The biological logic behind the approach rests on a concept called synthetic lethality. When a tumour cell loses the function of one gene, it can often survive because a closely related paralog compensates for the loss. Knock out both genes simultaneously, however, and the cell dies, while normal cells that retain functional copies of both genes remain unharmed. This therapeutic window has already produced clinical successes, most famously the PARP inhibitors used against BRCA-deficient breast and ovarian cancers. Yet the authors of the new study argue that the field’s dominant discovery tool, genome-wide CRISPR screening, is structurally blind to many of these interactions. Standard dependency maps, such as the widely used Cancer Dependency Map, largely capture single-gene effects. When two paralogs back each other up, knocking out either one alone produces no detectable fitness defect, and the underlying dependency goes unnoticed.

Combinatorial CRISPR screens, in which genes are knocked out in pairs, can reveal these masked dependencies, but they are expensive, labour-intensive and therefore limited in scale. More importantly, the paralog synthetic lethal effects they uncover are highly context-dependent, varying across tumour types and genetic backgrounds. A paralog pair that is jointly essential in HER2-amplified breast cancer may be entirely dispensable in lung cancer, and vice versa. To translate paralog synthetic lethality into clinical practice, researchers need to know not just which pairs are synthetic lethal in principle, but which pairs constitute actionable vulnerabilities in which specific cancer settings. That is the gap the Dublin-led team set out to close.

The researchers, led by Narod Daldal and corresponding author Colm J. Ryan, with contributions from Hamda B. Ajmal and David J. Adams, developed a machine learning classifier trained to predict cell-line-specific synthetic lethality between paralog pairs. Rather than relying on a single data type, the model integrates features drawn from transcriptomics, genomics, gene essentiality profiles and protein-protein interaction network context. The intuition is straightforward: if two paralogs are expressed at different levels in a given cell line, if one of them already shows partial essentiality, or if their interaction partners within the cellular network suggest functional compensation, the probability that the pair is synthetic lethal in that specific context changes accordingly. The classifier learns these patterns from pairs where combinatorial screening data already exists and applies them to pairs that have never been tested.

A central methodological strength of the study lies in its evaluation framework, which was deliberately designed to test generalisation across three biologically distinct scenarios. The first asks whether the model can predict synthetic lethality for paralog pairs it has seen before, but in cell lines it has never encountered. The second asks whether it can generalise to entirely unseen paralog pairs within cell lines it knows. The third, and hardest, asks whether it can make predictions for unseen pairs in unseen cell lines. This graduated testing matters because a model that merely memorises known pairs would be of little use for discovery, whereas a model that generalises to novel combinations can genuinely guide new experiments. The results showed that the model generalised to both unseen pairs and unseen cell lines, although the combination of both unknowns remained the most challenging scenario, as expected.

Underpinning the predictions is a finding about which signals carry the most information. The cell-line-specific expression levels and essentiality profiles of the paralogs themselves, and of their interaction partners, proved to be the most informative features for predicting synthetic lethal interactions. In other words, the model does not simply ask whether two genes are related; it asks whether, in a particular tumour cell, the pattern of activity across the paralog pair and its network neighbourhood is consistent with compensatory redundancy. This context sensitivity is precisely what single-gene dependency maps lack, and it is what allows the resulting resource to be stratified by disease subtype and biomarker status.

To assess how well the predictions held up, the team applied the model to 33,419 paralog pairs across 1,005 cancer cell lines and compared its output against independent combinatorial CRISPR screening data. One of the most striking conclusions from this validation is a sobering one for the field: the agreement between predicted and experimentally observed interactions was comparable to the agreement observed between independent experimental studies of the same pairs. This suggests that the ceiling on predictive accuracy may be set as much by the reproducibility of the experiments themselves as by the performance of the model. In a discipline where different laboratories screening the same gene pairs can reach different conclusions, a computational predictor that matches the level of inter-experimental consistency is performing at a biologically meaningful standard, and the authors suggest that improving experimental reproducibility and model accuracy will need to advance together.

The practical value of the framework is illustrated through its application to HER2-amplified breast cancer, a clinically defined subtype driven by amplification of the ERBB2 receptor gene. When restricted to this context, the model successfully recovered known synergistic paralog relationships already documented in the literature, providing an internal check on its reliability. More importantly, it also predicted novel biomarker-associated vulnerabilities that had not previously been experimentally characterised, generating a ranked list of candidate paralog targets specific to this tumour type. Each such prediction represents a hypothesis that can be prioritised for targeted combinatorial screening, dramatically narrowing the search space that would otherwise need to be explored experimentally.

To make the resource usable by the wider research community, the team has released all genome-scale predictions through an interactive web portal, available at cancergenetics.github.io/paralogmap. The portal allows researchers to query specific paralog pairs, filter by cancer type or cell line, and identify context-specific dependencies relevant to their disease of interest. The authors position the resource as a tool for hypothesis generation, for guiding targeted combinatorial screening campaigns, and for facilitating the identification of clinically actionable paralog targets. Because the underlying article is open access, the full dataset and predictions are available without restriction, lowering the barrier for laboratories that lack the computational infrastructure to build such models themselves.

The broader significance of the work lies in what it says about the next generation of precision oncology. Single-gene targets are finite, and many of the most obvious dependencies have already been mapped. Combination strategies aimed at synthetic lethal pairs multiply the therapeutic search space enormously, but only a systematic, context-aware prioritisation framework can make that space tractable. By combining machine learning with the rich molecular profiles now available for hundreds of cancer models, and by validating against the gold standard of combinatorial CRISPR screening, this study offers a template for how hidden, context-specific vulnerabilities can be surfaced at scale. If the predicted vulnerabilities withstand experimental follow-up, paralog pairs could become a rich source of biomarker-guided combination therapies, extending the synthetic lethality paradigm beyond its current clinical footholds and into tumour types where redundant gene pairs have so far kept their weaknesses safely out of sight.

Subject of Research: Machine learning prediction of context-specific paralog synthetic lethal vulnerabilities in cancer

Article Title: Systematic prioritisation of context-specific paralog pair vulnerabilities in cancer

Article References: Daldal, N., Ajmal, H. B., Adams, D. J., & Ryan, C. J. (2026). Systematic prioritisation of context-specific paralog pair vulnerabilities in cancer. Genome Medicine. https://doi.org/10.1186/s13073-026-01759-y

Image Credits: AI Generated

DOI: 10.1186/s13073-026-01759-y

Keywords: paralogs, synthetic lethality, precision oncology, machine learning, CRISPR screening, genetic dependencies, cancer cell lines, HER2-amplified breast cancer, dependency map, Genome Medicine, Systematic, prioritisation

The authors argue that the pharmaceutical industry’s dominant mental model, the linear pipeline in which discovery hands off to preclinical testing, then to phase I, II and III trials, is fundamentally mismatched to the complexity of human biology. In a pipeline, information flows forward and decisions are made at discrete gates with limited opportunity to revisit earlier assumptions. When a candidate fails in phase II, the most common reasons, according to decades of attrition analyses, are not pharmacokinetics or safety but lack of efficacy, meaning the underlying biological hypothesis was wrong, or the wrong patients were enrolled. By that point, hundreds of millions of dollars have been spent, and the evidence that could have corrected the hypothesis has often been discarded rather than systematically captured and fed back into the program.

Shalita and Amit propose reframing drug development as an iterative learning system driven by continuous data feedback, a structure they illustrate most thoroughly with immunotherapies, one of the most transformative and also most unpredictable areas of modern medicine. Cancer immunotherapies, from checkpoint inhibitors to CAR T cells, work by manipulating the immune system itself, and the immune system is exquisitely context-dependent: it differs between tissues, between individuals, between tumor types, and even within a single tumor across space and time. The authors contend that high-dimensional, multimodal molecular readouts, particularly single-cell and spatial profiling of therapeutic perturbations, can serve double duty. They can de-risk development by revealing early whether a drug is engaging its intended biological circuits in human patients, and they can deepen mechanistic understanding of how human immune responses actually work, knowledge that in turn sharpens the next round of therapeutic design.

The technical core of the framework is a feedback loop analogous in structure to reinforcement learning. In reinforcement learning, an agent acts, the environment returns a reward signal, and the agent updates its policy accordingly. In the proposed biomedical analogue, the intervention, whether a small molecule, antibody, cytokine or engineered cell, is the action; the patient’s molecular and clinical response, measured through single-cell sequencing, spatial transcriptomics, and longitudinal clinical assessments, is the feedback signal; and the mechanistic models of disease, sometimes framed as ‘virtual cell’ models, are the policies being updated. Each cycle of treatment and measurement refines the model, and the refined model guides the design of the next intervention. Crucially, the authors emphasize that human data, not animal or cell-line data, must provide the authoritative signal, because the immune circuitry that determines success or failure in patients is precisely what current preclinical systems most often fail to reproduce.

This emphasis on context is a direct response to a long-standing weakness in computational biology. Much effort has gone toward building a universal ‘virtual cell’, a comprehensive model of cellular behavior that would generalize across all conditions. Shalita and Amit argue instead for context-specific virtual cell models: models tuned to a particular disease state, tissue microenvironment, or patient population. The distinction matters because the same molecular circuit can behave oppositely in a healthy tissue and an inflamed tumor. Single-cell atlases such as the Human Cell Atlas and Tabula Sapiens, together with disease-specific atlases of tumors and inflamed tissues, provide the reference substrate, but the framework demands that these references be continuously updated with perturbation data, the readouts of what actually happens when a therapeutic is applied, rather than static snapshots of untreated biology.

The toolkit for generating perturbation data has matured rapidly. Pooled CRISPR screens coupled to single-cell RNA sequencing, first demonstrated in 2016, now allow researchers to map genotype-phenotype relationships at genome scale, and newer methods extend this to spatially resolved, multimodal readouts directly in tissue. Harmonized datasets of single-cell perturbation experiments are becoming shared community resources, and massively multiplexed chemical screens can profile thousands of drug candidates at single-cell resolution. The authors argue that these technologies should not sit solely in discovery labs but should be embedded throughout the development process, including in early-phase clinical trials, where serial biopsies and blood draws can reveal within weeks whether a therapy is reprogramming the intended immune cell states, long before conventional efficacy endpoints mature.

Immunotherapy case studies illustrate the cost of the linear model. Resistance to checkpoint blockade frequently traces to specific tumor or immune cell states, such as loss of interferon-gamma pathway genes or exclusion-prone cancer cell programs, that only single-cell analysis of patient samples could reveal. CAR T cell programs have encountered off-tumor toxicity when single-cell atlases later showed that the nominal tumor antigen was expressed in unexpected normal tissues, and neurotoxicities have been linked to mechanisms invisible to conventional monitoring. Conversely, single-cell analysis of patients’ pretreatment immune landscapes is increasingly able to predict who will respond to bispecific T cell engagers or CAR T therapy, pointing toward patient stratification as a first-class design input rather than a post hoc statistical adjustment. Each of these findings, the authors note, is exactly the kind of feedback signal that could have redirected a development program earlier had the loop been closed in real time.

The framework also carries implications for how development organizations are structured. The authors point to industry analyses showing that the best-performing research and development organizations succeed not by generating more candidates but by making better decisions about which hypotheses to pursue and when to abandon them, echoing the observation that quality of biological rationale, evidenced for example by the strong correlation between human genetic support and drug approval success, beats quantity. An iterative, data-centric system institutionalizes that principle: every trial is also an experiment, every measurement updates the model, and termination of a candidate is not a sunk cost but a data acquisition event that improves the next iteration. The authors draw a parallel to reinforcement learning from human feedback in artificial intelligence, where the model is continually aligned to human judgments; here, the model is aligned to human molecular and clinical reality.

Significant obstacles remain. Single-cell and spatial assays are still expensive and logistically demanding inside clinical trials, computational models are sensitive to batch effects and require rigorous validation, regulatory pathways are built around linear evidence generation, and access to high-quality longitudinal patient samples is uneven. The authors argue these are engineering and policy problems rather than fundamental barriers, and point to ongoing initiatives, from large-scale cell atlasing efforts to public virtual cell challenges, as evidence that the community is assembling the necessary infrastructure. Their perspective is ultimately a call to treat the drug development process itself as a learning system: mechanistically grounded, continuously updated by human data, and designed so that every patient treated generates knowledge that improves the next therapy. If the last several decades have proven that more data alone does not produce more approved drugs, the coming test is whether closing the loop between data and decisions can.

Subject of Research: Data-centric feedback loops that use single-cell and clinical data to iteratively improve immunotherapy development

Article Title: Data-centric feedback loops for next-generation immunotherapy development

Article References: Data-centric feedback loops for next-generation immunotherapy development. (n.d.). https://doi.org/10.1038/s41551-026-01785-6

Image Credits: AI Generated

DOI: 10.1038/s41551-026-01785-6

Keywords: immunotherapy, drug development, single-cell genomics, data-centric frameworks, feedback loops, reinforcement learning, virtual cell models, CAR T cells, checkpoint inhibitors, patient stratification, CRISPR screening, Nature Biomedical Engineering