The North Anatolian Fault Zone is a natural laboratory for this kind of work. It slices across northern Turkey for roughly 1,200 kilometers, accommodating the westward extrusion of the Anatolian plate as it grinds between the Eurasian and African-Arabian plate systems. The fault has a well-documented history of sequential ruptures, with large earthquakes migrating along its length over the past century, including the devastating 1999 Izmit and Duzce events. For the new study, the fault was divided into nine predefined segments, each treated as an analytical unit with its own seismic history and hazard profile. This fault-specific approach departs from broad regional hazard maps, which average risk over large areas and can obscure the concentrated danger posed by a particular locked segment.

At the heart of the framework is a dataset of 539 seismic events drawn from Turkish earthquake catalogs, including records maintained by the Turkish Accelerometric Database and Analysis System operated by AFAD, the country’s disaster and emergency management authority. From this catalog, Ozkan engineered a set of predictive features designed to capture what he describes as short-to-medium-term seismic memory indicators. These are statistical fingerprints of recent earthquake activity on a segment, such as patterns in the timing and magnitude of preceding events, that encode how the fault remembers its own recent behavior. The idea builds on decades of seismological thinking, from the classic Gutenberg-Richter frequency-magnitude relationship to the seismic gap hypothesis of McCann and colleagues in 1979, which holds that segments that have not ruptured recently accumulate the greatest pending hazard.

The machine learning core of the study is the Extreme Gradient Boosting algorithm, widely known as XGBoost. This ensemble technique builds a hierarchy of decision trees in sequence, with each new tree trained to correct the residual errors of its predecessors. Gradient boosting has become a workhorse of tabular data science because it handles heterogeneous features well, resists overfitting through regularization, and often outperforms deep neural networks on structured datasets of modest size. In head-to-head comparison within this study, XGBoost demonstrated superior predictive capability over alternative approaches, ultimately achieving a final test Area Under the Receiver Operating Characteristic Curve score of 0.8930. An AUC of that magnitude indicates that the model ranks a randomly chosen future major earthquake above a randomly chosen quiet period nearly nine times out of ten, a level of discrimination that is considered strong for a problem as notoriously difficult as earthquake forecasting.

Perhaps the most consequential part of the work is not the classifier itself but the calibration layer wrapped around it. Raw machine learning outputs are probabilities in name only; without calibration, a stated 40 percent chance may not correspond to any meaningful real-world frequency. Ozkan performed a high-resolution probability calibration and imposed what he calls a strict Zero Miss operational strategy. The logic is blunt: for disaster management, failing to warn of a major earthquake is far more costly than issuing occasional false alarms. The optimization therefore searched for the decision boundary that would capture every major event in the test set, even at the price of reduced precision. That optimal boundary was formally established at a 35 percent probability threshold.

The results at that threshold are striking. When the calibrated framework was evaluated, it successfully captured 100 percent of the major earthquakes in the test data, meaning no event of magnitude 5.6 or greater slipped through unflagged. At the same time, the model achieved a balanced accuracy of 87.3 percent, a metric that averages the detection rate for hazardous events and the correct rejection rate for quiet periods, thereby preventing a large class of negatives from inflating the score. In practical terms, the framework behaved like a smoke detector tuned to be extremely sensitive: it rarely stays silent when danger is present, while still avoiding the constant crying wolf that would erode institutional trust in any operational warning system.

With the classifier calibrated, the final step was deployment. Ozkan applied the framework to evaluate future hazard probabilities across each of the nine predefined fault segments for their upcoming statistical cycles. The output is a quantitative, segment-by-segment ranking of hazard potential, expressed as probabilities rather than vague qualitative categories. This kind of dynamic, data-driven assessment could give disaster management authorities a defensible basis for deciding where to position search-and-rescue capacity, pre-stage emergency supplies, prioritize retrofitting programs, and focus public preparedness campaigns. Static hazard maps, updated on decadal timescales, cannot offer that level of temporal responsiveness.

The study sits within a rapidly expanding field. Machine learning approaches to seismology have progressed from early artificial neural networks trained on magnitude time series and seismic electric signals, through radon monitoring models, to modern deep learning architectures such as LSTM networks applied to seismicity analysis and satellite data streams. Recent work has extended these methods to ground motion modeling for probabilistic seismic hazard analysis in subduction zones and to real-time prediction of spectral acceleration for earthquake early warning. What distinguishes the new framework is its explicit marriage of classification performance with operational threshold calibration, an element that much of the academic literature leaves as an afterthought but that determines whether a model can actually be used in an emergency management context.

Important caveats remain. The framework was developed and tested on a single fault system, and its 539-event dataset, while substantial for a fault-specific analysis, reflects the catalog completeness and magnitude homogeneity of the Turkish record. Earthquake prediction in the strict sense, meaning the deterministic specification of the time, place, and magnitude of a future event, remains beyond any existing method, and this study does not claim otherwise. What the framework offers is a probabilistic hazard potential, a statement about relative likelihood over statistical cycles, which is a fundamentally more modest and more defensible objective. The author also reports that no external datasets were generated or analyzed beyond the study itself, and the work was conducted without dedicated funding, underscoring the individual, exploratory nature of the research.

Even with those qualifications, the implications are considerable. Turkey’s population is concentrated along corridors that the North Anatolian Fault threads like a fuse, and the Marmara region in particular is regarded by many seismologists as overdue for a major rupture. A tool that can translate decades of cataloged seismicity into calibrated, segment-specific probabilities offers exactly the kind of quantitative guidance that logistical planners need, converting abstract seismological knowledge into actionable resource allocation. If the framework’s performance holds under independent validation and in other tectonic settings, threshold-calibrated machine learning could become a standard component of the seismic risk toolkit, sitting alongside probabilistic seismic hazard analysis and operational forecasting systems rather than replacing them. The study demonstrates that the bottleneck in earthquake forecasting may no longer be the raw predictive power of the algorithms, but the careful, operationally minded engineering that turns a good classifier into a trustworthy decision aid.

Subject of Research: Machine learning-based fault-specific seismic hazard assessment along the North Anatolian Fault Zone

Article Title: A threshold-calibrated machine learning framework for evaluating fault-specific seismic hazard potentials

Article References: Ozkan, E. (2026). A threshold-calibrated machine learning framework for evaluating fault-specific seismic hazard potentials. Earth Science Informatics, 19(11), Article 210. https://doi.org/10.1007/s12145-026-02265-w

Image Credits: AI Generated

DOI: 10.1007/s12145-026-02265-w

Keywords: seismic hazard, machine learning, North Anatolian Fault, XGBoost, earthquake forecasting, threshold calibration, probability calibration, fault segments, Turkey, disaster management, balanced accuracy, seismicity indicators

The CADD score works by training a machine-learning model to distinguish variants that have survived natural selection from those that have been purged from the human population over evolutionary time. Variants that look like the ones evolution tends to eliminate receive high scores, expressed on a PHRED-like scale where a score of 20 means the variant is predicted to be more deleterious than the top one percent of all possible mutations. In practice, laboratories often use a fixed cutoff, frequently around 20 or higher, to filter the flood of variants that emerge from whole-exome or whole-genome sequencing of a patient. Anything below the threshold is deprioritized; anything above it is flagged for closer scrutiny by a diagnostic team. The elegance of this approach is its simplicity. The problem, as the new paper makes clear, is that the simplicity is deceptive.

The distribution of CADD scores is not uniform across the genome. Different genes and different gene sets carry different baseline distributions of scores, meaning that a single global cutoff can behave very differently depending on the context. A threshold that works well for one disease panel may be too permissive for another, quietly letting potentially pathogenic variants slip through, or too strict, burying diagnosticians under false positives. Compounding the issue, scores can shift between genome builds, the two most common being GRCh37 and GRCh38, and between CADD releases, such as versions 1.6 and 1.7. A variant scored under one combination of build and release may land on a slightly different point of the scale when re-scored under another. For a field where a single misclassified variant can change a patient’s diagnosis, these are not trivial discrepancies.

CADD-ThresholdApp tackles the calibration problem by grounding threshold selection in real clinical evidence. The application draws on ClinVar, the public archive that collects and classifies variants according to their reported clinical significance, ranging from benign to pathogenic. By visualizing the CADD-score distributions of clinically classified variants, the app lets users see, for any gene or set of genes, how benign and pathogenic variants actually separate along the score axis. Rather than guessing where to place the cutoff, users can watch how the overlap between the two distributions changes as the threshold moves, and choose a boundary that reflects the genuine diagnostic trade-off between sensitivity and specificity in their specific context.

Beyond the visualizations, the app reports threshold-dependent performance metrics, quantifying how well a chosen cutoff performs at discriminating pathogenic from benign variants within the selected gene set. This turns what has often been an informal, intuition-driven decision into a measurable one. Users can define their own genes or gene lists, which is essential for laboratories working on rare diseases where only a handful of candidate genes matter. The app also integrates curated gene panels from PanelApp, a widely used resource of expert-reviewed gene panels for clinical genomics, allowing diagnosticians to calibrate thresholds directly against the panels they actually deploy in practice.

One of the most technically significant features of the application is its support for comparisons across genome releases and CADD versions. Because GRCh37 and GRCh38 differ in their coordinate systems and in some sequence content, and because successive CADD releases incorporate updated annotations and retrained models, the same variant can carry different scores in different pipelines. The app allows users to examine how these differences affect threshold behavior, which is particularly valuable for laboratories transitioning between genome builds or updating their scoring pipelines while maintaining consistency in variant interpretation. Being able to compare v1.6 and v1.7 side by side within the same interface means that a lab can assess the practical impact of an upgrade before committing to it.

The developers also emphasized extensibility. CADD-ThresholdApp is implemented in a modular framework, meaning that it can be adapted to incorporate additional variant resources beyond ClinVar and additional deleteriousness scores beyond CADD. This design choice matters because the variant-interpretation landscape is crowded with competing scores, including SpliceAI, REVEL, AlphaMissense, and many others, and laboratories increasingly combine multiple metrics. A tool that can grow to accommodate new evidence sources is far more durable than one hard-wired to a single score. The modular architecture also invites community contributions, allowing other groups to extend the app to their own scoring systems without rebuilding it from scratch.

Accessibility was clearly a priority. The application is available as a hosted web service at cadd-threshold.kircherlab.bihealth.org, requiring nothing more than a browser, which lowers the barrier for smaller diagnostic laboratories and research groups without dedicated bioinformatics support. For institutions with stricter data policies or custom requirements, the app can be installed locally via container images, from the Python Package Index through PyPI, through the Bioconda package manager, or directly from source code. This range of deployment options reflects a pragmatic understanding of how computational tools actually get adopted in clinical settings, where reproducibility and local control are often non-negotiable requirements.

The significance of this work extends beyond a single tool. Variant interpretation sits at the heart of modern genomic medicine, and the choice of scoring thresholds directly influences which patients receive diagnoses and which variants are dismissed as benign. As whole-genome sequencing becomes routine in healthcare systems worldwide, the volume of variants requiring classification is growing faster than the expert capacity to review them, making automated prioritization scores like CADD indispensable. But an indispensable tool calibrated carelessly is a liability. By providing a transparent, evidence-based framework for context-specific threshold calibration, CADD-ThresholdApp addresses a genuine blind spot in the diagnostic pipeline, one that has rarely been discussed openly because the field has defaulted to convenient universal cutoffs.

The study, published as open-access software research with no competing interests declared, represents a collaboration rooted in the lab that originally developed the CADD score itself, with Martin Kircher among the co-authors and acknowledgments extending to the broader CADD team and supporting bioinformatics units at the Berlin Institute of Health at Charité and the University of Washington. For a discipline built on the promise of precision, there is something fitting about a tool whose entire purpose is precision in calibration, ensuring that the numbers geneticists rely on mean the same thing, whether they are examining a cardiac gene panel in Berlin or a rare-disease exome in Tokyo. As genomic medicine scales up, tools like this one may prove to be the quiet infrastructure that keeps the entire diagnostic enterprise honest.

Subject of Research: Calibration of CADD variant deleteriousness score thresholds using ClinVar for clinical variant interpretation

Article Title: CADD-ThresholdApp: an interactive web application for calibrating CADD scores using ClinVar

Article References: Leifheit, C., Kircher, M., & Schubach, M. (2026). CADD-ThresholdApp: an interactive web application for calibrating CADD scores using ClinVar. BMC Bioinformatics, 27(1), Article 224. https://doi.org/10.1186/s12859-026-06675-x

Image Credits: AI Generated

DOI: 10.1186/s12859-026-06675-x

Keywords: CADD score, ClinVar, variant interpretation, genomics, bioinformatics, gene panels, threshold calibration, rare disease, web application, GRCh38, variant prioritization, machine learning