. '-1' to use all CPU cores. '-2' to use all CPU cores minus one.

MolMiner library

Autogenerated API documentation

Wrapper classes

For each of OSRA, ChemSpot and OPSIN there is a wrapper class. Some general options can be set in constructor. Each class has a process() method which takes path to input file or string and do the desired job with it. It returns an OrderedDict instance with results and can also write a CSV file. Classes, methods and their parameters are well documented in autogenerated API documentation. Unfortunately, the documentation builder (Sphinxdoc) is skipping the Python magic methods, which is besides the class constructor (__init__). For now, please refer directly to __init__ method docstrings in source code. They should be well-readable, because they are using numpydoc style.


from molminer import OSRA
from pprint import pprint

osra = OSRA(unpaper=2)
extracted = osra.process("path/to/document.pdf", output_file="path/to/output.csv", output_formats=["smiles", "inchi"],

Extractor class

This class combines OSRA, ChemSpot and OPSIN to extract chemical entities both from text and 2D structures. It has the same interface as wrapper classes. To constructor you can pass dicts with key-values mapping to wrapper classes constructor's named arguments.


from molminer import Extractor
from pprint import pprint

extractor = Extractor(chemspot_options={"max_memory": 16})
extracted = extractor.process("path/to/document.pdf")


  • ChemSpot itself is very memory-consuming so dictionary and ID lookup is disabled by default. Only CRF, OpenNLP sentence and multiclass models will be used by default. Maximum memory used by Java process is set to 8 GB by default. It is strongly recommended to use swap file on SSD disk when available memory is under 8 GB (see for more details). If you want to use dictionary and ID lookup in ChemSpot, pass --chs-dict and --chs-ids options. If you are using MolMiner library, pass path_to_dict="" and path_to_ids="" to ChemSpot class constructor.
  • If you are using conda package and want to add more Tesseract languages, download them and put them to <path_to_your_conda_env>/share/molminer/tesseract. <path_to_your_conda_env> is usually /home/<username>/miniconda3/envs/<your_env>. If you aren't using conda package, follow the instructions here for Tesseract.
  • Unfortunately, there wasn't enough time to write unit tests. I hope I will find time in future to do it.
  • We also wanted to test MolMiner's quality. That means mainly the completeness of extraction and ratio of false positives. Unfortunately, there aren't complex test data which will cover both textual and 2D structure chemical entities. We don't have enough time to prepare such a complex dataset manually, so for now you can separately look at ChemSpot and OSRA test results.
  • If you successfully compile all the dependencies for Windows, let me kindly know and I will add MolMiner package for Windows to Anaconda Cloud. Thank you!


Feel free to open issue ticket. This is the prefered way since we don't have mailing list.


MolMiner was the job description of my diploma thesis at University of Chemistry and Technology in Prague, Laboratory of chemistry and informatics. I would like to thank my supervisor Daniel Svozil for leading the work and Martin Sicho for helping me with conda distribution of MolMiner.

Citations of used software:

  • Rocktäschel, T., Weidlich, M., and Leser, U. (2012). ChemSpot: A Hybrid System for Chemical Named Entity Recognition. Bioinformatics 28 (12): 1633-1640.
  • "Optical Structure Recognition Software To Recover Chemical Information: OSRA, An Open Source Solution" J. Chem. Inf. Model., 2009, 49 (3), pp 740–743.
  • Chemical Name to Structure: OPSIN, an Open Source Solution Daniel. M. Lowe, Peter T. Corbett, Peter Murray-Rust, Robert C. Glen, Journal of Chemical Information and Modeling 2011 51 (3), 739-753