AI identifies potent inducers of breast cancer stem cell differentiation based on adversarial learning from gene expression data

Abstract

Cancer stem cells (CSCs) are a subpopulation of cancer cells within tumors that exhibit stem-like properties and represent a potentially effective therapeutic target toward long-term remission by means of differentiation induction. By leveraging an artificial intelligence approach solely based on transcriptomics data, this study scored a large library of small molecules based on their predicted ability to induce differentiation in stem-like cells. In particular, a deep neural network model was trained using publicly available single-cell RNA-Seq data obtained from untreated human-induced pluripotent stem cells at various differentiation stages and subsequently utilized to screen drug-induced gene expression profiles from the Library of Integrated Network-based Cellular Signatures (LINCS) database. The challenge of adapting such different data domains was tackled by devising an adversarial learning approach that was able to effectively identify and remove domain-specific bias during the training phase. Experimental validation in MDA-MB-231 and MCF7 cells demonstrated the efficacy of five out of six tested molecules among those scored highest by the model. In particular, the efficacy of triptolide, OTS-167, quinacrine, granisetron and A-443654 offer a potential avenue for targeted therapies against breast CSCs.

Keywords: artificial intelligence; breast cancer; cancer stem cells; domain adaptation; drug repurposing; transcriptomics.

Figure 1

Overview of the study. Single-cell GEPs of hiPSCs at various differentiation stages and drug-induced GEPs were fed to an adversarial learning model, which simultaneously learned differentiation features to be used in subsequent predictions (main task) and dataset-specific features to be avoided (adversarial task). The trained model was then used to score all the drug-induced profiles. A selection of six drugs among the top-scoring ones was experimentally validated.

Figure 2

Model development. (A) The DREDDA model architecture includes one encoder for each dataset and a shared decoder; the resulting profiles from the source domain are sent to the main task classifier (positively weighted in the overall loss function), while both source and target domain profiles are sent to the adversarial classifier (negatively weighted). (B) During training, the main task accuracy increases, while the adversarial task accuracy decreases. (C) Comparison between the embedding before (left) and after (right) domain adaptation shows that cells at the various differentiation stages tend to cluster together more, while LINCS drug-induced profiles tend to spread across the source domain. (D) The neighborhood of untreated cell profiles tends to be more enriched for LINCS profiles after domain adaptation (curve peaking to the right) as compared to before (curve peaking to the left).

Figure 3

Validation and characterization of the top hits. (A) The performance of DREDDA and the other five tested methods as measured by the MRR and the nDCG at four different thresholds (@50, 100, 150, 200) of the bottom 10 drugs with the lowest DECCODE scores (see Supplemental Methods for the detailed descriptions). Error bars displayed for DREDDA, DREDDA w/o DA and Random Prediction based on five independent runs. (B) Top (bottom) drugs as prioritized by DREDDA have low (high) DECCODE scores, which predict stemness features (C) Similar to (A) but using the top 10 drugs which resulted in the highest colony area and counts. (D) Drugs previously tested for inducing stemness tend to be ranked lower by DREDDA based on experimental evidence including stem cells’ colony count and size. (E) The classification diversity of the LINCS profiles by DREDDA and the five other comparing methods into the four states of hiPSCs (F, G) Summary of the positive and negative enrichments for pathways among the top levels of the ‘Biological Process’ and ‘Cellular Component’ Gene Ontology categories that are significantly dysregulated by the top 30 drugs. (H) Same analysis as in (F, G), but focused on the ‘Cell cycle’ and ‘Differentiation’ levels in the ‘Biological Process’ category.

Figure 4

Mammosphere assay in drug-treated BC cells. (A) Representative images of MCF7 and MDA-MB-231 cells pre-treated for 24 h with increasing doses of the indicated molecules and then cultured for 7 days as mammospheres in stem cell medium after the washing out of the drug. (B) Average number of mammospheres, their diameter and self-renewal capacity in three independent experiments. For MDA-MB-231 treated with triptolide and OTS-167, the number of single cells composing the mammospheres, as a measure of their growth, is reported instead of the mammosphere diameter. (C, D) Representative images of MCF7 and MDA-MB-231 cells pre-treated for 24 h with increasing doses of each molecule and then cultured for 7 days as mammospheres in stem cell medium after the washing out of the drug. (D) For granisetron, only the lower dose (300 μM) and its relative control (CTRL1) were shown. (E) Average number of mammospheres, their diameter and self-renewal capacity in three independent experiments. CTRL = DMSO 0.1%; CTRL1 = DMSO 0.6%; ^*, P < 0.05; ^**, P < 0.01; ^***, P < 0.001; ^****, P < 0.0001; ns, not significant.

Figure 5

FACS profiling of CD44 and CD24 expression in MDA-MB-231 and MCF7 cells treated with quinacrine, triptolide, OTS-167, granisetron and A-443654. (A) Representative dot plots for quinacrine-treated cells. (B) The mean values +/− SE of the CD44+/CD24- (blue bars), CD44+/CD24+ (green bars) and CD44-/CD24+ (orange bars) subpopulations were reported as a ratio relative to control (CTRL: DMSO 0.1%) for all treatments. (C) Representative dot plots of granisetron-treated MCF7 and A-443654-treated MDA-MB-231 cells. GNS, granisetron; A-44, A-443654. (D) The mean values +/− SE of the CD44+/CD24- (blue bars), CD44+/CD24+ (green bars) and CD44-/CD24+ (orange bars) subpopulations were reported as a ratio relative to control (CTRL: DMSO 0.1% for A-443654; DMSO 0.6% for granisetron) for all treatments. ^*, P < 0.05; ^**, P < 0.01; ^****, P < 0.0001.