The research, published on 16 October 2026, draws on publicly available data from the Single-cell Pediatric Cancer Atlas, or ScPCA, a resource that catalogs single-cell profiles from pediatric cancer patients. Kim reprocessed two public leukemia projects from the atlas using a common cell-type mapping strategy, ensuring that cells from different studies could be compared on the same biological vocabulary. From this harmonized foundation, a cross-project immune composition was constructed from three major compartments: T and natural killer cells, B-lineage cells, and myeloid cells. The goal was to see whether these proportions, tracked over time within individual patients, could reveal something about the trajectory of the disease that simpler categorical labels would miss.

Central to the study is the concept of ecological context. In cancer biology, the tumor microenvironment is often described in terms of discrete states, such as immune-hot versus immune-cold, or dominant cell type versus suppressed lineage. These coarse labels are convenient, but they risk discarding information. Kim’s framework treats immune composition as a continuous variable embedded in a stochastic longitudinal model, specifically an Ornstein-Uhlenbeck process, which is a mathematical model that describes a value fluctuating around a long-term mean with a tendency to revert toward that mean. In this context, the OU process models the leukemia-state trajectory over time, and the question is whether immune composition shifts the attractor, the value mu toward which the trajectory is pulled.

Among 104 compositionally eligible participants, coarse clustering with two groups was reproducible under bootstrap resampling, meaning that the split into two clusters was stable when the data were resampled repeatedly. However, this clustering showed only limited association with the project from which the data came or with the specific diagnosis. Within the 99 participants with B-cell acute lymphoblastic leukemia, the most common pediatric leukemia, the two clusters separated 73 participants with T and natural killer cell dominant profiles from 26 participants with B and myeloid shifted profiles. This separation was largely independent of which project the data originated from, suggesting a genuine biological structure rather than a batch effect.

Yet a critical complication emerged. When Kim compared cluster assignments derived from raw composition with those derived from centered log-ratio transformed composition, a standard transformation for compositional data that accounts for the constraint that proportions must sum to one, the concordance was poor. This means that the apparent ecological state a patient belongs to depends heavily on how the compositional data are represented mathematically. A patient might appear in one cluster under raw proportions and in a different cluster under log-ratio transformed proportions. This representation sensitivity undermines the reliability of coarse context labels as stable descriptors of biological reality.

Because of this instability, Kim chose to evaluate continuous ecological composition separately from coarse context labels in a corrected longitudinal analysis. The longitudinal dataset included 34 participants contributing 241 observations and 207 transitions between time points. The key finding was that higher B-lineage composition was associated with a lower effective OU attractor, the parameter mu, in both models restricted to B-lineage cells alone and models that included B plus myeloid compartments. In biological terms, this suggests that as B-lineage cells become more prevalent in the immune environment, the leukemia state trajectory is pulled toward a lower value, potentially reflecting a less aggressive or more controlled disease state.

In contrast, the coarse context effects, the discrete labels that would traditionally be used to categorize patients, were unresolved in the longitudinal models. They did not provide clear, consistent information about the direction or magnitude of the leukemia trajectory. This is a striking result because it suggests that the very labels clinicians and researchers often rely on to summarize immune context may be less informative than the underlying continuous composition they are meant to represent.

To assess whether the continuous composition genuinely adds predictive value, Kim employed leave-one-participant-out validation, a rigorous approach in which each participant is held out in turn and the model is trained on the remaining participants. In aggregate, the model including B-lineage composition modestly outperformed the OU-only model that lacked ecological covariates. However, the predictive differences were heterogeneous across participants, meaning that for some individuals the continuous composition helped substantially while for others it did not. Models that included only coarse context labels, or that combined ecology with context, provided no incremental advantage over the simpler approaches. This pattern reinforces the conclusion that the value lies in the continuous signal, not in the categorical summaries.

A secondary analysis using an ordinal-state model, which treats the leukemia state as an ordered categorical variable rather than a continuous one, revealed a directionally consistent B-associated signal. This consistency across different modeling frameworks strengthens confidence that the B-lineage association is not an artifact of a single analytical choice. Simulations using the observed irregular sampling schedules from the real data demonstrated that moderate ecological effects could be recovered even when observations were not evenly spaced in time, addressing a common concern in longitudinal single-cell studies. Posterior-predictive analyses further showed that including continuous ecological composition improved the calibration of uncertainty estimates, meaning that the model’s confidence intervals better reflected actual variability in the data.

The broader implications of this work extend beyond pediatric leukemia. The study demonstrates a general principle: stable clustering of compositional data does not necessarily imply that those clusters represent longitudinally predictive ecological states. Researchers across cancer biology, immunology, and microbial ecology often face the temptation to reduce continuous compositional information into discrete categories for ease of interpretation. Kim’s framework offers a rigorous way to test whether such reduction loses information, and it provides a template for distinguishing descriptive ecological coarse-graining from ecological covariates that genuinely add explanatory power to stochastic longitudinal models. Importantly, the author cautions that the OU mean reversion observed in the models should not be interpreted as direct evidence of stabilizing selection or of fixed leukemia ecotypes, a nuance that guards against overinterpretation of the mathematical framework. As single-cell technologies continue to generate rich longitudinal datasets, approaches like this one will be essential for extracting the full biological signal from the immune landscapes that surround cancer.

Subject of Research: Continuous immune composition as a longitudinal predictor in pediatric leukemia single-cell data

Article Title: Continuous immune composition retains longitudinal information beyond coarse ecological-context labels in pediatric leukemia

Article References: Kim, S.-H. (2026). Continuous immune composition retains longitudinal information beyond coarse ecological-context labels in pediatric leukemia. BMC Bioinformatics. https://doi.org/10.1186/s12859-026-06679-7

Image Credits: AI Generated

DOI: 10.1186/s12859-026-06679-7

Keywords: pediatric leukemia, single-cell RNA-seq, immune composition, Ornstein-Uhlenbeck process, Bayesian inference, tumor microenvironment, B-cell acute lymphoblastic leukemia, longitudinal modeling, compositional data, ScPCA, ecological modeling, BMC Bioinformatics

Feature selection is the computational step in which an algorithm sifts through potentially thousands of measured variables—gene expression levels, protein abundances, clinical measurements—and decides which subset carries the genuine predictive signal. It is essential for biomarker discovery and for building clinical predictive models, because models trained on too many noisy variables tend to memorize the quirks of their training data rather than learn generalizable patterns. Over the past decades, the methodological literature on feature selection has become rich and mature, offering a wide spectrum of algorithmic approaches, from filter methods that rank variables by statistical association, to embedded methods like LASSO that perform selection during model training, to wrapper methods that repeatedly retrain models on different variable subsets. But the ROOFS team observed a persistent gap: much of this methodological progress has not been fully adopted in applied biomedical research. Practitioners often pick a single familiar method, run it once, and report the top-ranked features, without asking whether those features would be stable, robust, or even real.

ROOFS tackles this problem with a benchmarking philosophy. Rather than prescribing one best method, the Python package takes the user’s own dataset and runs multiple feature selection methods on it in parallel, currently supporting classification tasks. It then generates reports summarizing a comprehensive set of evaluation metrics that together paint a far more honest picture of what each method has found. The central idea is that the choice of a feature selection method should be data-driven, adapted to the specific problem at hand, rather than inherited from habit or convenience. By automating this comparative evaluation, ROOFS lowers the technical barrier that has kept rigorous methodological benchmarking out of reach for many applied research groups.

Among the metrics ROOFS evaluates, downstream predictive performance is arguably the most important. A feature selection method is only useful if the features it selects actually improve the ability of a predictive model to classify new samples correctly. But estimating this performance naively is a well-known statistical trap. If the same data are used both to select features and to evaluate the resulting model, the reported accuracy is inflated by a statistical bias called optimism. ROOFS addresses this by estimating downstream predictive performance using optimism correction, a technique that quantifies how much the apparent performance overstates the true generalization ability and adjusts the estimate accordingly. This correction is what separates a defensible biomarker model from one that merely flatters its own training data.

Stability is the second pillar of the ROOFS evaluation. A genuine biomarker should appear consistently in the selected feature set regardless of small perturbations in the data—whether a few patients are added or removed, or the analysis is rerun with a different random seed. Methods that select wildly different features on nearly identical datasets are producing artifacts, not biology. ROOFS quantifies stability by repeatedly applying each feature selection method across resampled versions of the data and measuring how consistent the selected features remain. This matters enormously for translational value: a biomarker panel that cannot be reproduced across cohorts has little hope of surviving validation, let alone reaching clinical use.

Robustness of individual features adds a third dimension. Beyond asking whether the overall feature set is stable, ROOFS examines how reliable each individual feature is—how consistently a particular gene or protein earns its place in the selection across repeated analyses. This distinction is technically meaningful because a method can produce a stable overall panel while individual members fluctuate, or vice versa. For researchers deciding which candidates to carry forward into expensive experimental validation, knowing which individual features are robust and which are marginal can save years of wasted laboratory effort and millions in research funding.

Perhaps the most technically inventive component of ROOFS is its use of semi-synthetic data with a simulated outcome to estimate true positive and false positive rates. In real biomarker studies, researchers never know the ground truth—which features are genuinely predictive and which are noise—so they cannot directly measure how often a method correctly identifies real signals or falsely promotes random variables. ROOFS circumvents this by constructing semi-synthetic datasets in which the outcome is simulated and the truly predictive features are known by construction. Running each feature selection method on these controlled datasets reveals its true positive rate—how reliably it detects genuine signals—and its false positive rate—how often it promotes pure noise to the status of biomarker. This gives researchers a calibrated understanding of each method’s detection characteristics, something no single real-world analysis can provide.

The ROOFS team demonstrated the package’s utility on data from the PIONeeR biomarker study, a clinical research program aimed at identifying predictors of resistance to anti-PD-(L)1 immunotherapy in lung cancer. Immunotherapy with PD-1 and PD-L1 inhibitors has transformed the treatment of non-small cell lung cancer, but a substantial proportion of patients derive little benefit, and identifying who will resist treatment remains one of the most pressing unsolved problems in oncology. The PIONeR study, sponsored by Assistance Publique Hôpitaux de Marseille and conducted in accordance with Good Clinical Practice guidelines with approval from the French ethics committee, generated exactly the kind of high-dimensional clinical and molecular data where feature selection choices carry maximum consequence. Applying ROOFS to this data allowed the researchers to benchmark competing feature selection methods and evaluate which candidates for resistance biomarkers could withstand scrutiny across stability, robustness, and predictive performance metrics.

The implications reach well beyond lung cancer immunotherapy. The authors argue that comprehensive benchmarking with ROOFS has the potential to improve the reproducibility of feature selection discoveries and increase the translational value of clinical models. In an era when machine learning is being deployed across every domain of medicine—from radiology to genomics to drug response prediction—the reproducibility crisis in biomarker research represents both a scientific failure and a barrier to patient benefit. Tools that institutionalize rigorous evaluation at the point of discovery could reshape how applied biomedical research is conducted, making it standard practice to ask not just which features a method selects, but whether those features deserve to be trusted. The package is written in Python, supports classification tasks, and is freely available on the GitLab repository of Inria, the French national research institute for digital sciences, lowering the entry cost for research groups worldwide.

ROOFS arrives at a moment of reckoning for computational biomedicine. The methodological literature on feature selection is mature and abundant; what has been missing is the bridge that carries that maturity into everyday applied research. By automating multi-method benchmarking, correcting for statistical optimism, quantifying stability and robustness, and calibrating detection performance on semi-synthetic ground truth, ROOFS offers researchers a way to make feature selection choices that are defensible, transparent, and reproducible. For a field where fragile biomarker claims have too often collapsed under validation, that bridge could not come soon enough. Whether it succeeds in changing practice will depend on adoption—but the tool itself, open access and openly licensed, is now available for any research team willing to put its biomarker discoveries to the test.

Subject of Research: Biomarker discovery

Article Title: ROOFS: RObust biOmarker Feature Selection

Article References: Bakhmach, A., Dufossé, P., Charpigny, S., Monville, F., Greillier, L., Barlési, F., & Benzekry, S. (2026). ROOFS: RObust biOmarker Feature Selection. BMC Bioinformatics. https://doi.org/10.1186/s12859-026-06667-x

Image Credits: AI Generated

DOI: 10.1186/s12859-026-06667-x

Keywords: feature selection, biomarker discovery, machine learning, reproducibility, lung cancer, immunotherapy, BMC Bioinformatics, Python, ROOFS, RObust, biOmarker, Feature

Drug-target interaction prediction is a cornerstone of computational pharmacology. When researchers screen a new molecule, they need to know which proteins it might latch onto, because those binding events define both the intended benefit and the potential side effects of a drug. Experimentally testing every possible pairing is prohibitively slow, so computational models that can reliably rank which interactions are likely have become essential tools for shortening development cycles and cutting costs. Deep learning has made remarkable progress in this arena, but the authors of the new study argue that existing approaches leave significant information on the table.

The central problem the team identified is that drugs and proteins can be represented in fundamentally different ways. On one side sits sequence data: the linear string of amino acids in a protein, or the chemical notation describing a drug molecule. On the other side sits graph structure data, which encodes how atoms in a molecule are connected and how residues in a protein relate to one another spatially. Each representation captures different aspects of molecular identity, and most prior models struggle to fully exploit the complementary information between these modalities when handling multimodal data, according to the study.

DHF-DTI tackles this challenge with a purpose-built component called the Cross-Attention and Bilinear Fusion Module, abbreviated CABFusion. Cross-attention is a mechanism borrowed from modern transformer architectures that allows one data stream to selectively focus on the most informative parts of another. In practice, this means the model can align features from a drug’s chemical graph with the relevant segments of a protein’s sequence, achieving fine-grained fusion of local features while simultaneously capturing global contextual dependencies that span the entire molecule. The bilinear component of the module then multiplies feature representations together, letting the network model rich, higher-order interactions between the two modalities rather than simply concatenating them.

The second innovation addresses a subtler but equally important weakness in existing models. Most drug-target predictors are trained purely with classification loss, which teaches the network to output a binding probability but says little about how the learned internal representations are organized. The Foshan team found that relying solely on classification loss makes it difficult to learn highly discriminative feature representations in the latent space, which severely limits the robustness and domain generalization ability of the model. In other words, a model might perform well on familiar data but falter when asked to evaluate a drug or protein it has never seen before.

To fix this, the researchers introduced supervised contrastive learning into the training pipeline. This technique reshapes the geometry of the model’s internal feature space by pulling together representations of interacting drug-protein pairs while pushing apart non-interacting ones. The result is a latent space in which true interactions form tight, well-separated clusters, improving both inter-class separability and generalization. This matters enormously in real-world drug discovery, where the molecules of greatest interest are often precisely the ones least represented in existing databases.

The team evaluated DHF-DTI across three increasingly demanding testing regimes: intra-domain, cross-domain, and cold-start scenarios. Intra-domain testing measures performance when training and test data come from the same distribution, the easiest setting. Cross-domain evaluation probes whether the model transfers knowledge between different datasets or biological contexts. Cold-start scenarios are the harshest test of all, asking the model to make predictions for drugs or targets that were entirely absent from its training data. Across all three regimes, the study reports that DHF-DTI significantly outperformed state-of-the-art methods in both prediction performance and generalization, a combination that suggests the dual-modality fusion strategy is capturing genuinely useful molecular information rather than exploiting dataset artifacts.

Beyond raw benchmark numbers, the authors conducted case studies, drug feature visualization, and error analysis to validate the model’s application potential in practical drug discovery. Feature visualization allows researchers to inspect how the network organizes drug representations in its latent space, offering a window into whether the learned structure aligns with known pharmacological relationships. Error analysis, meanwhile, examines the cases where the model fails, providing insight into the remaining gaps between computational prediction and biological reality. This kind of transparency is increasingly seen as essential for models intended to guide laboratory experiments rather than merely top leaderboards.

The implications for the pharmaceutical industry could be substantial. A model that generalizes well to unseen drugs and targets can help prioritize which compounds advance to expensive wet-lab validation, potentially saving years of trial-and-error screening. The work also reflects a broader trend in computational biology: the recognition that no single data representation is sufficient, and that hybrid architectures capable of fusing sequence, structure, and other modalities are becoming the new standard for molecular machine learning. As multimodal approaches mature, they are expected to underpin more efficient and accurate drug discovery pipelines from target identification through lead optimization.

The research, led by Dongjie Chen with corresponding authors Zhihui Yang and Mingming Zhao, was supported by funding from the Basic and Applied Basic Research program of Guangdong Province, the Medical Science and Technology Research program of Guangdong Province, the Science and Technology Innovation Project of Foshan City, the National Natural Science Foundation of China, and the Guangdong Provincial Overseas Distinguished Scholar Program. The full study is available as an open-access article in BMC Bioinformatics, allowing researchers worldwide to examine, reproduce, and build upon the framework as the field moves toward more robust and generalizable models of molecular interaction.

Subject of Research: DHF-DTI: a dual-modality hybrid fusion framework for drug-target interaction prediction

Article Title: DHF-DTI: a dual-modality hybrid fusion framework for drug-target interaction prediction

Article References: Chen, D., Liu, Y., Yi, Y., Liu, Z., Zhang, R., Liu, D., Yang, Z., & Zhao, M. (2026). DHF-DTI: a dual-modality hybrid fusion framework for drug-target interaction prediction. BMC Bioinformatics. https://doi.org/10.1186/s12859-026-06673-z

Image Credits: AI Generated

DOI: 10.1186/s12859-026-06673-z

Keywords: drug-target interaction, deep learning, multimodal fusion, supervised contrastive learning, drug discovery, cross-attention, BMC Bioinformatics, machine learning, computational pharmacology, cold-start prediction, DHF-DTI, dual-modality

The method, called FLARE, short for Fourier–Laplace Analytical Risk Estimation, was introduced by Fahrettin Kaya of Sütçü İmam University in Kahramanmaraş, Türkiye, in an open access paper published in BMC Bioinformatics on 13 October 2026. FLARE combines two well-established mathematical techniques, Orthogonal Random Fourier Features and the Laplace approximation, into a unified framework for nonlinear clinical risk modeling with analytical uncertainty quantification. The result is a model that can match the flexibility of nonlinear machine learning while preserving the inferential transparency that has made logistic regression a staple of clinical research for decades.

The first pillar of FLARE is the use of Orthogonal Random Fourier Features, a technique that approximates kernel functions by mapping input data into a randomized feature space. This mapping allows the model to capture nonlinear relationships between clinical variables and outcomes without the computational burden of full kernel methods. Because the approximation is explicit and parametric, the framework retains access to the machinery of classical statistics, something that genuine kernel methods and tree ensembles generally forfeit. The orthogonality constraint on the random features improves the quality of the approximation and stabilizes the resulting estimates.

The second pillar is the Laplace approximation, a classical method for approximating posterior distributions around their mode. In FLARE, this approximation is used to estimate uncertainty arising from three distinct sources: the model parameters themselves, the kernel hyperparameters that govern the shape of the nonlinear function space, and the interaction between these two layers of uncertainty. This decomposition matters because clinical predictions are rarely endangered by a single source of doubt; parameter estimates, model choices, and their combined effects all contribute to how confident a physician should be in a given risk score.

What sets FLARE apart from standard Bayesian neural networks or Gaussian process classifiers is the breadth of analytical outputs it produces from a single fitted model. The framework delivers closed-form prediction uncertainty at the level of the individual patient, meaning that each risk estimate arrives with a mathematically derived confidence interval rather than a heuristic or resampling-based approximation. Through Jacobian-based projection, FLARE also provides Wald tests, confidence intervals, and odds ratios for model terms, restoring the familiar inferential vocabulary of clinical epidemiology to a nonlinear setting.

Interpretability receives equally systematic treatment. FLARE decomposes variable effects into linear and nonlinear components using path-dependent gradient analysis, allowing researchers to see precisely how much of a variable’s influence on risk flows through a straightforward linear channel and how much arises from curved, threshold-like, or otherwise nonlinear behavior. In addition, the framework performs analytical Hessian-based interaction analysis at the patient level, quantifying how pairs of covariates jointly influence risk for a specific individual rather than only for the average patient across a population.

The empirical evaluation covered four real-world clinical datasets and two synthetic datasets, providing both controlled and authentic tests of the method. Under data-generating conditions that were genuinely linear, FLARE produced results functionally equivalent to logistic regression, an important sanity check demonstrating that the added machinery does not distort inference when simplicity suffices. In nonlinear settings, FLARE yielded improved predictive performance, confirming that the random Fourier feature representation captures structure that a purely linear model would miss. Analytical confidence intervals achieved coverage close to nominal levels, meaning that the stated uncertainty bands contained the true values at approximately the promised rate, and attribution patterns derived from FLARE showed strong agreement with SHAP-based explanations, the de facto standard in explainable machine learning.

Perhaps the most striking finding emerged from the FLChain cohort, a real-world dataset of patients with monoclonal gammopathy. At the population level, certain interactions between covariates were statistically undetectable, the kind of result that would ordinarily lead analysts to conclude the interactions were negligible. Yet when FLARE’s patient-specific interaction analysis was applied, those same interactions were found to be substantially amplified in specific high-risk patients, with certain covariate pairs exhibiting markedly stronger local effects than their population-average counterparts suggested. This heterogeneity has direct clinical implications: a risk factor pairing that appears innocuous on average may be dangerous for a particular subgroup of patients, and only individual-level analytical tools can surface that danger.

The significance of this work lies in its refusal to accept the conventional separation between statistical inference and machine learning flexibility. In clinical practice, a prediction without an honest uncertainty estimate is difficult to act on, and a nonlinear model without interpretable components is difficult to trust or regulate. By providing closed-form uncertainty, parametric tests, linear–nonlinear decomposition, and patient-level interaction assessment within one analytical framework, FLARE offers a template for how predictive models in medicine might satisfy both the performance demands of modern machine learning and the accountability demands of clinical decision-making. The method’s reliance on publicly available, de-identified datasets and its open access publication should lower barriers to independent evaluation and adoption.

Challenges remain before such frameworks become routine in hospitals. The approach, like all Laplace-based methods, rests on a local approximation whose quality depends on the shape of the underlying posterior, and the random feature approximation introduces its own modeling choices. Nevertheless, the published results suggest that the era in which clinicians must choose between interpretability and nonlinear power may be drawing to a close. As uncertainty quantification and explainability move from optional extras to regulatory expectations in medical artificial intelligence, analytical frameworks like FLARE point toward models that can explain not only what they predict, but how confident they are, which variables drove the prediction, and where hidden interactions may lurk for the individual patient in front of the physician.

Subject of Research: Analytical uncertainty quantification and interpretable nonlinear modeling for clinical risk prediction

Article Title: FLARE: Analytical uncertainty quantification and linear-nonlinear decomposition for interpretable clinical risk prediction

Article References: Kaya, F. (2026). FLARE: Analytical uncertainty quantification and linear-nonlinear decomposition for interpretable clinical risk prediction. BMC Bioinformatics. https://doi.org/10.1186/s12859-026-06692-w

Image Credits: AI Generated

DOI: 10.1186/s12859-026-06692-w

Keywords: clinical risk prediction, uncertainty quantification, Laplace approximation, random Fourier features, interpretable machine learning, explainable AI, logistic regression, interaction analysis, SHAP, BMC Bioinformatics, biostatistics, FLChain cohort

The central problem the researchers identify is what they call phenotypic inflation: the systematic gap between how well a biomarker performs in a discovery dataset and how much genuine clinical predictive value it actually carries. When the team plotted phenotypic performance against clinical performance across their cohorts, they found a striking and reproducible pattern. Every 0.1-unit increase in the phenotypic area under the curve, a standard measure of classification accuracy, predicted a 0.079-unit larger drop in clinical performance, a relationship that explained roughly 37 percent of the variance and held with overwhelming statistical significance. Most remarkably, above a phenotypic AUC of approximately 0.55, a threshold that 95 percent of observations in their dataset exceeded, additional phenotypic performance conferred no additional clinical predictive value whatsoever.

This finding strikes at the heart of conventional biomarker discovery practice. If most published immune biomarkers already sit above the 0.55 phenotypic AUC threshold, then the endless pursuit of ever-higher classification accuracy within discovery datasets is, according to this analysis, largely a waste of effort. The uncoupled model, which removed the systematic relationship, explained only about 4 percent of variance, confirming that the inflation phenomenon is not statistical noise but a structured, quantitatively reproducible feature of bulk transcriptomic biomarker evaluation. The relationship replicated independently across ten external RNA-sequencing cohorts and held when the analysis was stratified by cancer type, where it explained more than half of the observed variance.

To understand why this inflation occurs, the authors turned to the biology of the tumor microenvironment. Bulk RNA sequencing measures the average gene expression across a mixture of tumor cells, immune cells, fibroblasts, and other stromal components. A gene that appears to be a strong marker of antitumor immunity in bulk data may in fact be expressed by a completely different cell type than assumed, or its apparent signal may be driven entirely by shifts in the proportions of cell populations rather than by genuine changes in per-cell expression. Either scenario can produce a biomarker that looks excellent in one cohort but collapses in another where the cellular composition differs.

The team’s first diagnostic instrument addresses this problem directly. The Source Fidelity Index quantifies how faithfully a gene’s expression can be attributed to T cells and natural killer cells, drawing on single-cell RNA-sequencing data from 2.2 million individual cells across 18 datasets. The results were sobering. Of 101 immune genes commonly invoked in biomarker studies, only 35, or 34.7 percent, showed high T/NK fidelity with an SFI of 0.6 or greater. In other words, roughly two-thirds of the genes that researchers routinely treat as proxies for cytotoxic immune activity cannot be reliably attributed to the immune cells they are assumed to represent. Well-known markers such as granzyme A, perforin 1, granulysin, and NKG7 fared better than many others, but the overall picture suggests that a substantial portion of the immune biomarker literature rests on shaky cellular attribution.

The second metric, the Biomarker Fragility Index, tackles the complementary problem of composition sensitivity. Using three-dimensional simulation across 82 synthetic cellular compositions for each gene, the framework measures how much a gene’s bulk expression signal changes when the mixture of cell types shifts. A fragile biomarker is one whose bulk-level behavior is dominated by composition rather than by biology, making it inherently unreliable across patient cohorts with different immune infiltration patterns. When the researchers tested whether fragility carried information beyond what spatial methods could capture, they found that BFI provided 18.6 to 27.3 percent independent explanatory variance for clinical outcomes even after controlling for a spatially derived source-fidelity estimate, with significance in four of five phenotype definitions. Notably, BFI also correlated independently with spatial myeloid co-localization, suggesting that the local positioning of myeloid cells relative to other populations contributes to how fragile a biomarker’s bulk signal becomes.

Combining these metrics with additional risk factors, the authors assembled a six-factor Operational Risk Stratification framework and validated it in 754 samples from eight immune checkpoint blockade cohorts. The composition-boundary rule, a simple decision criterion derived from the fragility analysis, achieved 82.8 percent accuracy in classifying whether a biomarker’s phenotypic behavior was reliable. The full six-factor framework recalled 93.3 percent of phenotypic failures, and this performance remained stable under leave-one-cohort-out validation, with mean recall of 95.0 percent across held-out cohorts. Perhaps most practically, BFI-guided selection reduced the risk of phenotypic validation failure by 82 percent compared with random biomarker selection, a difference that reached permutation-based significance at p equals 0.007. For research groups deciding which candidate genes to carry forward into costly validation programs, this represents a potentially enormous saving of time and resources.

The framework’s robustness was tested against a common criticism of such approaches: sensitivity to the statistical assumptions embedded in the analysis. The authors report that BFI’s classification performance was invariant across all tested distributional assumptions, meaning the tool does not quietly depend on a particular model of how gene expression data are distributed. To demonstrate generalizability, the team applied the framework pan-cancer across The Cancer Genome Atlas, identifying 168 gene-cancer combinations in the Unreliable tier spanning 32 cancer types. This catalog provides an immediate, actionable resource: any researcher considering one of these combinations as a biomarker in a given cancer type now has a quantitative warning that the marker’s bulk-level signal is likely to be composition-driven and fragile.

The clinical context of this work is the growing reliance on immune checkpoint blockade, the class of therapies that unleashes T cells against tumors and has transformed outcomes in melanoma, lung cancer, and other malignancies. Biomarkers such as PD-L1 expression, the cytolytic activity score, and gene signatures built from T cell markers are widely used to guide patient selection, yet their performance varies frustratingly across tumor types and treatment settings. The new study does not invalidate these tools, but it reframes the question researchers should ask. Rather than asking whether a biomarker achieves high accuracy in a discovery cohort, the framework argues, investigators should first ask whether the biomarker’s signal is genuinely attributable to the intended cell type and whether it can survive changes in cellular composition. A biomarker that passes those tests may justify validation; one that fails them is likely to join the long list of phenotypically impressive markers that never reach the clinic.

The implications extend beyond immunotherapy. Bulk transcriptomics remains the workhorse of large-scale biomarker discovery across medicine, from autoimmune disease to cardiology, and the composition-driven fragility mechanism the authors describe is not unique to cancer. Any bulk measurement drawn from heterogeneous tissue is vulnerable to the same inflation phenomenon, in which mixture effects manufacture apparent signal that dissolves under cohort variation. By providing open, quantitative diagnostics, the Source Fidelity Index and Biomarker Fragility Index offer a template for pre-validation screening that could be adapted to other fields. The authors, who note that generative artificial intelligence was used only for language refinement while all scientific contributions were human-generated, funded by Taiwan’s National Science and Technology Council, present the work as a practical addition to the biomarker development pipeline. If the framework’s predictions hold as widely as its validation suggests, the era of discovering biomarkers that look brilliant on paper and fail in patients may finally be drawing to a close, replaced by a more honest accounting of what bulk gene expression can and cannot tell us about the biology of disease.

Subject of Research: Computational diagnostics for predicting translation failure of bulk RNA-sequencing immune biomarkers in cancer immunotherapy

Article Title: Phenotypic Inflation and composition-driven fragility: a pre-validation diagnostic framework for immune biomarker translation failure

Article References: Pham, H. M. Q., Feng, P.-H., Chen, C.-L., Lee, K.-Y., & Lin, C.-F. (2026). Phenotypic Inflation and composition-driven fragility: a pre-validation diagnostic framework for immune biomarker translation failure. BMC Bioinformatics. https://doi.org/10.1186/s12859-026-06691-x

Image Credits: AI Generated

DOI: 10.1186/s12859-026-06691-x

Keywords: immune biomarkers, phenotypic inflation, biomarker fragility, bulk RNA-sequencing, tumor microenvironment, immune checkpoint blockade, single-cell RNA sequencing, biomarker translation, cancer immunotherapy, TCGA, transcriptomics, BMC Bioinformatics

The Model Context Protocol, or MCP, is an emerging standard for connecting artificial intelligence assistants to external tools and data sources. Rather than allowing a language model to improvise its way through an analysis, an MCP server exposes a defined set of capabilities that the model can invoke in a structured way. stMCP applies this idea to spatial transcriptomics by pairing a natural language interface with a modular orchestration layer that routes requests to predefined analytical tools. When a researcher asks a question about their tissue data, the language model interprets the intent, the orchestration layer selects the appropriate tool, and the computation runs on the user’s own machine. The results are then returned to the user through a web-based viewer, accompanied by comprehensive analysis logs that document every step.

The distinction between orchestration and execution is central to the design. In many AI-driven analysis systems, autonomous agents generate and run code in extended back-and-forth exchanges with a language model, which can be slow, expensive, and difficult to reproduce. stMCP takes a more constrained approach. Because the analytical operations are performed locally by established tools rather than by code the model writes on the fly, biological data never needs to leave the user’s system, a significant consideration for laboratories handling sensitive or proprietary datasets. The architecture also reduces repeated language-model interaction during computation, meaning the model is consulted for interpretation and routing rather than for every computational step, which improves both efficiency and reproducibility.

Accessibility was a driving concern for the developers. Spatial transcriptomics platforms such as Xenium generate rich maps of gene activity within tissue, but extracting biological insight from those maps has largely been the province of computational biologists. stMCP includes a streamlined installation workflow that reduces the software configuration burden required for deployment, aiming to bring the technology within reach of experimental laboratories without dedicated programming support. Session management allows users to carry on iterative analyses across multiple interactions, building on previous results rather than restarting from scratch each time, and the analysis logging ensures that the full chain of decisions remains auditable long after the session ends.

Benchmarking results reported in the paper suggest that the constrained architecture is robust across the common tasks of spatial transcriptomics analysis. The system handled standard workflows reliably and, importantly, recovered gracefully from ambiguous or unsupported requests, situations that often derail less structured AI tools. The team also demonstrated that stMCP could successfully analyze larger public Xenium datasets and support workflows across multiple spatial transcriptomics platforms, indicating that the framework is not tied to a single technology or data format. This platform flexibility matters in a field where instruments from different manufacturers produce data with distinct characteristics and processing requirements.

Perhaps the most thought-provoking portion of the study is a direct comparison with STAgent and SpatialAgent, two autonomous agent-based systems for spatial transcriptomics analysis. The comparison revealed meaningful differences in execution efficiency, language-model utilization, reproducibility, and deployment requirements between the constrained MCP-based orchestration of stMCP and the more free-form agent-based approach. Autonomous agents promise flexibility, but that flexibility comes at a cost: more calls to the language model, greater computational overhead, and analyses that can be harder to reproduce because the model may take different paths on different runs. stMCP’s predefined tools and structured routing trade some of that flexibility for predictability, a trade-off the authors argue is well suited to scientific work, where reproducibility is not optional.

The implications extend beyond spatial transcriptomics. The authors frame stMCP as a demonstration of how MCP-based architecture can connect natural language interfaces with scientific computing tools generally, balancing accessibility, computational control, and reproducibility. Many scientific domains face the same challenge: powerful, well-validated analysis pipelines exist, but they are locked behind command-line interfaces and steep learning curves. A protocol-driven server that lets a language model orchestrate those pipelines, without letting it rewrite them, offers a template for democratizing computational science without sacrificing rigor. If the pattern holds across fields, the role of the AI assistant shifts from an unpredictable code generator to a reliable translator between human intent and established software.

The work emerged from a collaboration spanning Case Western Reserve University and Duke University School of Medicine, with contributions from departments of biomedical engineering, pharmacology, ophthalmology, and integrative immunobiology. The research was supported by funding from the National Institutes of Health, and the spatial transcriptomics data used in the study was generated with support from the CWRU Imaging Core. The authors note that generative artificial intelligence tools were used to assist with code debugging and language editing during manuscript preparation, while all analytical decisions, data interpretation, and final manuscript content were reviewed and verified by the authors, a transparency practice that mirrors the careful human oversight built into the software itself.

For the spatial transcriptomics community, stMCP arrives at a moment of rapid growth. As instruments scale up and public datasets expand, the bottleneck is shifting from data generation to data interpretation. Tools that lower the computational barrier without lowering scientific standards could determine how broadly the technology’s benefits spread, from large genomics centers to individual laboratories studying development, disease, and tissue architecture. By keeping data local, preserving established methods, and logging every analytical step, stMCP makes a case that the future of AI-assisted science lies not in replacing the computational toolkit but in making it conversational. The study is published open access, allowing researchers everywhere to examine, adopt, and build upon the framework.

Subject of Research: A Model Context Protocol server enabling natural language analysis of spatial transcriptomics data

Article Title: stMCP: spatial transcriptomics with a Model Context Protocol server

Article References: Smith, J. J., Wang, X., McPheeters, M., Widjaja-Adhi, M. A. K., Littleton, S., Saban, D. R., Golczak, M., & Jenkins, M. W. (2026). stMCP: spatial transcriptomics with a Model Context Protocol server. BMC Bioinformatics. https://doi.org/10.1186/s12859-026-06693-9

Image Credits: AI Generated

DOI: 10.1186/s12859-026-06693-9

Keywords: spatial transcriptomics, Model Context Protocol, large language models, bioinformatics, Xenium, gene expression, reproducibility, AI agents, computational biology, open access software, natural language interface, BMC Bioinformatics

Genotype phasing matters because most sequencing technologies read a person’s genome as an unordered mix of genetic variants. At any position where the two chromosomes differ, standard sequencing tells you that a variant is present but not which parent’s chromosome carries it. Reconstructing which variants travel together along each chromosome, producing what geneticists call haplotypes, is essential for understanding inheritance, tracing ancestry, identifying disease-associated regions, and interpreting the functional consequences of mutations. Errors in phasing, particularly so-called phase switches where the assignment flips incorrectly between adjacent variants, can distort downstream analyses and lead to mistaken conclusions about how genetic variation contributes to health and disease.

The dominant phasing tools in use today, including Eagle, Beagle, and Mendel Impute, achieve high accuracy by comparing an individual’s genotypes against large panels of previously phased reference genomes. These methods exploit linkage disequilibrium, the non-random association of variants that arises because stretches of DNA are inherited together across generations. The catch is that the approach depends on having a reference panel that closely matches the ancestry of the individual being phased. Populations that are underrepresented in these panels, particularly those with greater genetic diversity such as African populations, tend to receive less accurate phasing. The reference panels themselves are also enormous, demanding substantial computational resources and limiting scalability as sequencing efforts expand worldwide.

RefFree-Phaser takes a fundamentally different route. Rather than consulting a library of other people’s genomes, the model learns the statistical structure of genomic inheritance directly from sequence data. At its core sits BigBird, a transformer architecture designed to handle long sequences efficiently through sparse attention. Standard transformers compute attention between every pair of positions in a sequence, an operation whose cost grows quadratically with length, which makes them impractical for genomic data where relevant interactions can span millions of base pairs. BigBird’s sparse attention restricts each position to attending over a limited set of neighbors plus a set of global tokens, dramatically reducing computational expense while still allowing the model to capture long-range dependencies across the genome.

The technical pipeline behind RefFree-Phaser reflects careful adaptation of transformer machinery to genomic data. Input sequences are first tokenized and enriched with positional embeddings so the model understands where each variant sits along the chromosome. These representations then pass through a twelve-layer BigBird transformer, producing contextualized hidden states that encode information about each variant in light of its genomic surroundings. From these hidden states, the model derives logits, confidence scores, and binary haplotype predictions that align with the unphased genotypes and true labels used during training. To make the output unambiguous, the researchers adopted a lexicographical haplotype-pairing strategy in which all possible haplotype configurations are sorted and systematically tagged according to their input genotypes, ensuring a consistent and reproducible mapping between predictions and the underlying data.

The evaluation spanned two cohorts, one drawn from a subset of the 1000 Genomes Project and the other from the Omni2.5 million common-variants dataset. Across diverse populations, RefFree-Phaser achieved average phasing accuracies of 92.60 percent on the 1000 Genomes data and 91.56 percent on the Omni2.5 million common variants, figures that place it in the same territory as the established reference-based methods it was compared against. Performance was strongest for individuals of European ancestry and other superpopulations, while accuracy dipped slightly for African individuals. The authors attribute this gap to higher phase-switch rates, greater genetic diversity, and limited representation in the training data, a pattern that mirrors known limitations of the field rather than a flaw unique to this model.

One of the most intriguing findings is that the model does more than simply match the accuracy of classical tools. RefFree-Phaser captured genomic behaviors that conventional methods do not explicitly report, most notably phase switches at heterozygous sites, which the model flags with distinguishable confidence scores. In principle, this means the system not only phases genotypes but also communicates how certain it is at each position, giving researchers a quantitative signal for identifying regions where phasing is unreliable. Such built-in uncertainty estimates could prove valuable for downstream applications, allowing analysts to weight or filter genomic regions according to the confidence attached to their phase assignments.

Perhaps the most striking evidence that the transformer has genuinely learned genomic biology comes from an analysis of its attention patterns. The researchers found that the model’s attention mechanism is driven primarily by genetic relevance, and that the positions receiving the maximum cumulative attention align precisely with linkage disequilibrium markers and with true haplotype crossover sites, the genomic locations where recombination has shuffled genetic material between chromosomes. In other words, without being explicitly taught about recombination, the model appears to have discovered where crossovers occur, validating the contextualized hidden states as a robust feature set that encodes real biological structure rather than superficial statistical patterns.

The implications extend beyond a single benchmark result. A phasing method that requires no external reference panel could be transformative for genomic medicine in populations that reference-based tools serve poorly, and for large-scale sequencing projects where the computational burden of maintaining and querying massive panels becomes prohibitive. It also fits into a broader trend of deep learning models, from protein structure prediction to regulatory element annotation, demonstrating that neural networks can internalize complex biological regularities when given sufficient data and appropriate architectures. The sparse-attention design of BigBird, originally motivated by the need to process long documents, turns out to be well suited to the analogous challenge of long genomic sequences.

Challenges remain before such tools reach routine use. The reduced accuracy for African ancestry individuals underscores a persistent problem in genomics: training data itself carries the biases of historical sampling, and a reference-free model is only as representative as the data it learns from. The study, supported by UTHealth startup funds and a National Human Genome Research Institute grant, was led by Muhammad Nadeem Cheema, with Anam Nazir and Arif Harmanci as co-authors, all based in the Department of Health Data Science and Artificial Intelligence at the D. Bradley McWilliams School of Biomedical Informatics. Their work, published open access on 29 August 2026, demonstrates that genotype phasing can indeed be performed without external reference panels using a transformer-based model that captures long-range genomic dependencies. As training datasets grow more diverse and transformer architectures continue to improve, reference-free approaches like this one may reshape how the field reconstructs the two intertwined stories that every genome tells.

Subject of Research: Reference-free genotype phasing using a BigBird transformer deep learning model

Article Title: RefFree-Phaser: a reference-free transformer-based framework for genotype phasing

Article References: Cheema, M. N., Nazir, A., & Harmanci, A. (2026). RefFree-Phaser: a reference-free transformer-based framework for genotype phasing. BMC Bioinformatics. https://doi.org/10.1186/s12859-026-06572-3

Image Credits: AI Generated

DOI: 10.1186/s12859-026-06572-3

Keywords: genotype phasing, haplotype inference, BigBird transformer, deep learning, genomics, linkage disequilibrium, 1000 Genomes Project, reference-free methods, sparse attention, recombination, BMC Bioinformatics, population genetics

The original AssayBLAST was built around a simple but powerful idea: use the well-established BLAST search engine to compare candidate primers and probes against large sequence databases, and thereby predict how they will perform in a polymerase chain reaction or related amplification assay. This approach allows researchers to screen oligonucleotides computationally against thousands or even millions of potential template sequences, flagging those that might bind to off-target sites or fail to amplify the organisms they were designed to detect. For multi-parameter assays, which combine many primer and probe pairs in a single reaction to detect numerous targets simultaneously, such computational pre-screening is not merely convenient; it is essential for keeping development costs and failure rates under control.

The new version addresses one of the most subtle and consequential failure modes in multiplex assay design: the spatial relationship between paired oligonucleotides. In a functioning amplification reaction, a forward and reverse primer must bind to opposite strands of the template DNA in the correct orientation and within a suitable distance of each other for the polymerase to generate a product. A primer pair that individually looks perfect can be rendered useless if the two primers bind to the same strand, or if their binding sites are too far apart or oriented incorrectly. AssayBLAST v2 introduces an integrated strand and proximity check that validates corresponding oligonucleotides against these geometric requirements, ensuring that predicted amplification events reflect biologically plausible configurations rather than coincidental sequence matches.

This check matters because earlier approaches could, in principle, report a match between a primer and a genomic sequence without verifying whether that match could actually support amplification. In a multiplex assay targeting dozens of pathogens or resistance genes at once, such false positives in the computational analysis can mislead researchers about the expected specificity and coverage of their design. By explicitly evaluating orientation and spacing, the updated software produces a more faithful picture of what the reaction will actually do. The developers describe this as enabling precise validation of corresponding oligonucleotides, and the improvement is backed by a comparative evaluation against the previous version of AssayBLAST, giving users a quantified sense of how much the analysis has matured.

Beyond geometry, the second major addition concerns the outcome of the reaction itself. AssayBLAST v2 does not stop at identifying where oligonucleotides bind; it uses the predicted interactions among all oligonucleotides in the assay to determine the theoretical amplification outcomes. In other words, the software constructs a computational benchmark of which template molecules should be amplified, which should be missed, and how the various primer and probe pairs in a multiplex panel are expected to behave together. This predicted outcome profile then serves as a reference point for downstream wet-laboratory validation, allowing experimentalists to compare observed assay performance against theoretical expectations and to correlate discrepancies with specific oligonucleotide interactions.

The practical implications of this benchmarking capability extend well beyond academic curiosity. Synthesizing primers and probes is a recurring expense in diagnostic development, and a poorly designed multiplex assay can burn through budgets in repeated rounds of synthesis, testing, and redesign. By catching suboptimal oligonucleotide combinations before synthesis, the software streamlines the assay development workflow and reduces the costs associated with suboptimal primer and probe synthesis. For laboratories developing diagnostic panels, particularly in fields such as infectious disease detection where panels must distinguish closely related organisms against a backdrop of host and environmental DNA, the difference between a well-screened and a poorly screened design can determine whether a product reaches the clinic at all.

The third pillar of the update is an adaptive optimization of BLAST parameters that responds dynamically to the size of the database being searched. BLAST searches are governed by numerous tunable parameters, including word size, scoring matrices, and statistical significance thresholds, and the settings that work well for a small custom database of target genomes may be poorly suited to a comprehensive repository containing billions of nucleotides. Fixed parameters force developers into a compromise: settings sensitive enough to detect weak but relevant off-target binding in large databases can become computationally expensive, while settings optimized for speed can miss biologically important interactions. The adaptive approach in AssayBLAST v2 scales its parameters with database size, improving both analytical sensitivity and computational performance simultaneously rather than trading one against the other.

This adaptive strategy reflects a broader trend in bioinformatics, where tools must increasingly operate across databases that grow at staggering rates. Sequence repositories double in size on timescales measured in years, and a screening tool with static assumptions about database scale risks becoming either obsolete or prohibitively slow. By building the scaling behavior directly into the search strategy, the AssayBLAST developers have future-proofed their tool to a degree, ensuring that the reliability of an in silico evaluation does not silently degrade as reference data expand. For users running routine re-evaluations of existing assay panels against updated databases, this consistency is a meaningful quality assurance benefit in its own right.

The team behind the software spans several Jena institutions, combining expertise in RNA bioinformatics, high-throughput analysis, and applied diagnostics. Tom Eulenfeld and Maximillian Collatz, who share first authorship, are based at Friedrich Schiller University Jena, with Collatz also affiliated with the university’s Bioinformatics Core Facility. Sascha D. Braun and Ralf Ehricht bring the diagnostic perspective from the Leibniz Institute of Photonic Technology, the InfectoGnostics Research Campus Jena, and associated translational research centers, including Jena University Hospital. This blend of computational and applied expertise is visible in the design of the update itself: the new features target precisely the points where computational predictions meet laboratory reality, from primer geometry to predicted amplification outcomes. The research was funded in part by the German Research Foundation under Germany’s Excellence Strategy, and the article is published open access under a Creative Commons Attribution license.

For the diagnostic and research communities that depend on molecular multi-parameter assays, the release of AssayBLAST v2 arrives at a moment when the stakes of in silico evaluation have never been higher. Multiplex panels now underpin syndromic testing for respiratory infections, sepsis diagnostics, food safety screening, and antimicrobial resistance surveillance, and each new panel multiplies the number of oligonucleotide interactions that must be anticipated and validated. A tool that can reliably predict amplification outcomes, verify the physical plausibility of primer binding, and adapt its search sensitivity to ever-growing databases addresses the central bottleneck in this workflow. The developers position the update as increasing the robustness and reliability of molecular diagnostics and research applications alike, and the comparative evaluation included in the publication offers users a transparent account of the gains. As computational screening continues to move from a supplementary check to a foundational step in assay design, tools of this kind are set to shape how quickly and how confidently new diagnostic panels reach the laboratory and, ultimately, the patient.

Subject of Research: In silico evaluation and design of molecular multi-parameter PCR assays using BLAST-based oligonucleotide analysis

Article Title: AssayBLAST v2: major update improving reliability and reporting of the in silico analysis of molecular multi-parameter assays

Article References: Eulenfeld, T., Collatz, M., Braun, S. D., & Ehricht, R. (2026). AssayBLAST v2: major update improving reliability and reporting of the in silico analysis of molecular multi-parameter assays. BMC Bioinformatics, 27(1), Article 205. https://doi.org/10.1186/s12859-026-06595-w

Image Credits: AI Generated

DOI: 10.1186/s12859-026-06595-w

Keywords: AssayBLAST, oligonucleotide design, molecular diagnostics, multiplex PCR, BLAST, in silico analysis, primer design, bioinformatics software, amplification prediction, diagnostic assay development, BMC Bioinformatics, computational biology

The study, led by Ryosaku Ota and Naoki Honda of Nagoya University Graduate School of Medicine together with colleagues at Kyoto University, Fujita Health University, and the University of Osaka, introduces a framework called Foldinsight. Its central premise is deceptively simple: protein function does not emerge from a one-dimensional string of amino acids, but from the physical and chemical interactions those amino acids form in three-dimensional space. If a computational model could see the protein the way physics sees it, the authors reasoned, it might both predict variant properties more meaningfully and explain where in the structure those properties come from. Existing machine learning approaches to protein engineering often rely primarily on sequence-derived features, which treat each position in the protein as an independent variable and can miss the spatial context that determines how a mutation actually behaves.

Foldinsight works by converting amino acid sequences into three-dimensional structures using AlphaFold2, the deep learning system that has reshaped structural biology by predicting protein conformations from sequence alone with remarkable accuracy. Once each variant’s structure is predicted, the framework aligns all of the structures in a common coordinate system so that equivalent regions of different variants occupy the same position in space. This alignment step is critical, because it allows the researchers to compare variants not as isolated molecules but as a family of related shapes that can be overlaid and interrogated systematically. The team also uses the pLDDT score, AlphaFold2’s per-residue confidence metric, to assess how reliable each predicted region is, and root-mean-square deviation measures to quantify how much the predicted structures differ from one another.

With the structures aligned, Foldinsight calculates molecular fields on a shared three-dimensional grid spanning the entire protein family. Two kinds of fields are computed: van der Waals fields, which capture the steric shape and volume of the protein at each grid point, and electrostatic fields, which describe the distribution of charge. The result is a fixed-length numerical descriptor for every variant, regardless of how its sequence differs from its relatives. This is the same conceptual machinery that medicinal chemists have long used to compare small drug molecules, where molecular field descriptors underpin classic quantitative structure-activity relationship models. Applying it to a large, flexible membrane protein like a channelrhodopsin is a considerably more ambitious undertaking, and the fixed-length nature of the descriptors is what makes them suitable for regression modeling.

To turn those descriptors into predictions, the researchers fitted regression models to a previously published dataset of channelrhodopsin variants with measured photocurrent properties. The study employed partial least squares regression, a technique well suited to situations where the number of descriptor variables vastly exceeds the number of measured samples, alongside Gaussian process regression, a flexible non-linear method that also provides uncertainty estimates. Model performance was assessed with cross-validation, using mean absolute error and root-mean-square error to gauge how well the models predicted properties they had not been trained on. According to the paper, the molecular-field descriptors captured predictive signals across the measured photocurrent properties, demonstrating that structure-derived features carry information relevant to channelrhodopsin function even when the structures themselves are predictions rather than experimentally determined crystals.

What sets the approach apart from many black-box protein models is its interpretability. Because each descriptor variable corresponds to a specific point in the shared three-dimensional grid, the fitted regression coefficients can be mapped back onto that grid, producing a spatial map of which regions of the protein are associated with which properties. When the authors performed this mapping for the channelrhodopsin dataset, the highlighted regions could be examined in relation to established functional and structural features of the protein, such as the ion-conducting pore and the light-absorbing retinal-binding pocket. This turns the regression model from a mere prediction engine into a hypothesis generator: instead of simply ranking variants, it points experimenters toward specific spatial neighborhoods where mutations are likely to matter.

The practical implications for protein engineering could be substantial. Channelrhodopsin variants are central tools in optogenetics, and variants with tailored properties, such as faster kinetics, red-shifted activation, or altered ion selectivity, are in constant demand across neuroscience and biotechnology. A workflow that prioritizes which variants to synthesize and test, and that explains its reasoning in spatial terms a structural biologist can evaluate, could shorten design cycles and reduce the burden of exhaustive screening. The authors are careful to frame Foldinsight as a tool for generating hypotheses for future mutational experiments rather than replacing them, a stance that reflects the inherent limits of models trained on datasets of finite size.

The study also illustrates a broader trend in computational biology: the repurposing of AlphaFold2 predictions as inputs for downstream machine learning. Predicted structures are approximations, and their accuracy varies across a protein, which is why the framework’s use of confidence scores and structural alignment matters. For channelrhodopsins, a family of seven-transmembrane proteins whose structures have historically been difficult to capture experimentally, the ability to generate consistent, comparable structural models for every variant in a dataset is itself a meaningful advance. It converts a heterogeneous collection of sequences into a coherent structural ensemble on which quantitative analysis becomes possible.

Limitations remain, as the authors acknowledge. The framework was validated on a published dataset rather than on newly generated experiments, and the quality of any prediction ultimately depends on both the accuracy of the predicted structures and the size and diversity of the training data. Cross-validated performance on retrospective data does not guarantee that the spatial maps identify causal mechanisms, only candidate regions worth investigating. Still, the work, which was partly supported by the Japan Agency for Medical Research and Development, JSPS KAKENHI, and the Moonshot R&D MILLENNIA Program, and computed in part on the ROIS National Institute of Genetics supercomputer, offers an open-access, interpretable template that other protein engineers can adapt. As predicted structures grow more accurate and variant datasets grow larger, structure-derived molecular fields may become a standard lens through which the properties of engineered proteins, channelrhodopsins included, are understood and designed.

Subject of Research: Interpretable machine learning modeling of channelrhodopsin variant properties using AlphaFold2-predicted structures and molecular field descriptors

Article Title: Interpretable modeling of channelrhodopsin variant properties using AlphaFold-derived molecular fields

Article References: Ota, R., Sakamoto, M., Aoki, W., & Honda, N. (2026). Interpretable modeling of channelrhodopsin variant properties using AlphaFold-derived molecular fields. BMC Bioinformatics. https://doi.org/10.1186/s12859-026-06625-7

Image Credits: AI Generated

DOI: 10.1186/s12859-026-06625-7

Keywords: channelrhodopsin, AlphaFold2, molecular fields, optogenetics, protein engineering, interpretable machine learning, partial least squares, Gaussian process regression, BMC Bioinformatics, structural bioinformatics, photocurrent properties, variant modeling

The tool, called EWF 2.0, was developed by Jaromir Sant of the University of Turin and the Collegio Carlo Alberto, together with Paul A. Jenkins, Jere Koskela and Dario Spanò of the University of Warwick and Newcastle University. It extends an earlier line of work known as the EWF algorithms, which allow researchers to draw exact samples from the Wright–Fisher diffusion — that is, to generate simulated allele frequency trajectories whose statistical properties match the theoretical model perfectly, without the approximation errors that plague standard numerical schemes. The crucial innovation in version 2.0 is that the population size is no longer required to be constant. Instead, the demographic history can change over time, which mathematically turns the drift coefficient of the diffusion into a time-inhomogeneous one: a process whose random dynamics depend explicitly on when you look at it.

To appreciate why this matters, it helps to understand what exact simulation means in this context. The Wright–Fisher diffusion is a stochastic differential equation, and the most common way to simulate such equations is to discretise time into small steps and approximate the motion between them. These Euler-type schemes are easy to implement but introduce bias: the simulated trajectories are not true draws from the model, and the bias can be difficult to quantify, particularly for quantities that depend on the fine structure of the path rather than just its endpoints. Exact algorithms sidestep this problem entirely. Using clever probabilistic constructions — often based on rejection sampling ideas in which a candidate path is proposed and then accepted or rejected according to criteria that guarantee the output has exactly the right distribution — they produce trajectories that are statistically indistinguishable from the real thing, whatever the intended application.

The difficulty with time-varying demography is that these rejection-based constructions were designed for diffusions with fixed coefficients. When the population size changes, the drift term of the allele frequency dynamics acquires a time-dependent factor, and the bounds and transformations that make the earlier algorithms valid no longer hold. Previous versions of EWF could handle time-varying mutation and selection in certain settings, but the demographic component had to remain static. Since virtually every dataset from natural populations — humans, livestock, pathogens, crop species — carries the imprint of a fluctuating census history, this limitation was a serious obstacle for anyone hoping to use exact simulation in realistic inference pipelines.

EWF 2.0 removes that obstacle. According to the authors, the algorithm accommodates a specified demographic history alongside mutation parameters, a selection function and a set of sampling times, and from those inputs it generates exact draws from the law of the corresponding Wright–Fisher diffusion. Perhaps more strikingly, it can also simulate diffusion bridges: trajectories that are conditioned to start at one allele frequency and end at another. Bridge simulation is a notoriously hard problem, because the conditioning fundamentally alters the dynamics of the process, and the authors note that existing methods cannot handle time-varying mutation and selection rates in this setting. Bridges are precisely what many statistical applications need, since inferring past evolutionary forces often amounts to asking which hidden paths most plausibly connect observed genetic data points collected at different times.

The applications of such a tool extend across modern population genetics. Ancient DNA studies, for example, now routinely provide snapshots of allele frequencies at multiple time points separated by hundreds or thousands of generations. Reconstructing the trajectories between those snapshots — and quantifying the uncertainty around them — requires simulating bridges under a model that includes the demographic history inferred from archaeological and genomic evidence. Similarly, forward-time simulations used to benchmark inference methods, to test the power of genome-wide association designs, or to explore how selection responds to environmental change all benefit from trajectories that are exact rather than approximate. Any bias in the simulator propagates into the conclusions drawn from it, so the guarantee of exactness is not a mathematical luxury but a practical safeguard.

The authors validated their implementation using distributional tests, including Kolmogorov–Smirnov tests and quantile-quantile plots, which compare the empirical distribution of simulated quantities against theoretical expectations. Agreement in these tests provides strong evidence that the algorithm samples from the intended law, since even subtle biases in a simulation scheme tend to show up as systematic deviations in such comparisons. Equally important from a practical standpoint is the performance result: despite its greater generality, EWF 2.0 retains the same runtime as previous versions of the algorithm. Adding the flexibility of time-varying demography therefore comes at no computational penalty, which the authors highlight as ensuring efficiency and scalability for large-scale simulation studies.

The software is written up as an open-access methods paper and the code is freely available on GitHub, lowering the barrier for other groups to adopt it. In a field where simulation underpins everything from method validation to likelihood-free inference, a tool that combines exactness, demographic realism and competitive speed is likely to find rapid uptake. The work was supported by funding from the Italian Ministry of University and Research through a PRIN 2022 grant financed by the European Union’s Next Generation EU programme, reflecting the increasingly international and computationally oriented character of statistical genetics research.

Looking at the broader picture, the advance illustrates a familiar pattern in computational science: the gap between the models researchers want to use and the models their tools can actually handle. The Wright–Fisher diffusion with time-varying demography has been part of the theoretical literature for decades, and approximate simulators have long been able to produce plausible-looking trajectories under it. What was missing was the guarantee — the assurance that a simulated history is a genuine draw from the specified model rather than an artefact of numerical approximation. By extending exact simulation into the time-inhomogeneous regime, EWF 2.0 brings a foundational class of population genetic models fully within reach of rigorous simulation-based inference, and it does so without asking practitioners to sacrifice speed. For studies probing how genes rose and fell across the shifting demographic tides of the past, that combination may prove hard to resist.

Subject of Research: Exact simulation of allele frequency trajectories under the Wright–Fisher diffusion with time-varying demography

Article Title: EWF 2.0: exact sampling of allele trajectories using the Wright–Fisher diffusion with time-varying demography

Article References: Sant, J., Jenkins, P. A., Koskela, J., & Spanò, D. (2026). EWF 2.0: exact sampling of allele trajectories using the Wright–Fisher diffusion with time-varying demography. BMC Bioinformatics. https://doi.org/10.1186/s12859-026-06576-z

Image Credits: AI Generated

DOI: 10.1186/s12859-026-06576-z

Keywords: Wright–Fisher diffusion, exact simulation, population genetics, allele frequency, demography, genetic drift, selection, mutation, diffusion bridge, computational biology, stochastic modelling, BMC Bioinformatics