STREAM, which stands for State Trajectory Representation and Evolution-Aware Monitoring, was developed by Ali Namvar, Sundaresh Ram, Wassim W. Labaki, Stefanie Galban, Njira L. Lugogo, and Craig J. Galban, and published in PLOS Digital Health. At its core, the framework embodies a deceptively simple conceptual shift: instead of asking whether any single vital sign or laboratory value has crossed a threshold, it treats each patient as a single point moving through a multidimensional physiological space. Twenty-six routinely collected clinical measurements—everything from blood pressure and heart rate to standard laboratory panels—define the axes of this space. As a patient’s condition evolves, their point moves along a trajectory through that space, and the geometry of that movement becomes a rich, interpretable signal of stability or decline.

The mathematical engine behind this geometric view is optimal transport theory, a branch of mathematics originally developed to find the most efficient way of moving distributions of mass from one configuration to another. In STREAM’s hands, optimal transport provides a principled way to compare physiological distributions and to discover, directly from the data, natural groupings of patients without any predefined diagnostic categories. Rather than forcing patients into boxes labeled by diagnosis or scoring system, the algorithm lets the structure of the data itself reveal distinct physiological states. Each incoming patient is then mapped to their nearest data-derived state, and their journey over time can be tracked as movement between and within these states. The result is a monitoring framework that is simultaneously data-driven and interpretable—a combination that has proven notoriously difficult to achieve in machine learning for medicine.

To build and test the framework, the team turned to two of the most extensively curated critical care datasets in existence. The development cohort came from the eICU Collaborative Research Database, a multicenter repository encompassing 158,294 ICU stays, which STREAM used to learn its physiological states and calibrate its predictions. External validation was performed on MIMIC-IV, an entirely separate dataset of 84,517 patients drawn from a different health system. This two-dataset design matters enormously in clinical machine learning, where models that perform brilliantly on their training data frequently collapse when confronted with patients, equipment, and care practices from a new institution. Demonstrating that a framework generalizes across centers is the gold standard for evidence that a tool might actually work in the real world.

What STREAM found when it sifted through this mountain of data was striking: five reproducible, data-derived physiological states, each with a distinct clinical signature. These states were not imposed by the researchers; they emerged from the geometry of the data itself, and they reappeared consistently across both databases. Some states corresponded to relatively stable physiology, others to progressively compromised profiles, and the boundaries between them carried real clinical meaning. Because the states are defined by recognizable patterns of vital signs and laboratory values, clinicians can in principle look at a state assignment and understand what it means physiologically—an antidote to the black-box problem that has slowed the adoption of artificial intelligence at the bedside.

The framework’s most dramatic finding concerns what the researchers call state outliers: patients who spend less than 10 percent of their ICU stay within the physiological state to which they were initially assigned. In plain terms, these are patients whose physiology refuses to stay where the data says it should be—a sign of instability that no single threshold would necessarily catch. In the development cohort, state outliers had an ICU mortality rate of 37.6 percent, a staggering sixteen-fold increase compared with the 2.3 percent mortality seen among patients who remained within their assigned states. The external validation cohort told the same story: outliers there experienced 33.5 percent mortality versus 3.2 percent for patients who stayed put, roughly a ten-fold difference. A simple geometric observation—how far a patient drifts from their expected state—turns out to be one of the most powerful mortality signals in critical care.

Beyond identifying dangerous trajectories, STREAM delivers quantitative mortality predictions with unusual precision. In the development data, the model achieved an area under the receiver operating characteristic curve—a standard measure of discriminatory power—of 0.863 at eight hours after ICU admission, rising to 0.903 at 72 hours. For context, values above 0.8 are generally considered strong, and anything approaching 0.9 is exceptional in clinical prediction. Just as important is calibration: a model must not only rank patients correctly but also produce risk estimates that match reality. STREAM’s expected calibration error was a remarkable 0.002, meaning its predicted probabilities were almost perfectly aligned with observed outcomes. External validation on MIMIC-IV preserved robust performance, with AUCs of 0.798 at eight hours and 0.857 at 72 hours—only a modest degradation despite the shift to an entirely different patient population.

Transparency is where STREAM aims to distinguish itself from the growing crowd of predictive models in medicine. Through feature importance analysis, the framework identifies which specific laboratory values and vital signs are most associated with a patient’s movement toward higher-risk states. A rising lactate, a drifting blood pressure, a deteriorating oxygenation profile—whichever features drive a given patient’s trajectory toward danger can be surfaced as an interpretable clinical explanation. This matters because clinicians are rightly skeptical of algorithms that issue risk scores without reasons. By linking the dynamics of state transitions to the underlying measurements that cause them, STREAM offers a monitoring narrative rather than a bare number, potentially giving care teams both an early warning and a starting point for clinical investigation.

The implications extend beyond any single ICU. Because STREAM relies exclusively on 26 routinely collected measurements, it requires no new sensors, no additional blood draws, and no changes to bedside workflow—only the computational machinery to analyze data that hospitals already gather. Its state-based language of stability and drift could complement, rather than replace, existing severity scores, offering a continuous, dynamic view where those tools provide only static snapshots. The framework’s strong discrimination, excellent calibration, and reproducibility across multicenter datasets satisfy many of the statistical preconditions for clinical usefulness, and the authors are explicit that the next step is prospective evaluation: testing STREAM in real time, on real patients, in live ICUs, where the true measure of any monitoring tool is whether it changes decisions and improves outcomes.

Still, the study’s authors and the broader field recognize the distance between retrospective validation and bedside reality. Models trained on historical databases inherit the biases, missing data patterns, and care practices of the institutions that produced them, and no amount of external validation on similar databases fully substitutes for a prospective trial. Yet the core insight of STREAM feels durable: a patient in an ICU is not a collection of independent numbers but a trajectory through physiological space, and the geometry of that trajectory—how far it strays from where it began—carries profound prognostic weight. If future prospective studies confirm what the eICU and MIMIC-IV analyses suggest, the era of threshold-based alarms may give way to something far more intelligent: monitoring systems that watch the shape of a patient’s journey and speak up, with reasons in hand, when that journey begins to bend toward danger.

Subject of Research: Data-driven physiological state monitoring and mortality prediction in intensive care unit patients

Article Title: STREAM: A data-driven framework for physiological state monitoring in ICU patients

Article References: Namvar, A., Ram, S., Labaki, W. W., Galban, S., Lugogo, N. L., & Galban, C. J. (2026). STREAM: A data-driven framework for physiological state monitoring in ICU patients. PLOS Digital Health, 5(10), e0001753. https://doi.org/10.1371/journal.pdig.0001753

Image Credits: AI Generated

DOI: 10.1371/journal.pdig.0001753

Keywords: STREAM, intensive care unit, physiological monitoring, optimal transport, machine learning, mortality prediction, eICU database, MIMIC-IV, patient trajectories, clinical decision support, critical care, predictive modeling

The study, conducted by Hugo Pimentel Tavares and Nilo Antônio de Souza Sampaio of the State University of Rio de Janeiro, is billed as the first application of Wasserstein distance to cross-continental river water quality benchmarking. The metric, formalised in its modern form by mathematician Cédric Villani, measures the minimum cost of transforming one probability distribution into another — literally, how much ‘earth’ must be moved to reshape one histogram into a second. Unlike a median or a compliance rate, it captures differences in location, spread, skewness and tail behaviour simultaneously. Two monitoring networks with identical medians can differ profoundly in their distributional architecture, and only a distributional metric can expose that divergence.

The evidence base is enormous. The researchers harmonised 646,356 physicochemical records from 141 stations operated by CETESB, the São Paulo State environmental agency, with 998,410 records from the Global River Water Quality Archive version 1.4, which aggregates monitoring from 23 European countries under the Water Framework Directive. Seven parameters common to both networks — pH, dissolved oxygen, biochemical oxygen demand, total phosphorus, turbidity, temperature and nitrate — were compared over a shared analytical window of 2016 to 2022. For each of ten São Paulo basins, the Wasserstein-1 distance between its empirical distribution and the pooled European distribution was computed for every parameter, then normalised by the 95th-to-5th percentile range of the combined data to allow cross-parameter comparison.

The results reveal a bifurcation that aggregate reporting has systematically hidden. The Afluentes do Paraíba do Sul basin records a composite normalised distance of just 0.245, and the Cantareira system 0.333 — both within the distributional range of lower-ranking European countries. At the opposite extreme, the Alto Tietê, which receives the cumulative discharge of the São Paulo Metropolitan Region, scores 2.060, a full 8.4 times the cleanest basin, with a dissolved oxygen distance of 7.661 milligrams per litre reflecting a substantial mass of observations near anoxia. The Médio Tietê follows at 1.392. Mann–Whitney U tests confirm statistically significant distributional differences for six of the seven parameters, with large Cliff’s delta effect sizes for temperature, dissolved oxygen and biochemical oxygen demand.

To convert these distances into rankings, the team built a decision matrix covering twenty entities — all ten Brazilian basins and the ten European countries with complete five-parameter median coverage — and applied TOPSIS, a multi-criteria method that scores each entity by its closeness to an ideal solution. Crucially, the criteria weights were not chosen by the analysts but derived objectively through Shannon entropy weighting, which assigns influence in proportion to each parameter’s cross-entity variability. Total phosphorus, spanning more than a thirty-fold range from Austria’s 0.030 milligrams per litre to the Tietê-Médio’s 1.000, dominated with a weight of 0.472. The resulting ranking places Afluentes do Paraíba do Sul seventh of twenty, ahead of Portugal, Poland, Germany and Belgium, while the Alto and Médio Tietê occupy the last two positions, below every complete-coverage European country.

Perhaps the most actionable component is a linear programming convergence model that translates distributional gaps into timelines. The model’s key innovation is its feasibility constraint: rather than assuming engineering targets, it anchors the maximum achievable annual improvement for each parameter to the 90th percentile of historically observed year-on-year improvements in the CETESB record itself, which extends back to 1978. Solved for each basin, the programme identifies which parameter is the binding constraint — the one whose gap-to-feasibility ratio is largest — and computes a critical-path convergence time. Dissolved oxygen emerges as the bottleneck for seven of the ten basins, with convergence times ranging from 4.4 years for the Afluentes do Paraíba do Sul to 14.8 years for the Alto Tietê under the optimistic feasibility assumption, stretching to 5.5 and 28.7 years respectively under a more conservative 75th-percentile assumption.

The contrast between absolute gap size and binding constraint is instructive. The Alto Tietê’s biochemical oxygen demand gap is the largest in the entire matrix at 11.4 milligrams per litre, yet it resolves in just 4.6 years at the historically sustained improvement rate of 2.5 milligrams per litre per year. Dissolved oxygen, with a smaller absolute gap but a far slower feasible recovery rate of 0.59 milligrams per litre per year, dictates nearly fifteen years of sustained effort. In other words, even if every other parameter reached European levels tomorrow, oxygen recovery alone would remain a generational project for the metropolitan corridor — a finding the authors argue supports differentiated governance strategies rather than basin-uniform targets.

One result stands out for its counterintuitive character: nitrate is the only parameter showing no significant difference between Brazilian and European distributions, with a p-value of 0.354 and a negligible effect size. The authors interpret this parity cautiously, as a plausible transient coincidence of trajectories rather than evidence of equivalent governance. Europe’s nitrate levels reflect decades of agricultural intensification followed by regulatory pressure from the Nitrates Directive; Brazil has no equivalent instrument, and the basins with the highest nitrate medians — Piracicaba, Capivari and Jundiaí, at 1.2 to 2.0 milligrams per litre — drain the state’s most intensive sugarcane, citrus and soybean country. Nitrate in the Jundiaí basin has risen more than 800 percent since the early 1990s. The aggregate parity conceals a steep internal gradient, from near-pristine headwaters at 0.200 milligrams per litre to sub-basins already exceeding European concentrations, signalling an emerging eutrophication risk.

The framework’s typological analysis, using principal component analysis and Ward hierarchical clustering, positions the twenty entities along a pollution axis and a nitrate axis that together explain 77.5 percent of the variance. Four clusters emerge: a predominantly European mainstream, an intermediate-impaired group containing the agricultural Piracicaba–Capivari–Jundiaí catchments, the metropolitan Tietê arc as a statistical outlier cluster with no European equivalent, and — the study’s most encouraging finding — a clean-reference group in which four São Paulo basins cluster alongside Portugal. The conclusions proved robust to the analytical choices: rankings held under equal-weight and CRITIC alternative weighting schemes, under wet-season and dry-season stratification, and under three levels of European benchmark stringency.

To test whether the method transfers beyond the network it was built on, the researchers applied the benchmarking and ranking components to 33 Polish and 26 Italian sub-basins under a leave-one-country-out design, recomputing the European reference each time with the target country excluded. Basin scores correlated negatively with an independent measure of anthropogenic pressure — population density from the European Commission’s Global Human Settlement Layer — in both countries, with Spearman coefficients of −0.44 and −0.42. That the association replicates across two hydrologically dissimilar settings that played no role in the method’s development is the substantive validation. The authors are careful to note what remains untested: the convergence-planning component requires multi-decadal basin-level records that the European archive does not provide, and the analysis covers only seven physicochemical parameters, excluding biological quality elements, heavy metals and emerging contaminants. Even so, the framework offers monitoring agencies something conventional indices cannot: a distribution-sensitive key performance indicator, a parameter-by-parameter decomposition of divergence, and a convergence horizon grounded in what the monitoring record itself has shown to be achievable.

Subject of Research: Cross-continental benchmarking of surface water quality monitoring networks using Wasserstein distributional distance and linear programming convergence planning

Article Title: Wasserstein distributional distance and linear programming convergence planning as monitoring network benchmarking tools: cross-continental surface water quality assessment across Brazilian and European river basins

Article References: Pimentel Tavares, H., & de Souza Sampaio, N. A. (2026). Wasserstein distributional distance and linear programming convergence planning as monitoring network benchmarking tools: cross-continental surface water quality assessment across Brazilian and European river basins. Environmental Monitoring and Assessment, 198(10), Article 1081. https://doi.org/10.1007/s10661-026-15914-w

Image Credits: AI Generated

DOI: 10.1007/s10661-026-15914-w

Keywords: Wasserstein distance, optimal transport, water quality monitoring, São Paulo river basins, Water Framework Directive, TOPSIS, linear programming, CETESB, GRQA, dissolved oxygen, total phosphorus, nitrate

A research team led by Yann Maugé and Elias Ventre has now introduced a computational framework designed to tackle this problem head-on. Writing in PLOS Computational Biology, the authors present CardamomOT, a method that builds on their earlier algorithm CARDAMOM and combines a mechanistic model of gene expression with the mathematical machinery of optimal transport. The work, published on June 10, 2026, addresses several long-standing limitations of previous approaches and, according to the authors, represents one of the first methods to explicitly integrate gene regulatory network inference, trajectory reconstruction, and generative modeling of single-cell data within a single unified framework. The implications reach across developmental biology, stem cell research, and drug development, wherever researchers seek to understand how cells change identity over time.

The core idea behind CardamomOT is elegantly simple to state, even if its implementation is mathematically demanding. Single-cell RNA sequencing experiments typically capture cells at a series of time points, producing static snapshots of gene expression across thousands of individual cells. What these snapshots lack is continuity: they show which cell states exist at each moment, but not how individual cells move from one state to the next. CardamomOT bridges these gaps by positing an underlying mechanistic model, a system of differential equations in which a gene regulatory network drives the dynamics of both messenger RNA and the unobserved proteins. The method then uses optimal transport, a mathematical framework for finding the most efficient way to move mass from one distribution to another, to connect the observed snapshots and infer the hidden trajectories that carry cells between them.

Optimal transport has become an increasingly popular tool in single-cell genomics precisely because it provides a principled way to compare distributions of cells across time. Rather than matching individual cells one-to-one, which is impossible when cells are destroyed during measurement, optimal transport computes a coupling between the cell populations at consecutive time points, effectively estimating the flow of cellular mass through state space. What distinguishes CardamomOT from purely statistical applications of this idea is its mechanistic backbone. The couplings are not free-floating mathematical objects; they are constrained to be consistent with a gene regulatory network whose parameters are being calibrated simultaneously. This joint estimation means that the reconstructed trajectories and the inferred network architecture reinforce one another, each constraining the space of plausible solutions for the other.

The new framework directly confronts three specific weaknesses of its predecessor. First, the original CARDAMOM algorithm could only exploit the relative ordering of time points, treating the intervals between measurements as unknown, which discarded valuable temporal information. CardamomOT incorporates the exact time labels of the experimental snapshots, anchoring the inferred dynamics to real physical time. Second, the earlier method relied on restrictive quasi-stationary assumptions about protein dynamics, effectively assuming that protein levels track their messenger RNA targets quickly enough that the transient behavior could be ignored. The new method relaxes this constraint, allowing protein trajectories to be reconstructed as genuine dynamic variables. Third, the original algorithm depended on multiple hyperparameters whose tuning was something of an art; CardamomOT substantially reduces this number, making the method more robust and easier to apply in practice.

A particularly notable feature of the new approach is its use of prior knowledge from the literature. Protein kinetic rates, which describe how quickly proteins are produced and degraded, have been measured in many experimental systems over decades of molecular biology research. CardamomOT allows researchers to incorporate these priors directly into the calibration process, providing biologically grounded constraints on otherwise unobservable parameters. This is a meaningful departure from purely data-driven approaches, which must estimate everything from the data at hand and are therefore prone to identifiability problems. By anchoring the model to known biochemistry, the framework narrows the space of possible solutions and improves the reliability of the inferred network, a strategy that mirrors how mechanistic modeling has long been practiced in systems biology more broadly.

The authors validated their framework on both simulated and experimental datasets, a two-pronged strategy that is standard practice in computational method development but executed here with particular thoroughness. On in silico data, where the ground-truth regulatory network is known because the data were generated from a model, CardamomOT demonstrated consistently improved performance over state-of-the-art methods in both network inference and trajectory reconstruction, while remaining computationally scalable to realistic dataset sizes. On experimental data, where the truth is only partially known, the method reconstructed cellular trajectories, velocity fields, and latent protein levels that were mutually consistent with one another, and the inferred gene regulatory network structures matched known biology from independent experimental studies. This internal coherence across multiple inferred quantities is an important sanity check, since methods that produce trajectories inconsistent with their own inferred networks are unlikely to be capturing real biology.

Perhaps the most forward-looking contribution of the work lies in its generative capabilities. Once the mechanistic model has been calibrated against experimental data, it becomes a predictive engine: researchers can perturb the model in silico, for example by simulating the knockout of a regulatory gene or the application of a differentiation-inducing drug, and observe how the modeled cell population responds. The authors show that the accuracy improvements achieved by CardamomOT make the calibrated model suitable for generating testable predictions of cellular responses to perturbations that have not yet been performed experimentally. This transforms the method from a descriptive tool into a hypothesis-generating one, potentially allowing researchers to prioritize experiments, screen perturbations computationally before committing laboratory resources, and explore reprogramming strategies that push cells toward desired fates.

The significance of this work must be understood against the backdrop of a broader shift in single-cell biology. The field has generated enormous datasets cataloging cell types and states across tissues, development, and disease, but converting these atlases into causal, predictive models of cellular behavior has proven far harder than cataloging itself. Most existing methods for trajectory inference or network inference operate in isolation, each addressing one piece of the puzzle with statistical tools that make no commitment to underlying biological mechanism. CardamomOT’s explicit integration of mechanism, dynamics, and generative modeling points toward a different vision, one in which the ultimate product of a single-cell time-series experiment is not a static map but a calibrated dynamical system that can be interrogated, perturbed, and refined. If approaches of this kind mature as their developers hope, the invisible protein layer that has long constrained the field may finally come into computational view, bringing the dream of predictive, engineering-grade models of cell behavior closer to reality.

Subject of Research: Computational inference of gene regulatory networks and cellular trajectories from single-cell RNA sequencing time-series data using mechanistic optimal transport

Article Title: CardamomOT: A mechanistic optimal transport-based framework for gene regulatory network inference, trajectory reconstruction and generative modeling

Article References: Maugé, Y., & Ventre, E. (2026). CardamomOT: A mechanistic optimal transport-based framework for gene regulatory network inference, trajectory reconstruction and generative modeling. PLOS Computational Biology, 22(10), e1014838. https://doi.org/10.1371/journal.pcbi.1014838

Image Credits: AI Generated

DOI: 10.1371/journal.pcbi.1014838

Keywords: gene regulatory networks, optimal transport, single-cell RNA sequencing, trajectory inference, computational biology, generative modeling, protein dynamics, systems biology, cell differentiation, PLOS Computational Biology, mechanistic modeling, perturbation prediction

Rare variant association testing is one of the most powerful strategies in human genetics for connecting genes to disease and to measurable traits such as blood pressure, lipid levels, and kidney function. Because individual rare variants are too uncommon to test one at a time with statistical confidence, researchers aggregate the variants within a gene and ask whether the burden of rare, presumably damaging variants differs between people with high and low trait values. The catch is that the word “presumably” hides an enormous amount of machinery. With tens of thousands of variants per gene and no experimental evidence for most of them, analysts rely on computational annotation methods to decide which variants to count, which to weight heavily, and which to ignore.

That is where the new study intervenes. The team evaluated annotations from five methods spanning three generations of technology: CADD versions 1.6 and 1.7, a long-standing supervised model trained on evolutionary conservation and genomic features; AlphaMissense, a deep learning system from Google DeepMind that predicts the pathogenicity of every possible missense change in the human proteome; ESM-1b, a protein language model that scores variants based on learned patterns of amino acid sequences; and GPN-MSA, a genomic pretrained model that incorporates multiple sequence alignments across species. These five scorers were plugged into four primary gene-based statistical tests and six annotation-level aggregation tests, creating a large grid of method combinations evaluated against the same trait data.

To compare the combinations fairly, the researchers developed a novel evaluation framework based on optimal transport, a mathematical technique originally developed to compare probability distributions and solve problems in logistics and image processing. Applied here, optimal transport allowed the team to quantify how well each annotation method partitioned genetic signal across labeled sets of variants, producing relative measurements of two distinct properties: calibration, meaning whether the test’s statistical behavior matches its theoretical expectations, and power, meaning its ability to detect true associations. Separating these two qualities matters because a test can appear successful simply by being miscalibrated in a way that generates more hits, a trap that has complicated earlier benchmarking efforts in the field.

The results were, by the authors’ own framing, markedly divergent. Tests built on CADD labels achieved the highest signal separation, meaning they did the best job of concentrating genuine genetic signal into the variants flagged as damaging. At the opposite end of the spectrum, tests using AlphaMissense labels showed the lowest calibration, indicating that the statistical behavior of gene-based tests fed with AlphaMissense scores departed most from theoretical expectations, even though AlphaMissense is widely regarded as one of the most accurate single-variant pathogenicity classifiers in clinical interpretation contexts. The lesson is uncomfortable but important: a scorer that excels at classifying individual variants as pathogenic or benign does not necessarily excel at the aggregate, gene-level task of boosting association power.

Perhaps the most visually striking result concerned evolutionary constraint. When the researchers examined which genes were flagged as significant by each annotation method, they found that hits from tests using GPN-MSA labels were strongly enriched among genes with high evolutionary constraint, reaching up to a 5.8-fold enrichment, compared with only 2.1- to 2.8-fold enrichment using the other methods. Evolutionary constraint measures how intolerant a gene is to mutation over millions of years of mammalian evolution, and constrained genes are known to be enriched for genuine disease biology. The GPN-MSA result suggests that a scorer grounded in cross-species sequence alignments directs statistical power toward exactly the genes where rare variant signals are most likely to be real.

Crucially, these differences in variant prioritization measurably shifted the patterns of genetic discovery. Different annotation methods did not merely reshuffle the rankings of the same genes; they changed which genes crossed the threshold of significance at all, meaning that a laboratory’s choice of scorer can determine whether a biologically important gene is reported or missed entirely. Yet the study also found reassuring signs of statistical health across the board: the tests showed low genomic inflation, the standard diagnostic for spurious associations driven by confounding, and similar replication rates when findings were tested in independent data. In other words, all the method combinations were behaving statistically responsibly; they were simply looking for different things.

The authors distill their findings into three conclusions with direct practical consequences for the field. First, no single combination of annotation method and statistical test is currently optimal for gene-level association testing, so there is no defensible default that a study can adopt without justification. Second, results from different methods and tests can and should be aggregated to increase study power, since the methods appear to capture partially overlapping but non-identical slices of the underlying biology, and combining them recovers signal that any one approach loses. Third, integrating the information content of variant annotations and gene-level annotations remains a key open problem, a gap that the authors identify as the central methodological challenge facing rare variant genetics.

The scale of the underlying data lends the conclusions considerable weight. The analysis drew on the UK Biobank under application number 44257 and additionally used data from the National Institutes of Health’s All of Us Research Program, with the authors acknowledging the hundreds of thousands of participants whose genomes and health measurements made the comparison possible. All authors were employees of Genentech at the time the work was performed, and several hold Roche stock, a disclosure that situates the study within an industrial research environment where rare variant discovery pipelines feed directly into drug target identification. For pharmaceutical companies, a method that systematically misses constrained genes is not an abstract statistical concern; it is a missed therapeutic hypothesis.

For the broader community, the study arrives at a moment when machine learning scorers are proliferating faster than the field can benchmark them. AlphaMissense, ESM-1b, and GPN-MSA represent a new wave of models built on deep learning architectures originally developed for natural language, and each embodies different assumptions about what makes a variant harmful: supervised training on known disease mutations, learned protein sequence grammar, or evolutionary conservation across species. The new results demonstrate that these assumptions translate into genuinely different discovery profiles at scale, not merely different numbers on a leaderboard. Researchers planning gene-based association studies should treat annotation choice as a first-order experimental design decision, ideally running multiple scorers and aggregating results, and method developers should recognize that the gene-level aggregation setting, not just single-variant classification accuracy, is where their tools will ultimately be judged. The full study is openly available under a Creative Commons license, with the DOI 10.1186/s12864-026-13379-2.

Subject of Research: Comparative performance of machine learning variant annotation methods in gene-level rare variant association testing

Article Title: Markedly divergent performance of variant annotation methods for gene-level association testing

Article References: Aguirre, M., Irudayanathan, F. J., Crow, M., Hejase, H. A., Menon, V. K., Pendergrass, R. K., McCarthy, M. I., & Fletez-Brant, K. (2026). Markedly divergent performance of variant annotation methods for gene-level association testing. BMC Genomics. https://doi.org/10.1186/s12864-026-13379-2

Image Credits: AI Generated

DOI: 10.1186/s12864-026-13379-2

Keywords: variant annotation, gene-level association testing, rare variants, UK Biobank, CADD, AlphaMissense, ESM-1b, GPN-MSA, optimal transport, genetic association study, machine learning, genomics

A new study published in Applied Intelligence by Jiaming Liu and Qi Tan of South China Normal University and Pei Yang of South China University of Technology tackles exactly this challenge. The team proposes a framework called Condition-Controlled Optimal Transport, or CCOT, which models the transformation between unperturbed and perturbed cellular distributions as a mathematical transport problem. Rather than trying to pair individual cells across conditions, which is impossible with the available data, the method learns a mapping between entire probability distributions of gene expression states. The work, published as volume 56, article 470 of the journal, was supported by the National Natural Science Foundation of China and the Guangdong Basic and Applied Basic Research Foundation.

Optimal transport is a branch of mathematics with roots stretching back to Gaspard Monge in the eighteenth century, concerned with finding the most efficient way to move mass from one distribution to another. In recent years it has found fertile ground in machine learning, and in single-cell biology specifically, where researchers have used it to align datasets across conditions, modalities, and time points. The intuition is appealing: if a drug pushes a population of cells from one transcriptional state distribution to another, the most plausible transformation between those distributions can be framed as an optimal transport map. Learning that map from data then allows one to generate predictions of how a new population of cells will respond.

The central innovation of CCOT lies in its condition-controlled mechanism, which the authors describe as bridging conditional and unconditional models within the optimal transport framework. The idea echoes classifier-free guidance, a technique popularized in the image generation community, where a model is trained to operate both with and without conditioning information and the two modes are interpolated at inference time. In CCOT, the conditioning signal is the drug perturbation itself. By learning transport maps that can function with or without the drug condition, the framework gains a knob that controls how strongly the perturbation shapes the predicted cellular response, allowing the model to balance faithful in-distribution prediction against robust generalization.

Under the hood, CCOT leverages partially input convex neural networks to parameterize the transport maps. Input convex neural networks, introduced in the machine learning literature in 2017, are architectures whose outputs are convex functions of their inputs, a property that makes them natural candidates for computing optimal transport maps via the dual formulation of the transport problem. Prior work, including the supervised training of conditional Monge maps and neural optimal transport approaches for single-cell perturbation responses, established this connection. CCOT extends it by making the condition control explicit and by aligning the learned transport with biological priors, one of three key innovations the authors list alongside the condition-controlled mechanism and cross-drug generalization.

Cross-drug generalization is arguably the property with the greatest practical stakes. Drug discovery pipelines routinely need to prioritize candidate compounds, and experiments are expensive; a model that can only describe drugs it has already seen is of limited use. CCOT is designed to generalize to novel therapeutic compounds, predicting responses out-of-distribution rather than merely interpolating within the training data. This is a notoriously difficult setting, because a new molecule may act through mechanisms only partially represented in the training set, and models can easily fail by defaulting to memorized responses or by producing biologically implausible expression profiles.

The evaluation was carried out on SciPlex3, a benchmark dataset drawn from the massively multiplex chemical transcriptomics work published in Science in 2020, which profiles gene expression at single-cell resolution across many cell types, drugs, and doses. The authors report that CCOT significantly outperforms state-of-the-art methods in capturing complex and nonlinear cellular responses, with its advantage most evident in capturing higher-order statistical properties and gene-specific responses for differentially expressed genes. Because no ground-truth paired control and perturbed cells exist, the evaluation relies on distributional metrics, including maximum mean discrepancy, energy distance, perturbation signatures, the coefficient of determination, and the Fréchet Inception Distance, each probing a different aspect of how closely the predicted distribution matches the observed one.

The appendix results add useful nuance. Across the full 977-gene space, not just the top 50 differentially expressed genes, CCOT retains its advantage on the distribution-level metrics, improving out-of-distribution maximum mean discrepancy by 66 percent over the next best model, while a competing method, ChemCPA, holds a marginal and statistically insignificant lead on some moment-based metrics. Ablation studies underscore the architecture’s dependencies: removing the drug encoder proves catastrophic, driving the coefficient of determination to a dismal negative value, while removing the condition-control mechanism or the cell type alignment degrades every metric. The authors also document a sensitivity analysis of the inner iteration count, finding stable performance for values of five or greater but a performance collapse at fifteen due to the known instability of adversarial min-max optimization, and they settled on ten inner iterations to match baseline configurations exactly.

Training dynamics receive careful treatment as well. Because the dual potential losses in the min-max optimization are adversarial and oscillate rather than decrease monotonically, the team monitors maximum mean discrepancy on the in-distribution test set as the true indicator of distributional alignment, evaluating every fifty batches and selecting the checkpoint with the minimum value as an implicit early stopping mechanism. Over a fixed budget of 100,000 outer iterations, matching the defaults of the CELLOT and CondOT baselines for a strictly controlled comparison, the losses converge from large negative values toward zero with narrowing oscillations, and the best checkpoint was selected at iteration 98,400 within a stable plateau rather than during an unconverged descent.

The broader significance of this work sits at the intersection of generative machine learning and pharmacology. If models like CCOT can reliably predict how cells respond to compounds they have never seen, the implications range from cheaper pre-screening of drug candidates to mechanistic insight into how perturbations propagate through cellular signaling networks. The authors position the framework as a tool for advancing drug discovery and understanding cellular mechanisms, and they have made the complete source code, including training scripts, preprocessing steps, and evaluation code, publicly available on GitHub, with the datasets themselves drawn from publicly accessible repositories. As with any computational prediction, validation against new laboratory experiments remains the ultimate test, but the study offers a principled mathematical route through one of single-cell biology’s most persistent obstacles: predicting the future state of a cell that must be destroyed in order to be measured.

Subject of Research: Machine learning-based prediction of single-cell drug perturbation responses using optimal transport

Article Title: Condition-Controlled optimal transport for cellular perturbation response prediction

Article References: Liu, J., Tan, Q., & Yang, P. (2026). Condition-Controlled optimal transport for cellular perturbation response prediction. Applied Intelligence, 56(15), Article 470. https://doi.org/10.1007/s10489-026-07521-6

Image Credits: AI Generated

DOI: 10.1007/s10489-026-07521-6

Keywords: single-cell sequencing, optimal transport, perturbation response prediction, drug discovery, input convex neural networks, classifier-free guidance, SciPlex3, gene expression, machine learning, computational biology, out-of-distribution generalization, ChemCPA

The study, published open access in the journal Machine Learning by Sintija Stevanoska, Christian L. Camacho Villalón, Sašo Džeroski and Katharina Dost, tackles a well-known pain point in applied machine learning. Labels are expensive: healthcare systems accumulate millions of patient records, yet expert-annotated outcomes remain scarce. Self-supervised learning promises to bridge that gap by defining pretext tasks from the intrinsic structure of data — masking features and reconstructing them, corrupting values and asking a model to spot the damage, or learning invariant representations across contrastive views. For tables, however, these methods inherit an awkward problem. Unlike images or sentences, tabular rows lack spatial or sequential structure, mix numerical and categorical features, and often contain missing values and tangled inter-dependencies. And pretraining on very large unlabeled pools is computationally expensive while potentially importing noisy or unrepresentative samples that actively degrade learning.

The researchers asked a deceptively simple question: is a small, well-chosen subset of unlabeled data sufficient for effective pretraining? To answer it, they ran what they describe as the first large-scale evaluation of unlabeled data subsampling for tabular self-supervised learning, spanning 25 benchmark datasets drawn from the OpenML-CC18 suite and covering domains from biology and finance to medicine and climate science. Four representative self-supervised models were tested — TabNet, VIME, SCARF and SAINT — deliberately chosen to span the field’s three main paradigms. TabNet and VIME learn by masking or corrupting features and reconstructing them, SCARF relies on contrastive views generated by resampling features from their empirical distributions, and SAINT combines contrastive and denoising objectives inside a transformer that attends over both rows and columns. This diversity of pretext tasks and encoder architectures ensured the conclusions would not hinge on one particular model family.

At the heart of the study lies a family of selection strategies borrowed and adapted from active learning, the branch of machine learning concerned with identifying the most informative examples to label. The team compared random sampling against three principled approaches. Disagreement sampling, based on the classic query-by-committee method, trains a committee of 100 decision trees on bootstrap samples of the labeled data, restricting each tree to a random subset of features in the manner of random forests, and scores every unlabeled point by the normalized entropy of the committee’s votes — a value from 0, indicating complete agreement, to 1, indicating uniform disagreement. Diversity sampling applies k-means clustering to the unlabeled pool and picks one representative per cluster, the point closest to each centroid, enforcing a minimum cluster size to avoid crowning representatives of tiny outlier groups. Optimal transport sampling treats the labeled and unlabeled sets as empirical distributions, computes an entropically regularized transport plan between them with the Sinkhorn algorithm, and assigns each unlabeled point its expected transport cost to the labeled distribution — a low score signaling strong alignment with the labeled data, a high score signaling underrepresented regions of the feature space.

Crucially, all of these strategies are deliberately model-agnostic. They are computed before any self-supervised pretraining begins, using no gradients, losses, embeddings or augmentations from the downstream model. Score-based methods partition the unlabeled pool into percentile brackets — low, medium and high — and samples are then drawn from a chosen bracket, either uniformly at random or via the diversity-based cluster representatives. This design allowed the authors to evaluate the sampling approaches consistently across all four pretext objectives, isolating the effect of data selection from the idiosyncrasies of any particular architecture.

The experimental scale was formidable. Labeled subsets of 100, 500 and 1000 examples were drawn under both unbiased, stratified conditions and a deliberately biased regime in which labeled examples were preferentially selected from one branch of a decision stump, simulating real-world selection bias. Unlabeled subsets of 100 to 1000 examples were drawn under every combination of strategy, subset size, bracket and five random seeds. Performance was measured through three lenses: downstream classification quality using macro-averaged area under the precision-recall curve, total wall-clock time including both selection and training, and representation quality assessed by linear probing — training a simple logistic regression classifier on frozen pretrained embeddings to test whether class information is linearly encoded in the learned latent space.

The headline finding is striking. Subsampling only 100 to 1000 unlabeled examples often matched or exceeded full-data pretraining across both labeled-data regimes. In the low-label, unbiased setting, the best sampling strategies reached up to 119 percent of the full-data baseline and 114 percent of a purely supervised baseline that skipped pretraining entirely. Under biased labeled selection the gains grew even more pronounced, climbing to 123 percent of the full-data baseline and 115 percent of supervised learning. The effect was strongest exactly where it matters most: when labels were scarcest, and the choice of pretraining data had the greatest opportunity to shape what the model learned. Optimal transport distance and disagreement-based sampling performed most consistently, but the humble random sampler was surprisingly competitive, hinting that for small labeled sets, simply reducing redundancy and avoiding overfitting to the unlabeled distribution may itself act as a form of regularization.

The computational story is equally compelling. Pretraining on the full unlabeled pool consistently demanded 60 to 70 percent more time than the sampling strategies, and the selection step itself added a negligible share of the total runtime. Random sampling typically finished in 29 to 42 percent of full-data time, with optimal transport close behind; diversity-based selection was the most expensive to compute, consuming up to 70 percent of the full-data baseline’s runtime without delivering consistent accuracy dividends. In other words, competitive or better predictive performance was achieved at roughly 30 to 40 percent of the computational cost — a margin that alone would justify subsampling in resource-constrained environments even without accuracy improvements. Linear probing reinforced the picture: targeted subsampling consistently matched or exceeded the representation quality of full-data pretraining, and in many configurations full-data self-supervision actually fell below the supervised baseline, an adverse effect the selective strategies appeared to mitigate, particularly under label bias.

To understand when and why subsampling pays off, the team conducted a meta-analysis linking dataset characteristics to observed gains. They extracted statistical and complexity-related metafeatures for each dataset — among them the performance of a nearest-neighbor classifier built on the single most important feature, the stability of local neighborhood structure across cross-validation folds, the minimum support of decision-tree leaves, the median absolute deviation of the features, and trimmed means capturing distributional asymmetry. The patterns that emerged were model-specific rather than universal. TabNet and SCARF gained most from subsampling on datasets where a single dominant feature carries the predictive signal, while SAINT and VIME tended to benefit more on structurally complex tasks; datasets containing a feature with strongly asymmetric central mass favored using all the data. No single metafeature predicted gains across all architectures, pointing to a decisive interplay between dataset geometry and each model’s inductive bias.

The authors are candid about limitations and open questions. The number and size of unlabeled subsets and the granularity of the score brackets were fixed by design, so the study does not determine an optimal data budget for a given dataset, and the meta-analysis, based on 25 datasets, is exploratory rather than confirmatory. Future directions include adaptive budget selection guided by validation performance or meta-learned dataset properties, curriculum-style ordering of the unlabeled pool, incorporating self-supervised model feedback into the selection loop, and training a meta-learner to recommend strategy and subset size automatically. Even so, the message for practitioners is already actionable and, in its own way, subversive: in tabular self-supervised learning, the brute-force assumption that more unlabeled data is always better does not survive scrutiny. A few hundred well-chosen — or even randomly chosen — examples can deliver richer representations and better predictions at a fraction of the cost, a result that could reshape how resource-conscious teams approach pretraining on the structured data that quietly runs the world.

Subject of Research: Unlabeled data selection for efficient self-supervised learning on tabular data

Article Title: Less is More: Unlabeled Data Selection for Efficient Tabular Self-supervised Learning

Article References: Stevanoska, S., Camacho Villalón, C. L., Džeroski, S., & Dost, K. (2026). Less is More: Unlabeled Data Selection for Efficient Tabular Self-supervised Learning. Machine Learning, 115(9), Article 208. https://doi.org/10.1007/s10994-026-07133-8

Image Credits: AI Generated

DOI: 10.1007/s10994-026-07133-8

Keywords: self-supervised learning, tabular data, data selection, unlabeled data, active learning, optimal transport, query-by-committee, computational efficiency, representation learning, machine learning benchmarks, TabNet, VIME