SOURCE METADATA
Project: CM4AI
Source ID: biorxiv_preprint
Source type: preprint
Source URL: https://www.biorxiv.org/content/10.1101/2024.05.21.589311v1
Raw file: data/raw/CM4AI/biorxiv_2024.05.21.589311v1_row4.pdf
--------------------------------------------------------------------------------
bioRxiv preprint

doi:

https://doi.org/10.1101/2024.05.21.589311

;

this version posted May 24, 2024.

The copyright holder for this preprint (which

was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made

available under a

CC-BY-NC-ND 4.0 International license
.

Cell Maps for Artificial Intelligence: AI-Ready Maps of Human Cell Architecture
from Disease-Relevant Cell Lines
Timothy Clark1†, Jillian Mohan2, Leah Schaffer2, Kirsten Obernier3, Sadnan Al Manir1,
Christopher P Churas2, Amir Dailamy2, Yesh Doctor2, Antoine Forget3, Jan Niklas
Hansen4, Mengzhou Hu2, Joanna Lenkiewicz2, Maxwell Adam Levinson1, Charlotte
Marquez2, Sami Nourreddine2, Justin Niestroy1, Dexter Pratt2, Gege Qian2, Swathi
Thaker5, Jean-Christophe Bélisle-Pipon6†, Cynthia Brandt7†, Jake Chen5†, Ying Ding8†,
Samah Fodeh7†, Nevan Krogan3†, Emma Lundberg4†, Prashant Mali2†, Pamela Payne-
Foster9†, Sarah Ratcliffe1†, Vardit Ravitsky10†, Andrej Sali3†, Wade Schulz7†, Trey
Ideker2*

1 - University of Virginia; 2 - University of California San Diego; 3 - University of
California San Francisco; 4 - Stanford University; 5 - University of Alabama at
Birmingham; 6 - Simon Fraser University; 7 - Yale University; 8 - University of Texas at
Austin; 9 - University of Alabama; 10 - University of Montreal
* corresponding author – tideker@health.ucsd.edu; † co-corresponding authors

Abstract

This article describes the Cell Maps for Artificial Intelligence (CM4AI) project and its
goals, methods, standards, current datasets, software tools , status, and future
directions. CM4AI is the Functional Genomics Data Generation Project in the U.S.
National Institute of Health’s (NIH) Bridge2AI program. Its overarching mission is to
produce ethical, AI-ready datasets of cell architecture, inferred from multimodal data
collected for human cell lines, to enable transformative biomedical AI research.

Introduction

1. Goals of CM4AI

Cell Maps for Artificial Intelligence (CM4AI) is an NIH-funded Bridge to Artificial
Intelligence (Bridge2AI)1 Data Generation Project (DGP) in the area of Functional
Genomics. It consists of three pillars – Data, People, and Ethics – organized into six
modules (Data: Data Acquisition, Tools, and Standards; People: Skills and Workforce
Development, and Teaming; and Ethics).

CM4AI’s objective is to deliver machine-readable hierarchical maps of cell architecture
as AI-Ready data produced from multimodal interrogation of 100 chromatin modifiers
and 100 metabolic enzymes involved in cancer, neuropsychiatric, and cardiac disorders
in disease-relevant cell lines under perturbed and unperturbed conditions, utilizing state-
of-the-art mass spectrometry based proteomics, spatial proteomics / cell imaging, and
genetic perturbations using CRISPR. The cell lines currently under investigation in
CM4AI are treated (with paclitaxel and vorinostat) versus untreated MDA-MB-468
breast cancer cell lines; and undifferentiated vs.  neurons and cardiomyocytes
generated from NIH-iPCS-1 KOLF2.1J induced pluripotent stem cells (IPSCs). CM4AI’s
software pipeline and AI-readiness framework provide important and reusable enabling
capabilities for this work.

CM4AI input data streams are generated using immunofluorescence (IF) subcellular
microscopy for spatial proteomics data; affinity purification mass spectroscopy (AP-MS)

1



bioRxiv preprint

doi:

https://doi.org/10.1101/2024.05.21.589311

;

this version posted May 24, 2024.

The copyright holder for this preprint (which

was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made

available under a

CC-BY-NC-ND 4.0 International license
.

and size exclusion mass spectroscopy (SEC-MS) methods for protein-protein
interaction (PPI) data; and single-cell CRISPR-Cas perturbation screens by cell type.
Input data streams are integrated via the multi-scale integrated cell (MuSIC) software
pipeline employing deep learning models and community detection algorithms2, and
output cell maps are packaged with provenance graphs and rich metadata as AI-Ready
datasets in RO-Crate format3,4 using an extended, client-server version of the
FAIRSCAPE framework5.

At a strategic level, CM4AI – as do other Bridge2AI DGPs – aims to ethically leverage
the most state-of-the art computational and experimental tools in biomedical science to
enable a new generation of  transformative artificial intelligence research to benefit
humanity. Its datasets and software are available to the research community under
minimally restrictive terms consistent with ethical guidelines, researcher attribution, and
research integrity.

2. What is a Cell Map?

Cell maps are hierarchical directed acyclic graphs (DAG), where each node represents
an assembly of proteins in proximity at a given scale, spanning from large assemblies
representing cell compartments (e.g. nucleus, mitochondria) to small assemblies of
proteins in close proximity (e.g. protein complexes or subunits). The data streams for
constructing these maps are generated from a panel of perturbed and unperturbed cell
lines, including treated and untreated cancer cell lines as well as differentiated and
naive induced pluripotent stem cells (iPSCs). Cell maps provide a foundation for
downstream applications in human genomics, including interpretation of genetic variants
and mutations. As an example, they have been used in AI tools for “visible machine
learning” or “visible neural networks” (VNNs) to interrogate how protein assemblies in
the cell affect cell-level phenotype predictions6–10.

In the CM4AI cell mapping process, affinity purification-mass spectrometry (AP-MS) and
size-exclusion mass spectroscopy (SEC-MS) techniques generate protein interaction
networks, while immunofluorescence (IF) staining and high-resolution microscopy
reveal protein localization and distribution within human cells11.

Cell maps are produced by integrating these localization and interaction data using self-
supervised embedding approaches from deep learning, followed by algorithmic
community detection at multiple resolutions to generate a hierarchy of protein
assemblies12. Cell maps are shared via the Network Data Exchange (NDEx)13  and can
be visualized in a web browser or accessed via tools such as Cytoscape14, HiView15,
and the Python ndex2 library16.
An example cell map from the Multi-Scale Integrated Cell (MuSIC)2 data set visualized
in HiView is shown in Figure 1.

3. What are Ethical, FAIR, AI-Ready Biomedical Data?

As defined by CM4AI, AI-Ready biomedical data are fully characterized FAIR data of
known provenance, which can be ethically and reliably processed by AI applications;
whose models and software are available and well-described for validation and re-use;
and whose predictions may be fully explained and interpreted to the user as needed.
CM4AI data are distinctive within Bridge2AI in that they are non-clinical (from tissue

2



bioRxiv preprint

doi:

https://doi.org/10.1101/2024.05.21.589311

;

this version posted May 24, 2024.

The copyright holder for this preprint (which

was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made

available under a

CC-BY-NC-ND 4.0 International license
.

Figure 1 – HiView visualization of a hierarchical cell map as a circle-packing diagram. (A)
A unified cell map of the interaction network in HEK293 cells generated by integrating
immunofluorescence images and affinity purification interactions for 661 proteins.2 The
hierarchical cell map contains 10 layers of depth representing communities at multiple
resolutions, such that communities at smaller distance (lighter circles) are nested inside larger
communities (darker circles) similar to physical compartments within a cell (the “root” structure).
HiView can be used to zoom in on nested communities, such as (B) the nuclear splicing speckle
(depth layer 4) and (C) the spliceosomal complex family (depth layer 5), to begin to resolve
smaller interaction communities and even individual proteins. For each community, HiView
provides the underlying interaction data that support it. For example, (D) shows the protein-
protein interaction network within the nuclear splicing speckle community, where yellow edges
represent similarities in the embedding from both protein image and affinity purification data.
HiView also provides a list of proteins within each community. For example, (E) shows the list of
proteins (15 total) located within the spliceosomal complex family.

cultures) and are considered to be de-identified as they cannot be matched, with current
knowledge, to a human subject.

AI-Readiness has been defined in multiple ways in the literature. Regardless of the
specific definitions adopted, in all cases significant requirements are placed on the
datasets used to train AI/ML models and on datasets analyzed using these AI/ML
models. Datasets that meet these requirements are called “AI-Ready”. We define the
following criteria for purposes of this discussion:

(cid:0)

  FAIR – Datasets, and the software used to prepare them for AI/ML analysis,

must comply with the FAIR Principles which have been outlined for data and for
software; interoperability is a particular requirement of AI-Readiness17. FAIRness
is an NIH requirement and is consistent with best current scholarly practice.

3




bioRxiv preprint

doi:

https://doi.org/10.1101/2024.05.21.589311

;

this version posted May 24, 2024.

The copyright holder for this preprint (which

was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made

available under a

CC-BY-NC-ND 4.0 International license
.

(cid:0)

(cid:0)

(cid:0)

  Provenanced – Provenance Graphs of computations, data, software, and
models used to prepare the dataset must be available as metadata18–20.

(cid:0)
  Characterized – Complete schemas, validation procedures, and data sheets21
for all datasets in the provenance graph; and model cards22 for all models used
to prepare the dataset, are resolvable from the dataset’s metadata.

  Explainable – Dataset must be fully provenanced (see above) and statistically

characterized within any limitations and constraints, including ethical
considerations and limitations; and this characterization is available in the
metadata23–26;

  Ethical – The dataset was obtained, validated, documented, licensed, and

distributed ethically27. Ethical conditions will include but are not limited to: ethical
treatment of subjects including protection of human subjects data and ethical
treatment of animals; proper non-biasing data analysis and scientific integrity of
methods, conclusions, and publications; full documentation of scientific methods
and reagents; and licensing and distribution of data, software, and models with
openness to the scientific community for reuse consistent with protection of
human subjects privacy and adhering to responsible research conduct;
moreover, this requirement involves anticipating potential applications of the data
and derived AI/ML systems, integrating them into data governance frameworks to
optimize benefits and mitigate potential risks, and promoting their use for the
collective good.

These criteria are interdependent. For example, ethical validation capability depends
upon explainability, explainability depends upon provenance, and so forth.

CM4AI uses an expanded version of the FAIRSCAPE framework to establish a basis for
AI-readiness. As we progress through the project, our AI-readiness features will become
ever more complete. At present they comprise: (a) establishing FAIRness including rich
metadata and persistent, globally unique identifiers; (b) computing a machine-readable
provenance graph, resolvable as metadata for all results, including inputs,
computations, software, and outputs; and (c) characterizing and validating all datasets
and data elements, and mapping data elements to public ontology vocabularies, where
appropriate, using JSON-Schema mini-data-dictionary descriptions resolvable from the
provenance graph metadata. Further AI-Readiness capabilities on our near-term
research agenda are described in the Methods section.

Methods

1. Cell Lines
CM4AI has elected to use the cancer cell line MDA-MB-46828 (+/- treatment with
paclitaxel and vorinostat) and the iPSC line KOLF2.1J29 (+/- neuronal and
cardiomyocytic differentiation), both of which have been ethically sourced.

MDA-MB-468 (RRID:CVCL_0419) is a triple negative breast cancer cell line established
from a metastatic site pleural effusion of a 51-year-old black female with a metastatic
mammary adenocarcinoma, available from ATCC. This cell line has been extensively
used to study triple-negative breast cancer and is well characterized with data such as
transcriptomic, mutational profile and whole-genome sequencing available.

4



bioRxiv preprint

doi:

https://doi.org/10.1101/2024.05.21.589311

;

this version posted May 24, 2024.

The copyright holder for this preprint (which

was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made

available under a

CC-BY-NC-ND 4.0 International license
.

The KOLF2.1J (RRID:CVCL_B5P3) iPSCs cell line was derived from a healthy male
Northern European donor, available from the Human Induced Pluripotent Stem Cells
Initiative (HipSci) resource30. It is available for access by non-for-profit organizations via
a simple MTA.

2. Data Acquisition

2.1 Protein-Protein Interaction

To map protein-protein interactions (PPIs) of 100 chromatin regulators under different
conditions (cancer: no treatment, paclitaxel or vorinostat; iPSC: undifferentiated,
neurons and cardiomyocytes), we employed AP-MS on endogenously tagged cell lines
and size-exclusion chromatography coupled to mass spectrometry (SEC-MS) for
proteome-wide complex/interaction mapping as two orthogonal mass spectrometry-
based approaches. We have endogenously tagged 17 genes in MDA-MB-468 and
acquired AP-MS data under three conditions (untreated, paclitaxel and vorinostat
treated) . We are currently in the process of tagging 34 additional genes inMDA-MB-468
cells . In addition, we have performed SEC-MS on MDA-MB-468 cells under three
conditions (untreated, paclitaxel and vorinostat treated), which enabled us to identify
72/100 chromatin modifiers, with 52 of them being integral components of protein
complexes. In addition to chromatin modifiers of interest, SEC-MS allowed us to map
complexes in MDA-MB-468 cells and investigate the impact of treatment proteome-
wide. We detected PPI profiles of over a thousand complexes in MDA-MB468 cells, with
thousands of proteins exhibiting differential elution profiles between control and
paclitaxel or vorinostat treated cells. SEC-MS has also been performed in KOLF2 iPSCs
and derivatives, uncovering more than 700 protein complexes in both parental iPSCs
and differentiated neurons.

2.2 Spatial proteomics mapping

For Year 1, we proposed to map the spatial subcellular organization of key chromatin
modifiers, their interactors and key signaling molecules involved in cancer using the
Human Protein Atlas resource of antibodies. We established automated fixation and
permeabilization protocols for the pipetting robot for MDA-MB-468 and KOLF2 cell lines,
and completed spatial proteomics mapping of 100 chromatin regulators in the MDA-MB-
468 cells (+/- paclitaxel or vorinostat), with another 500 proteins pending significant hits
from the genetic perturbations and PPI results. The first set of images was released as
input for the CM4AI Tools Module’s MuSIC pipeline and processed to RO-Crate
outputs.

2.3 Genetic perturbation mapping

For Year 1, we performed single-cell CRISPR screens perturbing 100 chromatin
regulators in the MDA-MB-468 cells under 3 conditions (no treatment, paclitaxel or
vorinostat) and KOLF2 iPSC in the undifferentiated state. We have designed a CRISPR
lentiviral library targeting 100 chromatin factors with 6 guide RNAs per gene. We
generated and characterized MDA-MB-468 and KOLF2 CRISPR lines expressing
inducible dCas9. Single-cell CRISPR screens of 100 chromatin regulators (+/– drug) in
MDA-MB-468 cells and in undifferentiated KOLF2.1J iPSCs were done using the 10x
Genomics 3’HT kit. The resulting data are currently being QCed.

5



bioRxiv preprint

doi:

https://doi.org/10.1101/2024.05.21.589311

;

this version posted May 24, 2024.

The copyright holder for this preprint (which

was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made

available under a

CC-BY-NC-ND 4.0 International license
.

3. Tools: Cell Map Data Integration Pipeline and VNN Tools

The Tools Module of CM4AI is responsible for producing and maintaining a dataset
integration and map production system, the Multi-Scale Integrated Cell (MuSIC)
pipeline, which produces integrated cell maps from the multiple, multi-modal input data
streams. The MuSIC pipeline is divided into multiple segments, each of which calls the
FAIRSCAPE client package to validate inputs and create the output RO-Crate package
for that segment which, in turn, is ready for dispatch to the FAIRSCAPE server for PID
creation and provenance graph entailment computation. FAIRSCAPE-generated PIDs
resolve to complete human- and machine-readable metadata including licensing
information, and provide links to the underlying datasets.

MuSIC pipeline segments and their functions are shown below and in Figure 2. The
steps are:

a)  Download PPI and Image Data

- PPI networks for the cell line and conditions are downloaded from the Krogan
Laboratory’s deposition archive and made available for further processing.

-  IF images for the cell line and conditions are downloaded from the Lundberg
Laboratory’s deposition archive and made available for further processing.

b)  Generate embeddings
- Downloaded PPI networks are processed using the node2vec deep learning model31
to reduce dimensionality, producing a PPI embedding that contains information about
the protein’s interactions.
- IF images are processed using a Human Protein Atlas deep learning model32 to
reduce dimensionality, producing an image embedding that contains information about
protein localization.

c)  Co-Embedding

- The PPI and image embeddings are integrated to obtain a co-embedding for each
protein using contrastive deep learning. This integration model learns co-embeddings
such that the original embeddings can be reconstructed with minimal information loss
and proteins with similar PPI and image embeddings have similar co-embeddings33.

d)  Protein Community Detection and Hierarchy Creation

- Community detection is performed based on the all-by-all similarities of pairs of
proteins in the co-embedding space. The resulting hierarchy of protein communities is
output as the final cell map.

e)  Hierarchy evaluation

- Cell maps are annotated in a two pronged approach. First, cell maps are aligned to
known protein function and pathway resources, including the Gene Ontology (GO)34 and
Reactome35, to determine protein assemblies in the cell map with high overlap with
known cell biology. Second, assemblies in the cell maps are annotated using a large
language model (LLM) approach that we developed to name sets of proteins and assign
a name confidence score.

6



bioRxiv preprint

doi:

https://doi.org/10.1101/2024.05.21.589311

;

this version posted May 24, 2024.

The copyright holder for this preprint (which

was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made

available under a

CC-BY-NC-ND 4.0 International license
.

Figure 2 – Tools Module: Data Integration and Map Production Pipeline

With the Standards and Data Acquisition Modules, we developed an alpha release of
software that interfaces the MuSIC pipeline with the FAIRSCAPE infrastructure and
data repository. The system includes data dictionaries, formatting standards, and a
FAIRSCAPE metadata and provenance API, implemented as a command line software
package.

f) Integrative structure modeling

We are currently exploring the feasibility of integrative structure modeling of the MuSIC
communities. The resulting structural models will increase our understanding of the
MuSIC communities and help with planning future experiments.

To begin, we developed a bioinformatics pipeline for annotating MuSIC communities by
the available structural information about the community members and their
interactions. This structural information includes data from the Protein Data Bank
(PDB)36, AlphaFold Protein Structure Database (AlphaFoldDB)37, crosslinking mass
spectrometry, and prediction of disordered sequence segments. We then ranked the
MuSIC communities by the amount of available structural information, serving as a
proxy for the feasibility of integrative structure modeling. Finally, we performed
integrative structure modeling of a few top ranked communities. This process uses our
standard integrative structure modeling framework, which proceeds through the
following four stages38–40: (i) gathering input information, (ii) representing subunits and
translating data into spatial restraints, (iii) configurational sampling to produce an
ensemble of models that satisfy the restraints, and (iv) analyzing and validating input
information and models. The modeling protocol was scripted using the Python Modeling
Interface package, a library for modeling macromolecular complexes based on our
open-source Integrative Modeling Platform (IMP) package version 2.18
(https://integrativemodeling.org).

7






bioRxiv preprint

doi:

https://doi.org/10.1101/2024.05.21.589311

;

this version posted May 24, 2024.

The copyright holder for this preprint (which

was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made

available under a

CC-BY-NC-ND 4.0 International license
.

4. Standards: AI-Readiness Packaging

The Standards Module of CM4AI is responsible for AI-Readiness packaging of data,
E
software, metadata and provenance graphs. This task is performed by the FAIRSCAPE
framework. The integration of FAIRSCAPE with CM4AI Data Acquisition and the Tools
data integration pipeline is shown in Figure 3.

Figure 3 – FAIRSCAPE AI-Readiness framework as applied to CM4AI datasets.

FAIRSCAPE consists of a client-side Python3 application, called either from the
command line or as a set of Python functions by the Tools Module’s data integration
pipeline, and a server application, also in Python3, which completes the packaging.

The client-side package, FAIRSCAPE-CLI, is called when any computation or coherent
set of computations in the pipeline is completed, and it is passed metadata which
defines schemas in JSON-Schema41 for the datasets in the computational unit, as well
as the inputs, computations, software, models, and outputs. FAIRSCAPE-CLI creates
an RO-Crate package with the datasets, metadata, and software – or resolvable
references to these components – and unique stubs for identifier creation on each of
these components.

The RO-Crates are then sent to the FAIRSCAPE server where they are registered and
assigned persistent, resolvable, globally unique persistent identifiers (PIDs). The RO-
Crates are then decomposed into their individual components – datasets, models,
software – which are also registered and assigned PIDs. The PID system currently in
use is the ARK scheme42 – with DOIs a future feature as supplementary PIDs for final-
state publishable work.

Lastly, the server computes end-to-end entailments on each RO-Crate’s provenance as
s
expressed in the EVI Evidence Graph Ontology and links them together where possible.
e.
g
A graphical view of an evidence graph presented as part of the human-readable landing
in
page for a CM4AI RO-Crate package, is shown in Figure 4. Alternate views serialized in
JSON-LD and in RDF-XML are also provided on the landing page. The complete RO-

8





bioRxiv preprint

doi:

https://doi.org/10.1101/2024.05.21.589311

;

this version posted May 24, 2024.

The copyright holder for this preprint (which

was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made

available under a

CC-BY-NC-ND 4.0 International license
.

Crate is provided as Supplemental Data, and all RO-Crates for the current release are
available in the University of

Figure 4. Evidence graph provenance visualization from human-readable landing page.
This landing page is displayed on resolution of the RO-Crate’s ARK globally unique persistent
ID. It was generated on an RO-Crate package from CM4AI’s 0.5 alpha data release. Landing
page navigation tabs allow full metadata on the package to be displayed in human readable
form, or alternatively in JSON-LD serialization.

Virginia’s LibraData data archive, an instance of Harvard’s Dataverse. Links to the
archived RO-Crates are provided in the Data and Software Availability Statement.

PIDs generated by the server resolve to machine- and human-readable landing pages
containing the metadata, expressed in the JSON-LD graph language using vocabularies
from Schema.org, EVI, and other well-defined public ontologies. The complete RO-
Crate dataset is attached as Supplemental Data.

Both the client-side and sever-side FAIRSCAPE packages are PIP-installable and are
freely available, under the MIT open-source license (see Data and Software Availability
Statement).

5. Teaming

The CM4AI project's Teaming Module supports integration and expansion of technical /
scientific expertise in CM4AI, emphasizing ethical considerations and diverse
interdisciplinary perspectives. It facilitates effective communication and collaboration
among CM4AI investigators and personnel, of varied geographic, disciplinary, and
cultural backgrounds.

Teaming supports open dissemination of CM4AI-generated data, maps, and tools via
the CM4AI web portal (http://www.cm4ai.org), using the U-BRITE platform43, to support
open and trustworthy data management and sharing, ensuring the broad dissemination
of the datasets, cell maps, and tools developed. This effort aligns with our commitment

9




bioRxiv preprint

doi:

https://doi.org/10.1101/2024.05.21.589311

;

this version posted May 24, 2024.

The copyright holder for this preprint (which

was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made

available under a

CC-BY-NC-ND 4.0 International license
.

to FAIR principles and high-performance access, making our deliverables both
accessible and useful to users.

Teaming has constructed an extensive scholarly network called a talent graph, based
initially on publications of CM4AI investigators, which links to each person’s peer-
reviewed research publications. Knowledge graphs involving this extended community
of experts, publications, data sets, and tools are captured through this effort.

Collaboration with the NIH Common Fund Data Ecosystem (CFDE) for data curation
helps integrate CM4AI contributions into the broader scientific community, and
engagement with the Bridge2AI Center helps us identify novel channels for
broadcasting our achievements, ensuring the outreach of our data, maps, and tools to
those who can utilize them fully.

The Teaming module fosters a culture of open science and community involvement with
mailing lists, shared online collaborating folders, shared documents, and remote
messaging and collaboration software within CM4AI, across the Bridge2AI program,
and in the broader scientific community, to promote an ethical, inclusive, and
collaborative scientific environment. In collaboration with the Skills and Workforce
Development module, it contributes significantly to enhancing the biomedical AI/ML
workforce and fostering innovation in the field.

6. Skills and Workforce Development

The Skills and Workforce Development module works to enhance the biomedical AI/ML
workforce by recruiting and training a diverse group of participants in the scientific
approach, data standards, and tools that can be used to accelerate innovation from
CM4AI. To recruit and prepare a biomedical workforce that is adept at both the data and
life sciences, new training that will allow researchers to leverage the data sets produced
by Bridge2AI and other programs is needed.  To achieve this goal, we aim to broadly
distribute the data and tools developed as part of the CM4AI project and provide broad
training in these components. We are leveraging a multimodal approach to content
delivery which includes asynchronous virtual training, hosted virtual events, and an in-
person internship being co-hosted at two of our sites.

A key focus of the Skills and Workforce Development module is to cultivate a diverse
biomedical AI/ML workforce.  Our recruitment approach has included communication
with investigators involved in the Bridge2AI program, integration with trainees at
organizations involved with Bridge2AI, and community outreach to other academic
organizations with an emphasis on those serving underrepresented communities. For
our inaugural virtual CodeFest event, we provided training to a total of 38 registrants
where 40% of the attendees identified as female and 30% came from underrepresented
communities. We have also developed a Diversity, Equity, and Inclusion (DEI)
Committee with representation from all Cores to direct initiatives which has at its core to
partner with historically black colleges and universities (Meharry Medical and
Morehouse School of Medicine) to increase the number of learners in the AI/ML field at
all levels:  graduate students:  masters and PhD, postdoctoral, and junior investigators.

Our in-person internship will be hosted at Yale University and the University of
California San Diego, with students jointly working on a final project between the two

10



bioRxiv preprint

doi:

https://doi.org/10.1101/2024.05.21.589311

;

this version posted May 24, 2024.

The copyright holder for this preprint (which

was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made

available under a

CC-BY-NC-ND 4.0 International license
.

sites and supervised by faculty from both institutions. This internship will provide
participants with the opportunity to learn 1) the impact of cell maps on biomedical
research applications, 2) how to interact with the data and standards used by CM4AI, 3)
how to develop a visible neural network using cell maps, and 4) how to integrate
external data into an existing cell map for the ethical development of AI/ML applications.

7. Ethics

The Ethics team worked closely with Data Acquisition, Tools, and Standards Modules to
develop a plan for ethical preparation, licensing, dissemination, and Data Access
supervision of CM4AI datasets. This plan balances the openness of the data produced,
with protection of the intellectual property, and supervision and monitoring of potential
efforts at commercialization. In addition to CM4AI data governance initiatives,
comprehensive guidelines and best practices are being established for the ethical
development of AI systems leveraging this data. The Ethics team employs a
methodology inspired by Value-Sensitive Design (VSD), facilitating the creation of
normative standards and expectations aimed at ensuring the responsible design of
datasets and AI technologies, in alignment with core values. This entails the Ethics
team engaging in conceptual efforts to craft an axiological repository—a repository of
values—to guide the articulation of design standards and expectations. Further
advancing this work, the Ethics team is instrumental in formulating the CM4AI Life
Cycle, a framework designed to elucidate governance milestones throughout the data
generation and AI system development processes. This lifecycle framework
underscores the commitment to embedding ethical considerations at every phase,
ensuring accountability and value alignment from inception through deployment.

Recognizing the importance of diverse perspectives in shaping ethical guidelines, the
Ethics team is dedicated to conducting mixed empirical research, combining qualitative
and quantitative methods. This research is critical for capturing a broad spectrum of
community insights, thus enriching the development of guidelines and best practices
with a multifaceted understanding of stakeholder values and concerns. Moreover, the
team is set to develop a suite of tools aimed at enhancing awareness and
comprehension of the ethical, legal, and social ramifications associated with CM4AI
data and the consequent AI systems. These tools are envisioned to bolster ethical
decision-making, providing vital support for navigating the complex landscape of AI
ethics. This comprehensive approach not only aims to elevate the standards of AI
development within CM4AI but also seeks to contribute to the broader discourse on
responsible AI, promoting a culture of ethical integrity, inclusivity, and societal benefit.

Results and Progress to Date

The CM4AI project has made significant progress with efforts by each of our six
modules (Data, Tools, Standards, Ethics, Skills, Teaming) relevant to each of the three
program pillars: Data – composed of the Data, Tools, and Standards Modules; People –
composed of the Skills and Teaming Modules; and Ethics – the Ethics Module.

The Data Module completed parallel maps of protein-protein interactions, protein
subcellular distributions, and transcriptional states associated with a near-
comprehensive set of chromatin regulators encoded by the human genome (Pillar:
Data). All of these data were generated in ethically sourced breast cancer cells (Pillar:

11



bioRxiv preprint

doi:

https://doi.org/10.1101/2024.05.21.589311

;

this version posted May 24, 2024.

The copyright holder for this preprint (which

was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made

available under a

CC-BY-NC-ND 4.0 International license
.

Ethics) and will be extended to ethically sourced human induced pluripotent stem cells
(iPS) in Year 2.

The Tools Module developed an early (alpha) release of the Multi-Scale Integrated Cell
(MuSIC) toolkit for assembling large-scale maps of human cells from fusion of protein
interaction and imaging data, including a highly automated pipeline that integrates with
Standards FAIRSCAPE toolkits. We expect to submit a mature version for publication
as a Protocols paper in early fall

The Standards Module established FAIR AI-Readiness metadata and validation toolkits,
which have been used to package and release our cell-line perturbation and
measurement data.

The Ethics Pillar established guidelines addressing socio-ethical challenges, positioning
us to now embark on stakeholder engagement and explainable AI development.
Preliminary work has been done to develop a first version of the Value Repository to
support the establishment of these guidelines44,45. The Ethics Pillar also supported the
work initiated by the Bridge2AI Ethics Working Group for supporting the Bridge2AI Data
Sharing and Dissemination Working Group in developing a Code of Conduct. The Code
of Conduct outlines basic requirements for participants in the Bridge2AI Open House to
access B2AI data (including CM4AI datasets), to attest and commit to it prior to
accessing the data.

The Workforce Development Module identified gaps in training of the functional
genomic and biomedical AI/ML workforce; and led the organization of a highly
successful CodeFest in March 2024, in collaboration with the Tools, Standards, Data,
and Teaming Modules. The CodeFest was attended by both within-project and external
participants. Our Workforce Development Module is now transitioning towards
collaborative content creation and delivery (Pillar: People).

Finally, the Teaming Module facilitated cross-disciplinary collaboration, completed the
open-access CM4AI web portal, and initiated a Diversity Equity and Inclusion (DEI)
Committee (Pillar: People). Beyond these specific activities, CM4AI has been actively
engaged in activities of the greater Bridge2AI Consortium, including a face-to-face
meeting in Pentagon City, Virginia; the overall Bridge2AI Steering Committee and
workgroups; and submission of numerous auxiliary concepts for supplemental funding
and a U-24 submission for cross-cutting integration across projects in the NIH Common
Fund Data Ecosystem.

Initial Datasets, Tools and Standards associated with all three parallel mapping
platforms have been released and are publicly available at our CM4AI web portal30.
These data were also made available at our CodeFest in March 2024, accompanied by
detailed tutorials explaining how the data were derived and integrated, and providing
production ready software for producing the datasets. This first data release was made
possible by intensive cross-module collaboration by dozens of staff as well as end-to-
end operation of all aspects of our CM4AI platform, from data generation to data
analysis by advanced toolsets to FAIR-compliant ethical AI-readiness standards for data
access and distribution.

12



bioRxiv preprint

doi:

https://doi.org/10.1101/2024.05.21.589311

;

this version posted May 24, 2024.

The copyright holder for this preprint (which

was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made

available under a

CC-BY-NC-ND 4.0 International license
.

Data packages from CM4AI including links to software used in preparing the data, input
datasets, dataset schemas, and deep provenance graphs, are available on the CM4AI
web portal and will soon be available on the Dataverse NIH approved generalist
repository.

All CM4AI integration and packaging software is freely available and licensed under
nonrestrictive open-source licenses.  Data packages are licensed under Creative
Commons CC-BY-NC-SA 4.0 license terms31, which include a requirement for
attribution to the dataset authors and the copyright holding institutions, and citation of
this article.  Commercial use requires a separate license negotiation with the copyright
holder (UCSD, Stanford, and/or UCSF depending upon the specific data package in
question). A Data Access Committee will supervise ethical matters related to dataset
distribution and potential dual licensing for commercial use.

Conclusions

The Bridge2AI Functional Genomics project, “Cell Maps for Artificial Intelligence”
(CM4AI), is a continuing effort which has already published valuable and re-usable
datasets and software, advanced the notions of ethical AI-readiness, developed a
strong teaming approach within the project including a shared portal, and provided
hands-on training to the research community in use of these tools.

This article presents the work as a whole in preprint form and will be continuously
updated as CM4AI progresses, with citations to detailed publications in each
contributing area.

With these measures, it is our intention that CM4AI become a highly transformative
platform for ethical, explainable, and interpretable biomedical AI, deepen our
understanding of processes of human disease and health, help to train a new
generation of researchers, enable development of novel cures, and assist researchers
and clinicians in equitably improving human lives.

Data and Software Availability Statement

The most recent data and metadata produced by CM4AI are licensed for reuse under
the Creative Commons Attribution Non-Commercial Share-Alike International 4.0
License  (https://creativecommons.org/licenses/by-nc-sa/4.0/) and are available in
LibraData, the University of Virginia’s archival data repository:

Clark T, Mohan J, Schaffer L, Obernier K, et al. Cell Maps for Artificial
Intelligence - Data Release", https://doi.org/10.18130/V3/DXWOS5, University of
Virginia Dataverse, V1

LibraData is the University of Virginia’s instance of Dataverse, an NIH-approved
generalist repository.

Attribution requirements for these datasets include attribution to the copyright holders
and the Cell Maps for Artificial Intelligence project, as referenced in the datasets, and
citation of the present article:

13



bioRxiv preprint

doi:

https://doi.org/10.1101/2024.05.21.589311

;

this version posted May 24, 2024.

The copyright holder for this preprint (which

was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made

available under a

CC-BY-NC-ND 4.0 International license
.

Clark T, Mohan J, Schaffer L, Obernier K, Al Manir S, Churas CP, Dailamy A,
Doctor Y, Forget A, Hansen JN, Hu M, Lenkiewicz J, Levinson MA, Marquez C,
Nourreddine S, Niestroy J, Pratt D, Qian G, Thaker S, Bélisle-Pipon J-C, Brandt
C, Chen J, Ding Y, Fodeh S, Krogan N, Lundberg E, Mali P, Payne-Foster P,
Ratcliffe S, Ravitsky V, Sali A, Schulz W, Ideker T. Cell Maps for Artificial
 Intelligence: AI-Ready Maps of Human Cell Architecture from  Disease-Relevant
Cell Lines. BioRXiv, May 2024.

Copyright © 2024 to these datasets is held by the Regents of the University of
California, except where otherwise indicated. Raw image data for spatial proteomics is
copyright © 2024 by The Board of Trustees of the Leland Stanford Junior University.

Software comprising the Tools data integration pipeline is available under BSD-3 open
source license in GitHub (for alpha level tools) or in the Zenodo long-term archive (for
production-ready tools). Links to these tools are packaged in RO-Crates as part of
CM4AI provenance representations.

The FAIRSCAPE AI-readiness framework is described with a tutorial and instructions
for installation at https://fairscape.github.io. Copyright © 2024 to this software is held by
the Rector and Board of Visitors of the University of Virginia. It is available under open
source MIT License.

The open source Integrative Modeling Platform (IMP) package is available at
http://integrativemodeling.org.

Software packages will be versioned as the project progresses, and the version used to
produce each dataset will be referenced in the dataset’s metadata.

Acknowledgements

This work was funded by the National Institutes of Health under awards
1OT2OD032742-01 (Bridge2AI Functional Genomics) and 5U54HG012513-02
(Bridge2AI Bridge Center), and by the Frederick Thomas Fund of the University of
Virginia.

We also acknowledge the helpful assistance of members of the other Bridge2AI
components, which helped to make this work possible, including but not limited to the
Bridge2AI Standards Working Group; the Bridge2AI Data Sharing and Dissemination
Working Group; and members of the Bridge2AI CHoRUS Critical Care Medicine Data
Generation Project.

REFERENCES

1.  National Institutes of Health. Bridge to Artificial Intelligence (Bridge2AI). Published

online February 7, 2023. Accessed February 9, 2023.
https://commonfund.nih.gov/bridge2ai

2.  Qin Y, Huttlin EL, Winsnes CF, et al. A multi-scale map of cell structure fusing

protein images and interactions. Nature. 2021;600(7889):536-542.
doi:10.1038/s41586-021-04115-9

14



bioRxiv preprint

doi:

https://doi.org/10.1101/2024.05.21.589311

;

this version posted May 24, 2024.

The copyright holder for this preprint (which

was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made

available under a

CC-BY-NC-ND 4.0 International license
.

3.  RO-Crate Community. Research Object Crate (RO-Crate). Published online October

26, 2023. https://www.researchobject.org/ro-crate/

4.  Soiland-Reyes S, Sefton P, Crosas M, et al. Packaging research artefacts with RO-

Crate. Published online August 13, 2021. doi:10.5281/ZENODO.5146227

5.  Levinson MA, Niestroy J, Al Manir S, et al. FAIRSCAPE: a Framework for FAIR and

Reproducible Biomedical Analytics. Neuroinform. 2022;20(1):187-202.
doi:10.1007/s12021-021-09529-4

6.  Ma J, Yu MK, Fong S, et al. Using deep learning to model the hierarchical structure
and function of a cell. Nat Methods. 2018;15(4):290-298. doi:10.1038/nmeth.4627

7.  Yu MK, Ma J, Fisher J, Kreisberg JF, Raphael BJ, Ideker T. Visible Machine

Learning for Biomedicine. Cell. 2018;173(7):1562-1565.
doi:10.1016/j.cell.2018.05.056

8.  Kuenzi BM, Park J, Fong SH, et al. Predicting Drug Response and Synergy Using a
Deep Learning Model of Human Cancer Cells. Cancer Cell. 2020;38(5):672-684.e6.
doi:10.1016/j.ccell.2020.09.014

9.  Zheng F, Kelly MR, Ramms DJ, et al. Interpretation of cancer mutations using a

multiscale map of protein systems. Science. 2021;374(6563):eabf3067.
doi:10.1126/science.abf3067

10. Ma J, Fong SH, Luo Y, et al. Few-shot learning creates predictive models of drug
response that translate from high-throughput screens to individual patients. Nat
Cancer. 2021;2(2):233-244. doi:10.1038/s43018-020-00169-2

11. Thul PJ, Åkesson L, Wiking M, et al. A subcellular map of the human proteome.

Science. 2017;356(6340):eaal3321. doi:10.1126/science.aal3321

12. Singhal A, Cao S, Churas C, et al. Multiscale community detection in Cytoscape.

Przytycka TM, ed. PLoS Comput Biol. 2020;16(10):e1008239.
doi:10.1371/journal.pcbi.1008239

13. Pratt D, Chen J, Pillich R, et al. NDEx 2.0: A Clearinghouse for Research on Cancer
Pathways. Cancer Research. 2017;77(21):e58-e61. doi:10.1158/0008-5472.CAN-
17-0606

14. Shannon P. Cytoscape: A Software Environment for Integrated Models of

Biomolecular Interaction Networks. Genome Research. 2003;13(11):2498-2504.
doi:10.1101/gr.1239303

15. Yu MK, Ma J, Ono K, et al. DDOT: A Swiss Army Knife for Investigating Data-Driven

Biological Ontologies. Cell Syst. 2019;8(3):267-273.e3.
doi:10.1016/j.cels.2019.02.003

15



bioRxiv preprint

doi:

https://doi.org/10.1101/2024.05.21.589311

;

this version posted May 24, 2024.

The copyright holder for this preprint (which

was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made

available under a

CC-BY-NC-ND 4.0 International license
.

16. Cytoscape Consortium. ndex2. Published online February 8, 2024.

https://ndex2.readthedocs.io/en/latest/?badge=latest

17. Wilkinson MD, Dumontier M, Aalbersberg IjJ, et al. The FAIR Guiding Principles for

scientific data management and stewardship. Sci Data. 2016;3:160018.
doi:10.1038/sdata.2016.18

18. Al Manir S, Niestroy J, Levinson MA, Clark T. Evidence Graphs: Supporting

Transparent and FAIR Computation, with Defeasible Reasoning on Data, Methods,
and Results. In: Glavic B, Braganholo V, Koop D, eds. Provenance and Annotation
of Data and Processes. Vol 12839. Lecture Notes in Computer Science. Springer
International Publishing; 2021:39-50. doi:10.1007/978-3-030-80960-7_3

19. Al Manir S, Niestroy J, Levinson M, Clark T. EVI: The Evidence Graph Ontology,

OWL 2 Vocabulary. Published online March 23, 2021.
https://doi.org/10.5281/zenodo.4630931

20. Lebo T, Sahoo S, McGuinness D, et al. PROV-O: The PROV Ontology W3C

Recommendation 30 April 2013. Published online 2013. http://www.w3.org/TR/prov-
o/

21. Gebru T, Morgenstern J, Vecchione B, et al. Datasheets for Datasets.

arXiv:180309010 [cs]. Published online March 19, 2020. Accessed August 4, 2021.
http://arxiv.org/abs/1803.09010

22. Mitchell M, Wu S, Zaldivar A, et al. Model Cards for Model Reporting. Proceedings
of the Conference on Fairness, Accountability, and Transparency. Published online
January 29, 2019:220-229. doi:10.1145/3287560.3287596

23. Arbelaez Ossa L, Starke G, Lorenzini G, Vogt JE, Shaw DM, Elger BS. Re-focusing

explainability in medicine. DIGITAL HEALTH. 2022;8:205520762210744.
doi:10.1177/20552076221074488

24. Combi C, Amico B, Bellazzi R, et al. A manifesto on explainability for artificial
intelligence in medicine. Artificial Intelligence in Medicine. 2022;133:102423.
doi:10.1016/j.artmed.2022.102423

25. Kundu S. AI in medicine must be explainable. Nat Med. 2021;27(8):1328-1328.

doi:10.1038/s41591-021-01461-z

26. Amann J, Blasimme A, Vayena E, Frey D, Madai VI, Precise4Q consortium.

Explainability for artificial intelligence in healthcare: a multidisciplinary perspective.
BMC Med Inform Decis Mak. 2020;20(1):310. doi:10.1186/s12911-020-01332-6

27. Floridi L. A Unified Framework of Ethical Principles for AI. In: The Ethics of Artificial

Intelligence. Oxford University Press.
https://doi.org/10.1093/oso/9780198883098.001.0001

16



bioRxiv preprint

doi:

https://doi.org/10.1101/2024.05.21.589311

;

this version posted May 24, 2024.

The copyright holder for this preprint (which

was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made

available under a

CC-BY-NC-ND 4.0 International license
.

28. Cailleau R, Olivé M, Cruciger QVJ. Long-term human breast carcinoma cell lines of

metastatic origin: Preliminary characterization. In Vitro. 1978;14(11):911-915.
doi:10.1007/BF02616120

29. Pantazis CB, Yang A, Lara E, et al. A reference human induced pluripotent stem cell

line for large-scale collaborative studies. Cell Stem Cell. 2022;29(12):1685-
1702.e22. doi:10.1016/j.stem.2022.11.004

30. Streeter I, Harrison PW, Faulconbridge A, et al. The human-induced pluripotent
stem cell initiative—data resources for cellular genetics. Nucleic Acids Res.
2017;45(D1):D691-D697. doi:10.1093/nar/gkw928

31. Grover A, Leskovec J. node2vec: Scalable Feature Learning for Networks. In:

Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge
Discovery and Data Mining. ACM; 2016:855-864. doi:10.1145/2939672.2939754

32. Le T, Winsnes CF, Axelsson U, et al. Analysis of the Human Protein Atlas Weakly
Supervised Single-Cell Classification competition. Nat Methods. 2022;19(10):1221-
1229. doi:10.1038/s41592-022-01606-z

33. Bao F, Deng Y, Wan S, et al. Integrative spatial analysis of cell morphologies and

transcriptional states with MUSE. Nat Biotechnol. 2022;40(8):1200-1209.
doi:10.1038/s41587-022-01251-z

34. Ashburner M, Ball C, Blake J, et al. Gene ontology: tool for the unification of biology.

The Gene Ontology Consortium. Nature Genetics. 2000;25(1):25-29.

35. Joshi-Tope G, Gillespie M, Vastrik I, et al. Reactome: a knowledgebase of biological

pathways. Nucleic Acids Res. 2005;33(Database issue):D428-32.
doi:10.1093/nar/gki072

36. Berman HM. The Protein Data Bank: a historical perspective. Acta crystallographica

Section A, Foundations of crystallography. 2008;64(Pt 1):88-95.
doi:10.1107/S0108767307035623

37. Varadi M, Anyango S, Deshpande M, et al. AlphaFold Protein Structure Database:
massively expanding the structural coverage of protein-sequence space with high-
accuracy models. Nucleic Acids Research. 2022;50(D1):D439-D444.
doi:10.1093/nar/gkab1061

38. Rout MP, Sali A. Principles for Integrative Structural Biology Studies. Cell.

2019;177(6):1384-1403. doi:10.1016/j.cell.2019.05.016

39. Sali A. From integrative structural biology to cell biology. Journal of Biological

Chemistry. 2021;296:100743. doi:10.1016/j.jbc.2021.100743

17



bioRxiv preprint

doi:

https://doi.org/10.1101/2024.05.21.589311

;

this version posted May 24, 2024.

The copyright holder for this preprint (which

was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made

available under a

CC-BY-NC-ND 4.0 International license
.

40. Russel D, Lasker K, Webb B, et al. Putting the pieces together: integrative modeling
platform software for structure determination of macromolecular assemblies. PLoS
Biol. 2012;10(1):e1001244. doi:10.1371/journal.pbio.1001244

41. Wright A, Andrews H, Hutton B, Dennis G. JSON Schema: A Media Type for
Describing JSON Documents. Internet Engineering Task Force (IETF); 2019.
https://datatracker.ietf.org/doc/html/draft-handrews-json-schema-02

42. Kunze J, Rodgers R. The ARK Identifier Scheme. Published online 2008.

https://escholarship.org/uc/item/9p9863nc

43. Cimino JJ, Liang WH, Wang J, et al. Empowering Team Science Across the
Translational Spectrum with the UAB Biomedical Research Infrastructure
Technology Enhancement (U-BRITE). In: 2020 IEEE 21st International Conference
on Information Reuse and Integration for Data Science (IRI). IEEE; 2020:194-200.
doi:10.1109/IRI49571.2020.00035

44. Victor G, Salem S, Bélisle-Pipon JC. A Scoping Review of Relevant Moral Values in

Health Sector AI Development. Published online 2023.
doi:10.17605/OSF.IO/BCVK3

45. Victor G, Bélisle-Pipon JC, Ravitsky V. Generative AI, Specific Moral Values: A

Closer Look at ChatGPT’s New Ethical Implications for Medical AI. The American
Journal of Bioethics. 2023;23(10):65-68. doi:10.1080/15265161.2023.2250311

18
