Integrating public single-cell transcriptomics and patient profiles to guide clinical development

Jack Russella Pollard¹, Joshua Whitener¹, Jordan Byck¹, Monika Manne¹, Hani Alostaz¹, Marianna Elia¹, Kamil Krukowski¹, Gabrielle Wong¹, Francisco Adrian¹, Liang Schweizer¹, Robert HI Andtbacka¹, & Christos Hatzis¹
¹HiFiBiO Therapeutics, Cambridge, MA USA
!

BACKGROUND

Single-cell technologies provide invaluable insights into disease biology and inform drug development by revealing
complex interactions among different cell types within patients. However, harnessing the potential of publicly
available single-cell data remains challenging due to the lack of integrated data across diverse single cell platforms.

To maximize the potential of single cell insights, we have created an AI/ML powered curation and data integration
process within our Drug Intelligence Science (DIS®) platform. This automated process identifies relevant published
studies and integrates single cell transcriptomic data from publicly available sources with in-house generated datasets
from our ex vivo translational efforts and our clinical programs.

MATERIALS AND METHODS

Dataset Search and Selection

Peer reviewed published datasets covering a variety of
human tissues in the healthy and diseased settings (cancer
and autoimmunity) are identified with a custom Python
based dataset crawler that utilizes keywords to surface
datasets and publications from various public repositories.

Dataset acquisition is aimed at improving indication selection
and combination strategies for our clinical programs. For
example, data from approximately 300 patients who received
standard of care (SOC) treatment including immune
checkpoint inhibitor (IO) therapy were selected for this
purpose.

After filtering datasets for the availability of raw count data
and vetting data quality via the recovery of the key findings
from the publications, our current dataset composition is
described in Figure 1.

Figure 1. DATASET COMPOSITION. ! Over 198 datasets were extracted from 186 publications (58 on
autoimmune disorders including 866 patients across 30 indications such as Crohn’s and Sjogren’s !
and 121 on cancer including 1476 patients across 39 indications such as lung adenocarcinoma and
renal cell carcinoma).

Metadata Extraction and Standardization

To structure the cell level data and render it suitable for
integration and analysis, we extract metadata from each
dataset/publication according to the guidelines for single cell
data reporting (1). Briefly, attributes detailing these aspects
are collected including:

• Sample processing-attributes (e.g., tissue dissociation, cell type
  enrichment, etc.)
• Data source (e.g., title, DOI, and number of patients), sample
  specific-attributes (e.g., indication, biopsy location, treatment, timepoint, response, etc.)
• Cell-level attributes (e.g., author annotated cell type/state, gene
  expression level, activation scores, clonotype, etc.)

These metadata are standardized where possible by
referencing appropriate external ontologies such as the
Disease Oncology (DO) for indications (2).

!Figure 2. CURATION AND STANDARDIZATION WORKFLOW. Once datasets are identified, manual
curation is used to extract both the raw data and the metadata in a standardized form. This
standardized form is then used to process and integrate the datasets in a fully automated fashion.

Generative Modeling for Single Cell Analysis Tasks

use of generative modeling for a variety of single cell tasks
such as integration, normalization and cell type/state
annotation. Typically, these tasks are done with best in
breed-tools (3) that then require rank-based or meta-analysis
to compare results across datasets. Generative modeling has
the advantage of using a single, integrated model for all
these tasks in an automated fashion and enables direct
comparisons across datasets for machine learning
applications.

For example, data integration and annotation harmonization
of diverse cell types/states across studies is a computational
challenge (4) that is ripe for generative approaches. Briefly,
using the autoimmune and cancer datasets, we pre-trained a
! model to annotate 31 immune and 4 non-immune cell
types/states in new data sets in an automated fashion. In the
latent space, new unlabeled data is integrated with data
from the pre-trained model. Transfer learning then assigns
the cell type/state annotation.

Figure 3. PRETRAINED MODEL FOR CELL TYPE / STATE CALLING. By integrating more than
377k cells drawn from 19 immune cell rich studies encompassing more than 411 samples
from 225 cancer and autoimmune patients and healthy controls, we have created an
algorithm to label cells from published and internal studies in an automated fashion.

RESULTS

Integrated Single Cell Database from Generative Modeling

Number of Cells By Indication

Representation of Cell Types in Diseases
!

! !Figure 4. CELL COMPOSITION. The current single-cell database comprises > 15 million cells analyzed in >5,000 samples collected from >2,500 patients
across 160 curated unique human studies, spanning oncology (53%), autoimmune disease (AID, 24%), and viral infection (14%), with the remaining being
from reference healthy tissues (9%). Only indications exceeding 0.09M cells are shown here.

The post-SOC resource was used to identify treatment settings where these mechanisms are enriched and markers co-expressed with targets of interest
(e.g., PD1, CTLA4, OX40, BTLA, TNFR2, etc.) in specific cell populations to inform combination strategies.

Cellect™: Multiuser Visualization and Interrogation Interface

Figure 5. CELLECT. To facilitate multiuser visualization and interrogation of the curated and integrated datasets, we have developed Cellect, a customized
version of the CellxGene Gateway (5). The system has a suite of tools that help both bench and computational scientists find, explore, and analyze our
curated single cell datasets. It includes several tools with features to engage with the data and aids in identifying cell types, clusters, and marker genes.
Shown in the figure is an integrated dataset of about 1-million cells from cancer patients treated with checkpoint inhibitors across a variety of indications.
Our generative modeling was able to standardize and integrate the data making comparisons across studies and treatments possible. This integrated
dataset was used to identify combination strategies for our clinical programs.

Post IO Combination Strategies Informed by Single Cell Data

APPLICATIONS

Figure 6. IDENTIFYING COMBINATIONS AND TARGET CELL TYPES. To explore the cell type/states associated to IO response and resistance, a collection of
about 475k cells from the primary tumor site of post IO treated patients was selected for analysis. Analysis across CD8 T-cells subsets that express several
surface targets revealed a distinct pattern of expression with respect to the validated PD1 target (PDCD1). Specifically, the B- and T-lymphocyte attenuator
(BTLA) and the Tumor necrosis factor (TNF) receptor 2 (TNFRSF1b) show patterns suggesting they maybe targets to activate distinct subsets of CD8 T-cells.
!

Post IO HFB200301 & HFB200603 May Target Distinct CD8 Subsets

HFB200301 Hypothesis

HFB200603 Hypothesis

Figure 7. HFB200301 & HFB200603 MAY TARGET DISTINCT CD8 SUBSETS. TNFR2 shows a pattern of expression consistent with it being an activating
receptor expressed on an alternat ive set of CD8 T-cells not expressing PD1, that if agonized with a compound such as HFB200301 these cells maybe able to
attack tumor cells. Alternatively, BTLA shows a pattern consistent with it being a coinhibitory receptor expressed with PD1 on CD8 cells in the post IO
treatment setting, that if antagonized with HFB200603 in combination with anti-PD1 might activate those cells to attack tumor cells.

CONCLUSIONS

We have presented an AI/ML guided approach to address the key challenge of integrating single-cell data across
platforms and demonstrated that relevant disease biology is retained upon integration. We outline a path for
deploying this solution at scale for bench and computational scientists to guide target as well as indication selection,
as was done for our ongoing clinical programs, including our first-in-class TNFR2 agonist (HFB200301, NCT05238883)
and our BTLA antagonist (HFB200603, NCT05238883).
!

References

1. Nature Biotechnology volume 38, pages 1384 1386.
2. Nucleic Acids Res. 2024 Jan 5; 52(D1): D1305––D1314.
3. Nat Rev Drug Discov. 2023 Jun;22(6):496-520.
4. Genome Biol 20, 194 (2019).
5. bioRxiv 2021.04.05; doi: https://doi.org/10.1101/2021.04.05.438318.