Unlocking Biomarker DiscoveryLarge Scale Application of Aptamer Proteomic Technology for Early Detection of Lung Cancer

Rachel Ostroff, William L. Bigbee, Wilbur A. Franklin, Larry Gold, Mike Mehan, York E. Miller, Harvey I. Pass, William N. Rom, Jill M. Siegfried, Alexandre F.R. Stewart, Jeffrey J. Walker, Joel L. Weissfeld, Stephen A. Williams, Dom Zichi, Edward N. Brody•View original

OverviewBalancedhelen voice

If lung cancer is caught early, surgery cures it. The five-year survival rate for patients diagnosed at stage one and operated on reaches eighty-six percent. But if it is caught late — which is most of the time — that number collapses to single digits. The difference between those two outcomes is almost entirely a matter of when the blood is drawn, when the image is taken, and when someone thinks to look. The brutal fact, as Ostroff and colleagues document, is that eighty-four percent of lung cancers are diagnosed at an advanced stage. About one and a half million people were diagnosed in two thousand eight. About one point three million died. That ratio has barely moved since nineteen sixty. The cure exists. The detection does not. Imaging has been the main attempt to fix this. Low-dose computed tomography scanning can find small tumors before they spread, and randomized trials have been testing it in high-risk smokers. But computed tomography has a serious signal-to-noise problem. In screening cohorts, roughly forty percent of long-term smokers show suspicious nodules on computed tomography. About ninety-seven percent of those nodules turn out to be benign. That means enormous numbers of people are funneled into follow-up procedures, added radiation exposure, anxiety, and cost — for very few true cancers. Given those limitations, researchers have been hunting for something that could work alongside or instead of imaging: a molecular signal in blood that reveals lung cancer before symptoms appear. Proteins are the natural candidates. They are the actual machinery of disease — secreted by tumors, shed into circulation, and measurable in a blood draw. The problem is that serum is an extraordinarily complex mixture, with proteins spanning concentrations that vary by a factor of ten billion from the most abundant to the rarest. Previous platforms — mass spectrometry, antibody arrays, and autoantibody arrays — could only cover a fraction of that range reliably. Ostroff and colleagues set out to change that by using a new aptamer-based proteomic technology to measure eight hundred thirteen proteins simultaneously in a single serum sample and to ask whether those eight hundred thirteen signals contain enough information to detect non-small cell lung cancer, or NSCLC, from a blood draw. Aptamers are short DNA molecules that fold into three-dimensional shapes and bind specific target proteins with high affinity — essentially synthetic antibodies, but manufactured from nucleic acids. The key capability here is multiplexing: because each aptamer is identified by its DNA sequence rather than by a label or an antibody, you can run hundreds of them simultaneously. The platform the team used measures each protein by capturing it with its cognate aptamer and then reading out the quantity of that aptamer on a hybridization array. The analytical performance is striking. The average limit of detection is one picomolar, with some targets detectable down to one hundred femtomolar. The dynamic range spans seven orders of magnitude — meaning the assay can simultaneously quantify proteins whose concentrations differ by ten million fold. Reproducibility was high, with a median coefficient of variation of five percent. Ostroff and colleagues argue, and the numbers support the argument, that this breadth and precision exceed what mass spectrometry and antibody-based platforms have achieved for broad serum profiling. The study was designed to put that capability through a serious clinical test. The team assembled one thousand three hundred twenty-six serum samples from four independent biorepositories: New York University, Roswell Park Cancer Institute, the University of Pittsburgh, and a commercial biorepository. Cases were two hundred ninety-one patients with biopsy-proven NSCLC whose sera were collected within eight weeks of diagnosis and before any tumor removal. Controls were one thousand thirty-five asymptomatic participants with at least ten pack-years of cigarette smoking — a deliberately difficult comparison group, because a test that can only distinguish cancer from healthy non-smokers isn't clinically useful. All one thousand three hundred twenty-six samples were processed blinded, randomly distributed across ninety-six well plates, and analyzed in a continuous run over eight days. In total, the study produced the equivalent of just over one million high-quality protein measurements. Finding the signal in that much data required a disciplined analytic pipeline. The team first screened all eight hundred thirteen proteins to remove any showing unexpected variation against internal controls — a step designed to catch preanalytical noise before it contaminates the discovery. They then ran six parallel Naive Bayes classifier discovery analyses across different control subgroups, which broadened the search while guarding against site-specific artifacts. Proteins were scored by Kolmogorov-Smirnov distance — the largest absolute gap between the cumulative frequency curves of cases and controls at any concentration — and by how often they appeared in greedy forward-search classifiers. This process produced forty-four candidate biomarkers from the original eight hundred thirteen. From those forty-four, the team built classifiers by adding one marker at a time and cross-validating at each step. Performance plateaued at twelve markers. The final twelve-protein panel — and it's worth reading these aloud: Cadherin-1, CD30 Ligand, Endostatin, HSP90 alpha, LRIG3, MIP-4, Pleiotrophin, PRKCI, RGM-C, SCF soluble receptor, soluble L-Selectin, and YES — was fed into a Naive Bayes classifier that combines each protein's individual signal into a single diagnostic call. Cross-validated training performance showed ninety-one percent sensitivity and eighty-four percent specificity, with an area under the ROC curve of zero point ninety-one. Sensitivity means the fraction of real cancer cases the test correctly flags. Specificity means the fraction of non-cancer controls the test correctly clears. Crucially, the classifier performed similarly for early and late stage disease — stage one sensitivity in training was ninety percent. That last point matters enormously. A screening test that only catches advanced cancer isn't screening — it's confirming what imaging already found. The real credibility test was the blinded verification set. Before any analysis began, the team randomly set aside twenty-five percent of all samples — three hundred forty-one subjects, seventy-eight cases, and two hundred sixty-three controls — and locked their identities behind a key held by a third-party statistician unaffiliated with any study site. When the classifier trained on the remaining data was applied to that hidden set, it returned eighty-nine percent sensitivity and eighty-three percent specificity. Almost identical to training. That's the pattern you want: a small, predictable drop, not a collapse. Stage one sensitivity in verification held at eighty-seven percent. Equally important was what happened across sites. The four collection centers varied in sample size and case distribution — Roswell Park had seventy-two cases, New York University had eighty-eight, Pittsburgh had eighty-eight, and a smaller site had forty-three. The classifier performed consistently across all four. This matters because site-to-site variation in how blood is collected, processed, and stored is one of the most reliable ways that biomarker studies fail to replicate. The team tested this directly, comparing protein-level differences between sites against the NSCLC versus control differences. The proteins most affected by handling — including complement C3 variants and coagulation factor IXab — were distinct from the NSCLC biomarkers. That separation is not guaranteed; it's a result. And it's why the biomarker selection pipeline explicitly removed analytes showing unexpected site-driven variation from the start. The twelve proteins that survived that process tell a biologically coherent story. Five are down-regulated in cancer cases — Cadherin-1, LRIG3, RGM-C, SCF soluble receptor, and soluble L-Selectin — and seven are up-regulated: CD30 Ligand, Endostatin, HSP90 alpha, MIP-4, Pleiotrophin, PRKCI, and YES. Together they implicate cell adhesion, angiogenesis, immune signaling, and oncogenic kinase activity — the machinery of tumor growth, invasion, and host response. Endostatin is an angiogenesis inhibitor. HSP90 alpha is a chaperone that stabilizes oncoproteins. Pleiotrophin drives cell proliferation and vessel formation. PRKCI and YES are kinases linked to oncogenic signaling. The panel isn't a random collection of elevated proteins; it reflects multiple aspects of what NSCLC does to the body and how the body responds. The authors are direct about the study's limitations. These are archived samples from tobacco-exposed populations, not a prospective screening trial. The controls are heavy smokers, not the general population. The study doesn't demonstrate organ specificity for most markers, and it doesn't include samples collected before clinical detection — so it can't yet answer whether the signal appears months or years before diagnosis. Independent validation in prospective cohorts remains the necessary next step. What this study establishes, with one thousand three hundred twenty-six subjects across four sites and a blinded verification set, is that eight hundred thirteen-protein aptamer profiling can identify a twelve-marker blood signature that detects NSCLC with high sensitivity at early stage, that the signal is reproducible, and that it outperforms published protein and gene expression panels. Ostroff and colleagues state they have initiated clinical validation studies aimed at a blood-based test for earlier lung cancer diagnosis. The bottleneck has always been detection. This is a serious attempt to break it. This lecture was created by ennepō. Go to https://ennepo.ai to Discover, Create and Follow the latest research in your field. Read when you can. Listen when you want to.

Given those limitations, researchers have been hunting for something that could work alongside or instead of imaging: a molecular signal in blood that reveals lung cancer before symptoms appear. Proteins are the natural candidates. They are the actual machinery of disease — secreted by tumors, shed into circulation, and measurable in a blood draw. The problem is that serum is an extraordinarily complex mixture, with proteins spanning concentrations that vary by a factor of ten billion from the most abundant to the rarest. Previous platforms — mass spectrometry, antibody arrays, and autoantibody arrays — could only cover a fraction of that range reliably. Ostroff and colleagues set out to change that by using a new aptamer-based proteomic technology to measure eight hundred thirteen proteins simultaneously in a single serum sample and to ask whether those eight hundred thirteen signals contain enough information to detect non-small cell lung cancer, or NSCLC, from a blood draw.

Aptamers are short DNA molecules that fold into three-dimensional shapes and bind specific target proteins with high affinity — essentially synthetic antibodies, but manufactured from nucleic acids. The key capability here is multiplexing: because each aptamer is identified by its DNA sequence rather than by a label or an antibody, you can run hundreds of them simultaneously. The platform the team used measures each protein by capturing it with its cognate aptamer and then reading out the quantity of that aptamer on a hybridization array. The analytical performance is striking. The average limit of detection is one picomolar, with some targets detectable down to one hundred femtomolar. The dynamic range spans seven orders of magnitude — meaning the assay can simultaneously quantify proteins whose concentrations differ by ten million fold. Reproducibility was high, with a median coefficient of variation of five percent. Ostroff and colleagues argue, and the numbers support the argument, that this breadth and precision exceed what mass spectrometry and antibody-based platforms have achieved for broad serum profiling.

The study was designed to put that capability through a serious clinical test. The team assembled one thousand three hundred twenty-six serum samples from four independent biorepositories: New York University, Roswell Park Cancer Institute, the University of Pittsburgh, and a commercial biorepository. Cases were two hundred ninety-one patients with biopsy-proven NSCLC whose sera were collected within eight weeks of diagnosis and before any tumor removal. Controls were one thousand thirty-five asymptomatic participants with at least ten pack-years of cigarette smoking — a deliberately difficult comparison group, because a test that can only distinguish cancer from healthy non-smokers isn't clinically useful. All one thousand three hundred twenty-six samples were processed blinded, randomly distributed across ninety-six well plates, and analyzed in a continuous run over eight days. In total, the study produced the equivalent of just over one million high-quality protein measurements. Finding the signal in that much data required a disciplined analytic pipeline. The team first screened all eight hundred thirteen proteins to remove any showing unexpected variation against internal controls — a step designed to catch preanalytical noise before it contaminates the discovery. They then ran six parallel Naive Bayes classifier discovery analyses across different control subgroups, which broadened the search while guarding against site-specific artifacts.

Proteins were scored by Kolmogorov-Smirnov distance — the largest absolute gap between the cumulative frequency curves of cases and controls at any concentration — and by how often they appeared in greedy forward-search classifiers. This process produced forty-four candidate biomarkers from the original eight hundred thirteen. From those forty-four, the team built classifiers by adding one marker at a time and cross-validating at each step. Performance plateaued at twelve markers. The final twelve-protein panel — and it's worth reading these aloud: Cadherin-1, CD30 Ligand, Endostatin, HSP90 alpha, LRIG3, MIP-4, Pleiotrophin, PRKCI, RGM-C, SCF soluble receptor, soluble L-Selectin, and YES — was fed into a Naive Bayes classifier that combines each protein's individual signal into a single diagnostic call. Cross-validated training performance showed ninety-one percent sensitivity and eighty-four percent specificity, with an area under the ROC curve of zero point ninety-one. Sensitivity means the fraction of real cancer cases the test correctly flags. Specificity means the fraction of non-cancer controls the test correctly clears. Crucially, the classifier performed similarly for early and late stage disease — stage one sensitivity in training was ninety percent. That last point matters enormously. A screening test that only catches advanced cancer isn't screening — it's confirming what imaging already found.

The real credibility test was the blinded verification set. Before any analysis began, the team randomly set aside twenty-five percent of all samples — three hundred forty-one subjects, seventy-eight cases, and two hundred sixty-three controls — and locked their identities behind a key held by a third-party statistician unaffiliated with any study site. When the classifier trained on the remaining data was applied to that hidden set, it returned eighty-nine percent sensitivity and eighty-three percent specificity. Almost identical to training. That's the pattern you want: a small, predictable drop, not a collapse. Stage one sensitivity in verification held at eighty-seven percent. Equally important was what happened across sites. The four collection centers varied in sample size and case distribution — Roswell Park had seventy-two cases, New York University had eighty-eight, Pittsburgh had eighty-eight, and a smaller site had forty-three. The classifier performed consistently across all four. This matters because site-to-site variation in how blood is collected, processed, and stored is one of the most reliable ways that biomarker studies fail to replicate. The team tested this directly, comparing protein-level differences between sites against the NSCLC versus control differences. The proteins most affected by handling — including complement C3 variants and coagulation factor IXab — were distinct from the NSCLC biomarkers.

That separation is not guaranteed; it's a result. And it's why the biomarker selection pipeline explicitly removed analytes showing unexpected site-driven variation from the start. The twelve proteins that survived that process tell a biologically coherent story. Five are down-regulated in cancer cases — Cadherin-1, LRIG3, RGM-C, SCF soluble receptor, and soluble L-Selectin — and seven are up-regulated: CD30 Ligand, Endostatin, HSP90 alpha, MIP-4, Pleiotrophin, PRKCI, and YES. Together they implicate cell adhesion, angiogenesis, immune signaling, and oncogenic kinase activity — the machinery of tumor growth, invasion, and host response. Endostatin is an angiogenesis inhibitor. HSP90 alpha is a chaperone that stabilizes oncoproteins. Pleiotrophin drives cell proliferation and vessel formation. PRKCI and YES are kinases linked to oncogenic signaling. The panel isn't a random collection of elevated proteins; it reflects multiple aspects of what NSCLC does to the body and how the body responds. The authors are direct about the study's limitations. These are archived samples from tobacco-exposed populations, not a prospective screening trial. The controls are heavy smokers, not the general population.

The study doesn't demonstrate organ specificity for most markers, and it doesn't include samples collected before clinical detection — so it can't yet answer whether the signal appears months or years before diagnosis. Independent validation in prospective cohorts remains the necessary next step. What this study establishes, with one thousand three hundred twenty-six subjects across four sites and a blinded verification set, is that eight hundred thirteen-protein aptamer profiling can identify a twelve-marker blood signature that detects NSCLC with high sensitivity at early stage, that the signal is reproducible, and that it outperforms published protein and gene expression panels. Ostroff and colleagues state they have initiated clinical validation studies aimed at a blood-based test for earlier lung cancer diagnosis. The bottleneck has always been detection. This is a serious attempt to break it. This lecture was created by ennepō. Go to https://ennepo.ai to Discover, Create and Follow the latest research in your field. Read when you can. Listen when you want to.

Related lectures