THIRD_PARTY_NOTICES.md

Third-party notices

The study weights, PCA components and loading adapters use the MIT license in LICENSE, with the Microsoft copyright notice. Third-party files retain the licenses and copyright notices listed below.

Paper code and gene order

• Project: https://github.com/microsoft/scFM-dataselection
• Included gene-order file: public-resources/sctab/adata_var.csv.
• License: MIT, Copyright (c) Microsoft Corporation.
• Full notice: licenses/scFM-dataselection-MIT.txt.

SSL implementation

• Project: https://github.com/v-mahughes/ssl_in_scg
• Included source files: public-resources/ssl-source/base.py and public-resources/ssl-source/cellnet_autoencoder.py.
• Authors credited by the project: Till Richter, Mojtaba Bahrami, Yufan Xia and Felix Fischer.
• License: MIT, Copyright (c) 2024 Till Richter.
• Full notice: licenses/ssl_in_scg-MIT.txt.

The loading adapters in tools/ use these source files to construct the models.

Geneformer tokenizer files

• Source: https://github.com/lcrawlab/Geneformer
• Original project: https://huggingface.co/ctheodoris/Geneformer
• Original work: Christine V. Theodoris and collaborators.
• Included dictionaries: token_dictionary.pkl, gene_median_dictionary.pkl and gene_name_id_dict.pkl in public-resources/geneformer/.
• License: Apache-2.0.
• Full terms: licenses/Geneformer-Apache-2.0.txt.

The Geneformer weights are models trained for this study, not upstream Geneformer pretrained weights.

SCimilarity network implementation

• Project: https://github.com/Genentech/scimilarity
• Included source: tools/_scimilarity_networks.py, the unmodified scimilarity/nn_models.py from version 0.4.1.
• License: Apache-2.0, Copyright 2023 Genentech, Inc.
• Full notice and terms: licenses/SCimilarity-Apache-2.0.txt.

The SCimilarity encoder and decoder weights were trained for this study; they are distributed under MIT, separately from the upstream implementation.

Dependencies and data

Runtime packages are installed separately and retain their own licenses. The scTab corpus is not included. See https://github.com/theislab/scTab for dataset access and terms, and the paper code for preprocessing.