Enhancing Engineered Cell Therapies for CNS

1. Overview

Rationale: Cellular replacement efforts in the central nervous system (CNS) have been employed to treat neurodegenerative diseases, stroke, brain injuries, and other ailments. Preclinically and clinically, efforts to replace neurons or provide neurotrophic support have been hindered by limitations in the ability of the cells to differentiate, survive, integrate, and mature. Genetically engineering cells has been proposed – and utilized – to enhance the ability of cells to overcome these limitations. This project aims to narrow a dataset of genetic perturbations using publicly available datasets and machine learning techniques. This dataset (roughly 250 genes and/or pathways) will be useful in enabling future in vivo genetic screening efforts to enhance the engraftment efficacy of cell therapies for treating a variety of CNS conditions and insults.

Ethics, Timeline, and Data-Sharing: I estimate that this project will take approximately 5 months to complete, pending time to access datasets and obtain funding for computational credit purchasing. Primary focus during the time course will be on developing the analysis pipeline, cleaning data for interpretability, and refining outputs of the pipeline to be as relevant to future utilization in in vivo studies as possible. No human or animal subjects will be utilized in this trial and all data employed is publicly available or available by request; as such, no ethics approval will be needed.

Updates on the progress of the project, including updates to the code, changes to protocols or procedures, and results will be posted on a monthly basis. I affirm that all raw code scripts, requirements, data, and final outputs will be made publicly available for sharing with the scientific community.

2. Introduction

Background: Neurodegenerative diseases (NDDs), traumatic brain injuries (TBIs), stroke, spinal cord injuries (SCIs), and other CNS insults – including general brain aging – generate, among other effects, massive amounts of neuronal loss. With the brain’s limited ability to compensate for this loss, the death of these neurons can lead to irreversible impairments of basic functions, cognition, and personality. To reverse the resulting phenotypes of these insults, exogenous transplantation of cells to replace dead neurons and repair damaged neural circuits is a promising approach. However, many preclinical and clinical efforts have faced challenges with ensuring transplanted cells differentiate, survive, integrate, and mature to provide a direct functional restoration of lost ability^1–3.

Engineering of cells to enhance their functionality in overcoming these hurdles could potentially improve outcomes for patients. However, investigating the modulation of the whole genome is logistically infeasible¹. As such, computational analysis of existing genetic perturbation screens in neuronal cells under various stress conditions could reveal key transcriptomic patterns underlying resilience and thus inform subsets of perturbation candidates. A strong example of such data enabling this analysis is available from the lab of Dr. Martin Kampmann at UCSF in the CRISPRbrain platform^4–6. Additionally, advances have been made in in silico prediction of transcriptomic state changes to cells undergoing genetic perturbations. One such example of this is the Arc Institute’s STATE model⁷. Predicting transcriptomic states based on perturbations that possess beneficial effects in stress-induced environments, and isolating unique transcriptomic features and/or pathways using machine learning (ML) methods could provide insight into key genetic perturbations that may enhance cellular resilience to the types of stress intracranial transplantation subjects them to, thus improving outcomes.

Research Question: Can ML-based analysis of large-scale genetic perturbation datasets, including the utilization of silico cell state prediction models, be leveraged to identify key perturbations that enhance neuronal cell survival, integration, and resilience in stress conditions relevant to CNS transplantation?

H1: Perturbation-derived transcriptomic profiles (using a transformer such as STATE) will reveal common beneficial features and subsequent resilience associated pathways (e.g., stress response, synaptic plasticity, metabolic regulation) that are predictive of long-term resilience, functional integration, and maturation of transplanted neurons.

H2: Machine learning (LASSO-regularized logistic regression) models trained on pathway enrichment from perturbed cell state data can generalize to predict novel perturbations and combinations of perturbations to enhance cellular phenotypes, thus accelerating the discovery of edits that improve cell therapy outcomes in neurodegenerative diseases and brain injuries.

Impact Statement: If successful, the list of genes identified in this initial process will be brought into pending pooled screen experiments in vitro and in vivo. This will including translational efforts to identify/validate edits, create stable cell lines, and conduct the ability of such cells to enhance behavioral outcomes.

3. Methods

Participants/Sample

Data from the CRISPRbrain platform (Kampmann Lab, UCSF) will be manually sorted for inclusion based on the screen’s relevance to neurological conditions, disease state microenvironments, stress-response, or apoptosis. The final screen datasets included in the initial data are as follows; (+)/(-) indication determines whether the positive “phenoscores” (provided by the platform as a measure of barcode enrichment in a given cell state) will be taken. This is for the purpose of identifying perturbations that have a positive effect on cell outcomes, which does not always mean a “positive” phenoscore and must be manually annotated:

Glutamatergic Neuron-CellRox-CRISPRi (–)
Glutamatergic Neuron-Day-14-Survival-CRISPRi (+)
Glutamatergic Neuron-Day-21-Survival-CRISPRi (+)
Glutamatergic Neuron-Day-28-Survival-CRISPRi (+)
Glutamatergic Neuron-FeRhoNox-CRISPRi (–)
Glutamatergic Neuron-Liperfluo-CRISPRi (–)
Glutamatergic Neuron-No Antioxidants-Survival-CRISPRa (+)
Glutamatergic Neuron-No Antioxidants-Survival-CRISPRi (+)
Glutamatergic Neuron-PSAP KO-No Antioxidants-Survival-CRISPRa (+)
Glutamatergic Neuron-PSAP KO-No Antioxidants-Survival-CRISPRi (+)
Glutamatergic Neuron-PSAP KO-Survival-CRISPRa (+)
Glutamatergic Neuron-PSAP KO-Survival-CRISPRi (+)
Glutamatergic Neuron-Survival-CRISPRa (+)
Glutamatergic Neuron-Survival-CRISPRi (+)
iPSC-Survival-CRISPRi (+)
iTF-Microglia-Survival-Proliferation-CRISPRi (+)

Extracted data will include the perturbation names, the barcodes for the perturbation, the type of perturbation (inhibition or activation), the phenoscore, the significance (p-adj value), the cell type screened on, and the assay.

Power analysis: To ensure accurate power both for the ML regression and likelihood ratio test (LRT), we will want to produce approximately 200 single cell transcriptomic profiles (scTPROs) per perturbation, thus enabling informed binary and multi-group leiden clustering.

Materials and Procedures

CRISPRbrain (Kampmann Lab, via API): I will obtain perturbation data (as noted above) from survival and stress-response screens of iPSC-derived human glutamatergic neurons, iPSCs, and iTF-microglia. Metadata fields will include perturbation name, phenoscore, adjusted p-value, assay context, and screen type (CRISPRi/a).

Reference transcriptomes: baseline single-cell expression profiles from GEO accession GSE124703 (Tian et al., 2019), specifically GSM3543622 / SRR8400813 iNeuron CROP-seq data, used as a reference for STATE/scTPRO transformations.

Pretrained Transformer Models: Arc Institute’s STATE framework (ST-Tahoe pretrained checkpoint, or CZI’s scGenePT model, both available from HuggingFace implementation), and cell embeddings to project perturbation-induced transcriptomic profiles.

Compute Resources: analyses implemented in Python/R; packages include Scanpy, Seurat (normalization, clustering, integration), Harmony (batch correction), sklearn (ML), and statsmodels/R glm (statistical testing). Analyses will be run on local and cloud GPUs (Google Colab T4 instances or AWS instances (pending funding) for STATE training/inference).

Planned Analyses

Analyses:

Screen selection and API extraction – screens will manually be annotated and ported into a matrix. All code will be prepared to extract information about the screens and to filter the desired perturbations (“positive” phenotype, p-adj ≤ 0.05). Control cells from the baseline transcriptome will also cleaned and loaded.

scTPRO transformation – using STATE and the pretrained transformer (ST-Tahoe checkpoint), control cell single cell expression matrices will be referenced to model the output scTPROs following the list of perturbations extracted from step 1. Resulting files will be annotated with metadata from the CRISPRbrain assay(s) that the perturbation was extracted from.

Integration – normalization (dropping zero-var genes) and concatenation of the expression profiles will occur (retaining annotations). Principal Component Analysis (PCA) will be run using Scanpy on the concatenated scTPROs and Harmony will be used for correction.

Clustering – unsupervised multi-group leiden clustering will be conducted to gather information (Seurat). Additionally binary clustering will be conducted based on Shannon’s entropy to maximize the diversity of assays represented in the scTPROs as well as the overall number of cells. The “best” cluster (highest entropy and number of cells) will be saved while all other cells are concatenated and included in an “other” cluster.

Pathway enrichment analysis – Gene Set Enrichment Analysis (GSEA) will be conducted on both clusters, and pathway enrichment rankings are saved.

ML feature extraction – a LASSO-normalized logistic regression model will be trained to identify gene pathway enrichment patterns that determine membership in the “best” binary cluster. A likelihood ratio test (LRT) will be conducted as an in silico mutagenesis screen to rank the pathway enrichment patterns that are identified.

Gene Ontology and Selection – top 250 genes in the enriched pathways (identified by the ML model and ranked in the LRT) will be extracted and will undergo Gene Ontology (GO) Analysis. Genes with relevant GO terms (relevance to neurological conditions, disease state microenvironments, stress-response, or apoptosis, neuronal development, axonal guidance, synaptic integration, and others) will be extracted, and metadata (including how the gene was perturbed in the origin screen) will be furthered.

Hypothesis-Specific Contingents:

H1: (1) IF Adjusted Rand Index (ARI) is <0.75 (indicated less meaningful/stable clusters) THEN increase n scTPROs produced from perturbations and increase number of control cells included in scTPRO production. Re-run PCA with tuned params. (2) IF pathway discovery returns limited hits (less than 5 significant pathways per assay at FDR ≤ 0.10), THEN switch to rank-based enrichment (AUCell/GSVA).

H2: (1) IF Events per Variable (EPV) is <20 (EPV = n_scTPROs/x_ML-features), indicating a potential overfitting of the model THEN reduce the number of features, collapse number of assays by combining or removing, reduce gene set by restricting only to top % of high variance genes.

Ethics and Data Management

No human or animal subjects will be utilized in this trial and all data employed is publicly available or available by request; as such, no ethics approval will be needed.

All data is deidentified and derived purely from in vitro experimental assays with no patient information included in any of the collection or processing. Data and scripts will be made fully available.

4. Pilot Data (Optional)

Working with Dr. Sarah Rozelle at UMass Lowell (my current institution), I have begin laying out some o the pipelines for the transformations. As for the machine learning and feature extraction, prior works have utilized LASSO-regularized regression models with a LRT for ranking to assess defining features of clusters before. A notable example of this method is from the CellCartographer pipeline developed in the lab of Dr. George Church to identify driver transcription factors that influence cell differentiation⁸.

5. Budget

Item	Units / Qty	Unit Cost (USD)	Subtotal (USD)	Notes
AWS p4d (8×A100) – model training/fine-tuning	80 GPU-hours	32	2560	STATE runs; LASSO regression; LRT runs
GPU for inference & clustering (A10G/T4 class)	300 GPU-hours	1	300	STATE infer; PCA; Harmony; Leiden; AUCell/GSEA
Google Colab Pro+ (dev notebooks)	6 months	50	300	Iterative development; QC; quick visualizations
AWS EBS block storage (analysis volumes)	1 TB × 3 months	80	240	Solid workspace for working fast with remodeling matrices and datasets
AWS S3 object storage (archives)	1 TB × 3 months	23	69	Durable storage for raw pulls; model training checkpoints; outputs
Data transfer / egress	750 GB	0.1	75	S3 ↔ local snapshots and figure builds
External NVMe SSD (4 TB)	1 unit	230	230	Local cache and backup for scTPRO, transformer model training data, and outputs
Visualization & tooling (figures, reports)	1 package	200	200	Diagrams, cluster maps, and figure export utilities, biorender
ML experiment tracking (e.g., W&B/Neptune)	~3 months	180	180	Run tracking; reproducibility; config audit trail
Contingency (overages)			346	Covers price changes and extra GPU hours
TOTAL			4500