The complexity and variability of biological data has promoted the increased use of machine learning methods to understand processes and predict outcomes. These same features complicate reliable, reproducible, interpretable, and responsible use of such methods, resulting in questionable relevance of the derived outcomes. Here we systematically explore challenges associated with applying machine learning to predict and understand biological processes using a well characterized in vitro experimental system. We evaluated factors that vary while applying machine learning classifers: 1) type of biochemical signature (transcripts vs. proteins), data curation methods (pre- and post-processing), and 3) choice of machine learning classifier. Using accuracy, generalizability, interpretability, and reproducibility as metrics, we found that the above factors significantly modulate outcomes even within a simple model system. Our results caution against the unregulated use of machine learning methods in the biological sciences, and strongly advocate the need for data standards and validation tool-kits for such studies.

Universal and early recognition of pathogens occurs through recognition of evolutionarily conserved pathogen associated molecular patterns (PAMPs) by innate immune receptors and the consequent secretion of cytokines and chemokines. The intrinsic complexity of innate immune signaling and associated signal transduction challenges our ability to obtain physiologically relevant, reproducible and accurate data from experimental systems. One of the reasons for the discrepancy in observed data is the choice of measurement strategy. Immune signaling is regulated by the interplay between pathogen-derived molecules with host cells resulting in cellular expression changes. However, these cellular processes are often studied by the independent assessment of either the transcriptome or the proteome. Correlation between transcription and protein analysis is lacking in a variety of studies. In order to methodically evaluate the correlation between transcription and protein expression profiles associated with innate immune signaling, we measured cytokine and chemokine levels following exposure of human cells to the PAMP lipopolysaccharide (LPS) from the Gram-negative pathogen Pseudomonas aeruginosa. Expression of 84 messenger RNA (mRNA) transcripts and 69 proteins, including 35 overlapping targets, were measured in human lung epithelial cells. We evaluated 50 biological replicates to determine reproducibility of outcomes. Following pairwise normalization, 16 mRNA transcripts and 6 proteins were significantly upregulated following LPS exposure, while only five (CCL2, CSF3, CXCL5, CXCL8/IL8, and IL6) were upregulated in both transcriptomic and proteomic analysis. This lack of correlation between transcription and protein expression data may contribute to the discrepancy in the immune profiles reported in various studies. The use of multiomic assessments to achieve a systems-level understanding of immune signaling processes can result in the identification of host biomarker profiles for a variety of infectious diseases and facilitate countermeasure design and development.

Optical whispering gallery modes (WGMs) arise from total internal reflection of light from an optical cavity, such as microspheres. Whispering gallery mode phenomena produce resonant frequencies at the reflected wavelengths that appear as peaks in spectrographic data. In microspheres, whispering gallery mode spectra vary by the size and material used. These differences allow for distinct labeling of many targets using only a single fluorescent input and output. In this work we use resonant microspheres that, conjugated to antibodies, label bacterial and mammalian cell populations in vitro.

Early detection of pathogens using nucleic acids in clinical samples often requires sensitivity at the single-copy level, which currently necessitates time-consuming and expensive nucleic acid amplification. Here, we describe 1) a redesigned flow cell in the shape of a trapezoid-subtracted geometric stadium, and 2) modified experimental procedures that allow for the measurement of sub-attomolar analytes in microliter quantities on a fluorescence-based waveguide biosensor. We verified our instrumental sensitivity with a 200-μL sample of a fluorescent streptavidin conjugate at 100 zM (100 zeptomolar, or 100·10−21 mol L−1) and theoretically explored the applicability of this modified sensing platform in a sandwich immunoassay format using a Langmuir adsorption model. We present assays that demonstrate specific detection of synthetic influenza A DNA (in buffer) and RNA (in saliva) oligonucleotides at the single-copy level (200 μL at 10 zM) using a fluorescent molecular beacon. Lastly, we demonstrate detection of isolated genomic influenza A RNA at a clinically relevant concentration. This work constitutes a sensitivity improvement of over twelve orders of magnitude compared to our previous nucleic acid detection work, illustrating the significant enhancements that can be gained with optimized experimental design.

Mycobacterium ulcerans is the causative agent of the chronic and debilitating neglected tropical disease Buruli ulcer (BU) which mostly affects children. The early detection and treatment of M. ulcerans infections can significantly minimize life-long disability resulting from surgical intervention. However, the disease is characterized by relatively few systemic systems as a result of complex host-pathogen interactions that have yet to be fully characterized, which has limited the development of both diagnostic and therapeutic approaches to treat BU. In this work, we study the interactions of the host immune system with two principle M. ulcerans virulence factors: mycolactone, an amphiphilic macrolide toxin, and lipoarabinomannan (LAM), a cell wall component of most mycobacterial pathogens. We observe that human lipoproteins have a profound effect on the interaction of both mycolactone and LAM with the immune system. Individually, both molecules are pro-inflammatory in the absence of serum and immunosuppressive in the presence of serum. When combined, mycolactone and LAM are immunosuppressive regardless of serum conditions. We also show that Toll-like receptor 2 (TLR2), a macrophage pathogen pattern recognition receptor, is critical for LAM immune stimulation but aids in mycolactone immunosuppression. These findings are a first step towards unraveling mycolactone-mediated immunosuppression during BU disease and may facilitate the development of effective diagnostics and therapeutics in the future. Author Summary Buruli ulcer (BU) is a neglected tropical disease caused by the pathogen Mycobacterium ulcerans . The principal virulence factors associated with it are the macrolide toxin mycolactone and the major cell wall component lipoarabinomannan (LAM). Here, we examine the impact of the amphiphilic biochemistry of mycolactone and LAM on their interaction with the human immune system. We show that both mycolactone and LAM associate with serum lipoproteins, and that this association is critical for the immune evasion seen in early-stage M. ulcerans infections. In the absence of serum, mycolactone is pro-inflammatory. Immunosuppression occurs only in the presence of human serum lipoproteins. In the presence of LAM, mycolactone is immunosuppressive, regardless of serum conditions. Immunosuppression is a hallmark of BU disease, and understanding the mechanisms of this immunosuppression can support the development of effective diagnostic and therapeutic strategies.

Lipopolysaccharide (LPS) is the main molecular cause of sepsis, a severe bacterial infection responsible for killing as many as 11 million people per year. Upon bacterial infection, LPS is released by Gram-negative bacteria into the host bloodstream, where it is bound by LPS-binding protein (LBP). When LBP presents LPS to immune cells, it can cause the inflammatory response that leads to sepsis. However, when LBP associated with the carrier molecule high density lipoprotein (HDL) binds to LPS, the bacterial endotoxin is carried to the liver, where it is detoxified.

Detection methods that do not require nucleic acid amplification are advantageous for viral diagnostics due to their rapid results. These platforms could provide information for both accurate diagnoses and pandemic surveillance. Influenza virus is prone to pandemic-inducing genetic mutations, so there is a need to apply these detection platforms to influenza diagnostics. Here, we analyzed the Fast Evaluation of Viral Emerging Risks (FEVER) pipeline on ultrasensitive detection platforms, including a waveguide-based optical biosensor and a flow cytometry bead-based assay. The pipeline was also evaluated in silico for sequence coverage in comparison to the U.S. Centers for Disease Control and Prevention's (CDC) influenza A and B diagnostic assays. The influenza FEVER probe design had a higher tolerance for mismatched bases than the CDC's probes, and the FEVER probes altogether had a higher detection rate for influenza isolate sequences from GenBank. When formatted for use as molecular beacons, the FEVER probes detected influenza RNA as low as 50 nM on the waveguide-based optical biosensor and 1 nM on the flow cytometer. In addition to molecular beacons, which have an inherently high background signal we also developed an exonuclease selection method that could detect 500 pM of RNA. The combination of high-coverage probes developed using the FEVER pipeline coupled with ultrasensitive optical biosensors is a promising approach for future influenza diagnostic and biosurveillance applications.