The collection of transcriptomic data is expensive, resulting in datasets with a small number of sample sizes (in the hundreds) but thousands of variables. As a result, several techniques that are making significant strides in the imaging space, such as deep neural networks, are not suitable for these datasets, as they require a large number of samples. Furthermore, the heterogeneity of the patient population and the complexity of diseases found in oncology requires going beyond the labels. The development of techniques that can explain the driving variables behind patient subpopulations is tremendously valuable in identifying and developing novel therapeutic agents—this is particularly relevant for mapping out heterogeneous diseases such as lung cancer.

Lung cancer is the leading cause of cancer mortality worldwide, with non-small cell lung cancer (NSCLC) accounting for 85% of all lung cancers [1]. NSCLC can be divided into three histological subtypes with distinct phenotypes and prognoses: adenocarcinoma (ADC), squamous cell carcinoma (SCC), and large cell carcinoma (LCC) [2, 3]. The histological differences across these subtypes suggest that distinct molecular mechanisms underlie the observed phenotypic differences. Although the differential gene expressions across NSCLC subtypes have been of increasing interest, the therapeutic implications on how these pathways interact are only more recently being investigated [4]. The remarkable degree of genetic variability within each histological subtype further highlights the importance of molecular biology and genotyping for NSCLC [5, 6].

Fortunately, machine learning (ML) advancements have served as promising tools for stratifying NSCLC, predicting transcriptional mutations based on histological slides, and discriminating NSCLC subtypes through genomic expression levels. The bulk of ML efforts has focused on image analysis for predicting the stage of NSCLC [7–10]. However, the growing body of evidence highlighting the molecular abnormalities that underlie the genomic subtypes of NSCLC can train ML algorithms to identify novel biomarkers for NSCLC, moving towards precision medicine [11–13]. For instance, previous reports have identified that ADC is associated with increased expression of genes related to protein transport and cell junctions, while SCC is associated with increased expression of genes related to cell division and DNA replication [14]. An analysis of gene expression profiles between ADC and SCC using ML has been previously reported, identifying several genes including cystatin-A (CSTA), tumor protein p63 (TP63), serpin family B member 13 (SERPINB13), chloride channel accessory 2 (CLCA2), bicaudal D cargo adaptor 2 (BICD2), P53 apoptosis effector related to PMP22 (PERP), FAT atypical cadherin 2 (FAT2), basonuclin 1 (BNC1), ATPase phospholipid transporting 11B (ATP11B), family with sequence similarity 83 member B (FAM83B), keratin 5 (KRT5), par-6 family cell polarity regulator gamma (PARD6G), and plakophilin 1 (PKP1) which were differentially expressed in ADC and SCC [15].

Other computational methods for discriminating genetic drivers of NSCLC have been previously investigated. A k-means clustering method was used to classify genetic subtypes of ADC [16]. Healthy and ADC tissue was then classified using a support vector machine followed by input into a self-organizing map neural network. The neurons in the output layer were categorized using a hierarchical clustering method to divide ADC tumours into four genetic subtypes. Two subtypes were found to have high expression levels of immune-related genes, suggesting the existence of heterogenous subpopulations of NSCLC. In another study, researchers also used hierarchal clustering of copy number variations to derive insights into NSCLC drug response [17].

Several ML frameworks have been previously developed specifically tailored for small datasets. A one-shot learning approach called CancerSiamese has been used to predict cancer types while highlighting several marker genes to predict metastatic or primary tumour signatures [18]. A second ML approach has shown promise in deriving insights into immune cell populations in a rare disease application [19].

In order to identify novel driving genes that distinguish these two broad subtypes, a combination of ML tools was designed to learn from patient datasets to analyze gene expression data derived from ADC and SCC NSCLC patients. Because large datasets are critical for most contemporary ML methods such as deep neural networks, there is a need for alternative techniques when data banks are insufficient to train the model. In addition, significant features found within small datasets may become diluted by more obvious statistical features and hence over-represented in large datasets. As such, ML methods must be carefully used and complemented by statistical methods that allow for the discovery of non-linear ways in which groups of genes may interact to drive disease heterogeneity. The methodology presented here is designed for small datasets—a novel way of hypothesizing genetic subpopulations that may result in pathogenesis. For example, the ML framework proposed here has been previously used on a small genetic dataset consisting of Alzheimer’s disease brain samples [20]. Several genetic pathways associated with Alzheimer’s disease were uncovered, suggesting that even with a small dataset, there exists a high degree of genetic complexity within pathophysiology. Similarly, the findings presented here support genes previously reported to distinguish ADC and SCC subtypes. However, the novelty of this work lies in the ability to discover previously unknown subpopulations that are defined by several genes at a time. These findings shed light on the different mechanisms at play within these subtypes.

Using publicly available NSCLC datasets with a suite of ML techniques appropriate for small datasets had an excellent signal for separating ADC and SCC. The main philosophy followed here is that for smaller datasets, where the patients are unlikely to reflect the distribution of patients in the totality of reality, one can allow ML methods to generate hypotheses about the population available in a small dataset. This allows a researcher to benefit from the power of statistics, in that they can test the hypothesis and derive some measure of confidence. Proprietary methods like the NetraAI empower this hypothesis testing paradigm, but the method described above is also capable of expressing hypotheses in the form of patient clusters.

Here, patient clusters were compared using statistical methods suitable for a dataset with so few samples in order to avoid overfitting that often comes with utilizing contemporary ML methods with small datasets. Further, the transparency of the driving factors is important so that human experience can be used to evaluate what is being ‘discovered’ by the machine.

This study highlights the genetic heterogeneity within NSCLC subtypes. Using this dataset, a set of 10 genes that distinguish ADC and SCC were identified (Table 2). Within these 10 genes, 9 have been previously reported to be associated either with NSCLC or a specific subtype of NSCLC, validating this ML approach. These findings were aligned with previous reports on SCC genes being associated with the organization and assembly of cell and gap junctions, glutathione conjugation and the redox stress response, ECM organization and collagen-related proteins, interferon and cytokine signaling, and HLA downregulation and ADC genes associated with ECM organization proteins and complement, interferon and cytokine signaling, and collagen-related genes and proteins for ECM organization [44]. Another study identified epidermis development, cell division, and epithelial cell differentiation as the most common categories characterizing SCC, and cell adhesion enrichment, biological adhesion, and coagulation for ADC [45]. However, some of the genes identified have not been previously associated with NSCLC or a specific subtype and represent areas that warrant greater investigation for the advancement of precision medicine in NSCLC.

The first of the previously reported NSCLC-associated genes identified was DSC3, which plays a role in epidermal morphology and keratinocyte proliferation [22]. There are several studies that report on DSC3 distinguishing ADC from SCC, with a higher expression in SCC [36, 46–48]. Notably, there has been a report on the association between DSC3 and tumour suppressor activity in NSCLC mediated by inhibition of EGFR [49]. However, there remain contradictory associations between DSC3 and prognosis, with elevated levels associated with increased metastatic risk in melanoma and better prognosis in lung and colon cancer [40]. This suggests that the same protein may have differential effects in the tumour microenvironment (TME), which presents an interesting field of research to understand how DSC3 expression correlates with NSCLC subtypes depending on where they originate in the lung. Reports of upregulation of desmosomes and gap junctions in SCC and tight junctions in ADC suggest that SCC and ADC are characterized by a distinct set of adhesion molecules [21].

In the results presented here, ADC has been reported to be characterized by tight junctions and was identified by CGN and SCC has been characterized by gap junctions and was identified by DSC3 (Figure 3). Males have been reported to have a significantly poorer NSCLC prognosis compared to females, shifting efforts towards sex-based approaches to diagnosis, prognosis, and therapeutic interventions [50, 51]. Additionally, estrogens have been associated with an increased risk of ADC in women despite equal expression of estrogen receptors α and β; however, the role remains unclear [52]. While there are several reports on the sex-based differences in cancer mechanisms, including differences in metabolism, immunity, and angiogenesis, differences in CGN and DSC3 expression have not been previously reported to the best of our knowledge [53]. Gap junction proteins, also known as connexins, serve as channels that connect the interior of adjacent cells, facilitating intracellular homeostasis and coordination of activities via second messengers [54]. Desmosomes primarily provide mechanical strength via a structural network. In contrast, tight junctions form a barrier around the cell, regulating the permeability of the paracellular space [55, 56]. These molecules play critical roles in epithelial-to-mesenchymal transition, a process involved in cancer metastasis. Aside from the current work relating CGN expression to females, no sex-based differences have been previously reported. This presents a unique field of research, as there may be different druggable targets for males and females. The variability of adhesion molecule expression across sex warrants further investigation to elucidate the details of the correlation and advance toward gender related precision medicine.

Interestingly, SLC6A10P was the single gene that was found to distinguish between two specific subpopulations of ADC. SLC6A10P was previously found to be a marker for aggressive ADC [43], and recently, implicated within the Notch signaling pathway [57]. These findings suggest that SLC6A10P warrants further investigation as a genetic biomarker in the context of the ADC patient subpopulation. This demonstrates the power of machine intelligence to reveal etiologies within complex diseases, even when a small number of samples are present. However, the methods must be used to reveal subpopulations that can then be compared using appropriate statistical methods suitable for comparing small groups.

With respect to the SCC patient population, CLCA2 and IRF6 were found to distinguish between two distinct SCC subpopulations. CLCA2 has been reported to be highly expressed in SCC, suggesting that it may serve as a diagnostic marker to differentiate SCC from ADC. Female patients with CLCA2-negative SCC exhibited significantly poorer prognoses [37]. Furthermore, SCC expression was correlated with tumour grade upon histological characterization. In particular, CLCA2-negative samples were associated with poorly differentiated tumours [37].

Most noteworthy, phosphatidylinositol glycan anchor biosynthesis class gene PIGX, was the only gene identified that has not been previously associated with NSCLC. However, there have been reports that PIGX promotes cancer cell proliferation by suppressing EHD2 and ZIC1 in breast cancer [58]. The authors reported that PIGX expression was associated with shorter recurrence-free survival. In the present study, PIGX was found to be a driver of ADC and SCC differentiation, being overexpressed in SCC patients (Figure 2). As a novel gene associated with NSCLC or a specific subtype, this highlights an area that warrants further investigation for the advancement of precision medicine in NSCLC.

In order to create robust predictive models with machine intelligence, large datasets are required, but this study utilized the ability for some of these methods to create hypotheses instead, and then use methods appropriate for small data to test these hypotheses. This method uncovered several genetic subtypes of ADC of SCC, including those driven by SLC6A10P, CLCA2, and IRF6, respectively. Furthermore, these data suggest that the expression levels of adhesion proteins encoded by CGN and DSC3 may play a role in sex-based differences in NSCLC. Finally, this study uncovered a statistically significant driver of NSCLC heterogeneity, PIGX, which warrants further investigation.

This report highlights the use of a novel set of ML techniques that are appropriate for small datasets. The primary aim of using such techniques is to encourage other researchers to explore small datasets that are often otherwise skipped with ML as there may be hidden valuable information within them. Adopting these approaches, one can extract meaningful insights with the techniques described here to move closer toward precision medicine.