Poster L28 - Integrating CRISPR and Machine Learning to Achieve Mature Chondrocyte Genetic Profiles in MSCs

Introduction: Traumatic and degenerative cartilage injuries are a leading cause of pain worldwide leading to large medical expenses and reduced quality of life. Mesenchymal stromal cells (MSCs) are a promising cell source for cartilage regeneration; however, clinical translation is challenged by limited chondrogenic capacity in vivo and phenotype instability. CRISPR gene editing adds fine control of cell fate through precise manipulation of gene expression but the downstream effects of CRISPR on broader gene regulatory networks (GRNs) are not well understood making it difficult to understand how a CRISPR therapy will affect cell phenotype and drive MSC chondrogenesis. To address this, deep learning allows for the interpretation of highly complex GRNs that dictate cell phenotype; however, no models exist to harness this predictive ability for selecting CRISPR target genes. Our main objective is to construct a synthetic musculoskeletal GRN using a deep neural network (DNN) framework and employ this model to iteratively evaluate potential CRISPR target genes for their capacity to induce a chondrogenic phenotype in MSCs. We hypothesize that insights from our synthetic GRN will elucidate critical relationships between MSCs and cartilage tissues to inspire machine learning informed CRISPR strategies for reliable control over cell fate and robust chondrogenic commitment in MSCs.

Materials and

Methods: Bulk RNAseq data describing musculoskeletal tissues was downloaded from the sequence read archive (SRA) using the SRA toolkit and aligned to the GRCh38.p14 genome assembly using the Kallisto algorithm. Differential expression analysis elucidated the top differentially expressed genes (DEGs) in either fetal cartilage or mature cartilage compared to MSCs. The 1000 significant (p < 0.01) DEGs with the highest fold change in fetal cartilage compared to MSCs were identified as gene candidates for CRISPRa induced MSC chondrogenesis. Alternatively, the top 5000 most significant DEGs, lowest adjusted p value, were identified as key regulatory genes that will be used to assess the effects of gene activation. Gene candidates and key regulatory genes were then set as input and output nodes respectively in a DNN connected with three hidden layers and two dropout layers. The DNN model was trained using keras built on top of tensorflow with an gaussian error linear unit (gelu) activation function, learning rate=0.0001, β1=0.9, β2=0.999, ε=1e-8, batch size = 50, validation split=0.15, mean squared error loss function, 200 training epochs and early stopping with a patience of 10. Following model training, CRISPRa-mimetic perturbations are made in MSC RNAseq datasets resulting in a unique dataset for each gene candidate. These datasets are then input into the trained DNN and model predictions are compared to mature cartilage RNAseq datasets to determine optimal CRISPRa target genes for MSC chondrogenesis.

Results, Conclusions, and Discussions: Differential expression analysis yielded DEGs that correspond to previous reports of MSC chondrogenesis with CNMD, UCMA, COL2A1, TGFB3, and SOX9 ranking among the most upregulated genes. These and remaining DEGs were then sorted to yield input and output genes for the DNN. DNN training resulted in a model that explained over 93% of the variance in output gene expression using the expression of input genes. CRISPRa minetic perturbations and subsequent assessment yielded ACKR3, BTG2, RPS27P19, PPP4R4, and SOX9 as the top performing genes. SOX9 is known as the master chondrogenic transcription factor, therefore validating model predictions and re-affirming the promise of activating this transcription factor(1). ACKR3, also known as CXCR7, has been reported to regulate the osteogenic and chondrogenic differentiation of stem cells and is known to regulate inflammation in cartilage tissues(2). BTG2 is involved in BMP/Smad signaling pathways and slows replication suggesting that it may be a powerful target specifically for MSC chondrogenesis as it must slow the proliferation of these cells while inducing chondrogenesis(3). PPP4R4 is predicted to be downstream of the EZH2 transcription factor which is well characterized for the methylation of anti-chondrogenic genes(4,5). Finally, RPS27P19 is a ribosomal pseudogene that is uncharacterized in cartilage tissues. In all, we present an in silico modeling approach for the prediction of optimal CRISPR target genes for driving MSC chondrogenesis. Top scoring genes show obvious correlations to known chondrogenic genes and we expect that experimental validation will confirm their affinity for driving chondrogenesis. Furthermore, adaptation of this strategy will allow CRISPR gene target prediction across a range of tissue and cell types for the design of a broad range of successful biomedical therapies. Future studies will focus on the in vitro and in vivo validation of model performance by activating the identified genes in MSCs and assessing transcriptomic profiles and cartilage extracellular matrix deposition. This novel approach allows for an in silico high throughput screening of CRISPR targets and the discovery of new therapeutic gene editing strategies for enhanced musculoskeletal repair and clinically relevant regenerative strategies.

Acknowledgements (Optional): 1. Liu, C.-F. & Lefebvre, V. The transcription factors SOX9 and SOX5/SOX6 cooperate genome-wide through super-enhancers to drive chondrogenesis. Nucleic Acids Res. 43, 8183–8203 (2015).
2. Wei, S.-T. et al. Gain of CXCR7 function with mesenchymal stem cell therapy ameliorates experimental arthritis via enhancing tissue regeneration and immunomodulation. Stem Cell Res. Ther. 12, 314 (2021).
3. Park, S. et al. B-Cell Translocation Gene 2 (Btg2) Regulates Vertebral Patterning by Modulating Bone Morphogenetic Protein/Smad Signaling. Mol. Cell. Biol. 24, 10256–10262 (2004).
4. Lui, J. C. et al. EZH1 and EZH2 promote skeletal growth by repressing inhibitors of chondrocyte proliferation and hypertrophy. Nat. Commun. 7, 13685 (2016).
5. Lachmann, A. et al. Massive mining of publicly available RNA-seq data from human and mouse. Nat. Commun. 9, 1366 (2018).