Genialis, a privately held biotechnology artificial intelligence company, announced on October 22, 2026, new research demonstrating that its biology-first molecular models improved cancer treatment response prediction across three small retrospective oncology datasets containing 41 to 66 patients. The findings, scheduled for presentation at the European Society for Medical Oncology Congress 2026, showed that models built using biologically structured gene-expression features outperformed conventional gene-expression models under held-out testing. The company also described a separate retrospective analysis of four tumour RNA sequencing samples collected over 25 months from a patient with osteosarcoma, demonstrating an approach to comparing molecular changes within an individual cancer case.
The research addresses a persistent challenge in precision oncology: developing reliable predictive biomarkers when the number of available patient samples is small relative to the enormous complexity of molecular data. Early-stage cancer trials frequently enrol dozens rather than thousands of patients, yet each tumour sample can generate measurements covering approximately 20,000 genes. Conventional machine learning approaches may struggle to distinguish genuine treatment-related biological patterns from statistical noise under these conditions.
Genialis is attempting to address this problem through a molecular foundation model trained on large collections of cancer transcriptomic data. Rather than asking a predictive algorithm to learn directly from thousands of individual gene measurements within a small trial, the company’s approach first converts the measurements into representations of underlying biological processes. These features are intended to make the resulting predictions more stable and biologically interpretable.
The findings could have implications for pharmaceutical research, particularly when developers need to identify responsive patient populations before investing in larger clinical trials. However, the evidence remains retrospective, and the reported performance does not establish that the models improve actual treatment decisions or patient outcomes. Prospective validation and evidence of clinical utility would be required before the approach could support routine treatment selection.
How does the Genialis Supermodel reduce the complexity of RNA sequencing data?
The Genialis Supermodel is a large molecular model developed using hundreds of thousands of transcriptomic samples. Transcriptomic data describe the activity of genes within biological material, providing information about the molecular processes occurring in tumour tissue. RNA sequencing is commonly used to generate these measurements because it can capture gene-expression patterns across thousands of genes simultaneously.
A conventional predictive model may attempt to identify relationships between individual gene-expression measurements and treatment outcomes. This becomes statistically challenging when the number of molecular variables greatly exceeds the number of patient samples. A model trained on 50 patients and 20,000 genes can identify relationships that appear convincing within the training data but fail when applied to new patients.
This problem is associated with high-dimensional data and the risk of overfitting. When too many variables are available relative to the number of observations, an algorithm can learn patterns that reflect accidental characteristics of the dataset rather than reproducible biology. Standard statistical controls can reduce this risk, but the underlying data limitations remain important.
Genialis approaches the problem by translating gene-expression measurements into what it calls biomodules. These are lower-dimensional representations of biological processes, including oncogenic signalling, immune activation and DNA damage responses. The idea is to provide predictive models with a more structured description of tumour biology before using limited clinical data to estimate treatment response.
The resulting models are designed to learn relationships between these biological features and observed outcomes. Reducing the number of variables may help limit the potential for overfitting, particularly if the biomodules retain clinically relevant information. This strategy also offers an interpretability advantage because model predictions can potentially be associated with identifiable biological pathways.
However, reducing data complexity is not automatically beneficial in every setting. A simplified representation may discard information that is useful for predicting particular treatment responses. The quality of the biomodules and the biological relationships they capture must therefore be assessed independently rather than presumed to be superior because the model has fewer input features.
The Supermodel’s training on a large transcriptomic collection provides background molecular knowledge, but the relevance of that knowledge to a particular drug remains dependent on validation. Differences in tumour type, specimen processing and treatment mechanism can affect predictive performance. A useful model must therefore demonstrate that its learned biological features generalise beyond the datasets used to develop them.
What did the ESMO 2026 research demonstrate across melanoma, gastric cancer and lung cancer?
The ESMO research evaluated three separate clinical datasets containing pretreatment RNA sequencing information and treatment outcomes. The cohorts included 41 patients with melanoma treated with anti-PD-1 immunotherapy, 45 patients with gastric cancer treated with pembrolizumab and 66 patients with KRAS G12C-mutated non-small cell lung cancer treated with sotorasib. Each dataset represents a different cancer setting and treatment mechanism.
The analysis compared models built using Genialis biomodules with models developed from conventional gene-expression measurements. According to the company’s announcement, the biomodule-based approaches achieved stronger predictive performance across all three cohorts under held-out testing. The approach was intended to determine whether structured biological representations could make better use of limited training data.
The strongest numerical example disclosed involved the non-small cell lung cancer cohort. The biomodule model achieved an area under the receiver operating characteristic curve, or AUROC, of 0.80 for predicting response to sotorasib. The best conventional gene-expression model achieved an AUROC of 0.60 in the same comparison.
AUROC measures a model’s ability to distinguish between patients belonging to different outcome categories across possible prediction thresholds. A value of 0.50 indicates discrimination comparable to chance, while 1.00 represents perfect discrimination within the assessed data. An AUROC of 0.80 therefore indicates stronger discrimination than 0.60, but it does not mean that the model correctly predicts outcomes for exactly 80% of patients.
This distinction is important because diagnostic and predictive performance depends on more than AUROC. Clinically useful models must operate at appropriate sensitivity and specificity thresholds and produce results that can be interpreted within the relevant treatment setting. The proportion of patients who respond to treatment also influences measures such as positive predictive value.
The announcement did not provide equivalent numerical AUROC comparisons for the melanoma and gastric cancer cohorts. Although Genialis reported that biomodule models outperformed gene-expression comparators in each dataset, the magnitude of improvement across those two cohorts cannot be quantified from the disclosed summary. The full poster may provide additional information about confidence intervals, model training and evaluation procedures.
The findings therefore support the research hypothesis that biological feature representations can improve prediction under small-data conditions. They do not yet establish a broadly validated predictive test suitable for use across different cancers or medicines.
Why are the sotorasib results important for biomarker development in KRAS G12C lung cancer?
Sotorasib is a targeted therapy developed to inhibit the KRAS G12C mutant protein, an oncogenic driver found in a subset of non-small cell lung cancers. The presence of the KRAS G12C mutation provides a molecular basis for treatment selection, but not every patient with the mutation experiences the same treatment response. Differences in tumour biology and resistance mechanisms can influence clinical outcomes.
The Genialis analysis investigated whether additional transcriptomic information could identify biological characteristics associated with response. According to the company, patients predicted to respond to sotorasib remained on treatment for a median of 337 days, compared with 178 days among predicted nonresponders. This difference provides a clinically relevant association within the analysed dataset because treatment duration can reflect several aspects of disease control and tolerability.
However, time on treatment is not equivalent to progression-free survival or overall survival. Patients may stop or continue therapy for reasons including adverse effects, clinician decisions, access to treatment and personal circumstances. The reported difference therefore supports further investigation but does not independently demonstrate that the predictive model identifies patients who live longer.
The analysis also generated biological explanations for the predictions. Genialis reported that predicted responders showed molecular patterns associated with dependence on MAPK signalling, while predicted nonresponders displayed PI3K-associated bypass programmes. These pathways are relevant because cancer cells may use alternative signalling mechanisms to sustain growth when a particular oncogenic target is inhibited.
The association between predicted response and underlying signalling provides a possible mechanistic explanation for the model’s performance. Rather than producing only a numerical probability, the biomodule approach attempts to identify biological characteristics that may influence treatment sensitivity. Such information could be useful for developing hypotheses about resistance and potential combination therapies.
Nevertheless, the biological interpretation remains dependent on the validity of the underlying measurements and modelling assumptions. A pathway signature associated with poor response does not establish that the pathway caused resistance in each patient. Experimental and prospective clinical research would be necessary to determine whether modifying the identified pathway changes treatment outcomes.
The relatively small cohort size also limits the precision of the performance estimate. Even when held-out testing is used, results from 66 patients may be sensitive to differences in patient selection, response definitions and sample processing. Independent validation in a substantially larger population would be important before relying on the model for clinical decision-making.
Can biology-first AI improve patient selection in early-stage oncology clinical trials?
Early oncology trials frequently face difficult decisions about which patients should receive an investigational therapy. A medicine targeting a particular mutation or pathway may produce responses only in a subset of patients whose tumours share the relevant biological characteristics. Identifying that subset can influence trial design, dose expansion and the allocation of research resources.
Conventional biomarker discovery often relies on individual gene mutations, protein expression or molecular signatures identified through exploratory analyses. These approaches can be valuable, but complex treatment responses may depend on interactions across several biological pathways. A model that captures coordinated molecular activity could potentially identify patterns not apparent from a single biomarker.
Genialis’ approach may be particularly relevant when a development programme has sufficient samples for exploratory modelling but too few for conventional high-dimensional machine learning. Using biological representations learned from larger datasets could reduce the amount of trial-specific information needed to build a preliminary predictive model. The resulting hypotheses might then inform patient-selection strategies for subsequent studies.
However, the timing of model development is important. A biomarker discovered retrospectively after trial outcomes are known may appear promising but fail when evaluated prospectively. Clinical developers must therefore distinguish between exploratory discovery and a model fixed in advance for validation.
For a predictive biomarker to influence treatment decisions reliably, it must demonstrate that it can identify patients likely to benefit from a particular intervention. This is different from a prognostic biomarker, which may predict outcomes regardless of treatment. Establishing predictive value can require treatment-specific comparisons and evidence that the biomarker meaningfully modifies expected treatment benefit.
The ESMO results suggest that biomodule-based models could support the discovery stage of such research. They do not establish that the models are already suitable for selecting patients in registration trials. Additional work would be needed to determine how models perform across independent populations, analytical platforms and clinical environments.
The approach also introduces a potential commercial opportunity for companies supporting pharmaceutical research. Drug developers may use validated molecular modelling tools to prioritise combinations, investigate resistance and design biomarker-enriched studies. However, the October announcement did not disclose commercial contracts, new revenue commitments or prospective regulatory qualification of the reported models.
What does the 25-month osteosarcoma case study reveal about longitudinal tumour analysis?
Alongside the cohort-level results, Genialis described a retrospective analysis involving GitLab co-founder Sid Sijbrandij, who was diagnosed with osteosarcoma in November 2022. Osteosarcoma is a rare bone cancer, and the publicly disclosed case involves extensive efforts to characterise the disease through clinical and molecular testing. The research team associated with the case has collected and shared more than 30 terabytes of clinical and molecular information.
Genialis analysed four tumour RNA sequencing libraries obtained over a period of 25 months. The samples originated from different specimens and were processed by three clinical laboratories. This created an opportunity to investigate whether the Supermodel could provide comparable biological interpretations despite differences in sample collection and laboratory processing.
The company reported that its biomodule representations enabled consistent comparisons of tumour biology across the four samples, even where the raw gene-expression measurements were not directly comparable. The analysis allowed researchers to examine how molecular pathway activity changed over time. It also provided a basis for revisiting therapeutic hypotheses as new tumour information became available.
This type of longitudinal analysis can be valuable because cancer biology is not static. Tumours may change following treatment, acquire resistance-associated characteristics or develop different cellular populations over time. Molecular information obtained from a single biopsy may therefore become less representative as the disease evolves.
A model capable of integrating sequential tumour samples could help researchers investigate these changes and identify new biological questions. However, the value of this approach depends on whether the model’s measurements reflect real disease evolution rather than differences caused by specimen handling, tumour composition or analytical methods. Longitudinal consistency must be established through appropriate technical and biological validation.
The case study should also be distinguished from evidence of treatment efficacy. Genialis did not report a controlled clinical trial showing that Supermodel-guided decisions improved the patient’s survival or response to therapy. The analysis demonstrates a method for interpreting molecular information within one individual case, not proof that a particular experimental treatment worked.
Because the study involves a single patient, its findings cannot establish how reliably the approach would perform across the broader osteosarcoma population. The case nevertheless illustrates a relevant research application: tracking tumour biology when multiple samples become available during an individual disease course.
Why does analysing one patient’s tumour differ from developing a clinically validated AI diagnostic?
The osteosarcoma case illustrates how individualised molecular interpretation could support research into rare and difficult-to-treat cancers. However, generating a biological hypothesis is different from producing a clinically validated diagnostic result. A research model may identify pathways associated with tumour behaviour without establishing that its predictions are accurate enough to guide treatment.
Clinical validation requires evidence that a tool performs reliably in the population and setting for which it is intended. Such evidence may involve independent datasets, prospective evaluation and assessment of clinically relevant errors. Researchers must also demonstrate that results remain consistent across laboratory methods and patient populations.
Clinical utility introduces a further requirement. Even if a model predicts treatment response accurately, it must be shown that using the prediction improves decision-making or patient outcomes compared with existing approaches. A model that generates interesting molecular explanations but does not alter clinical management may still be valuable for research, but its clinical role would remain limited.
The distinction is especially consequential when considering single-patient applications. A model trained using information from many cancers may identify molecular patterns within an individual tumour, but the absence of a comparable patient population makes validation more difficult. Researchers must avoid interpreting model-generated treatment hypotheses as confirmed evidence that a particular intervention will benefit the individual.
For Genialis, the immediate value of the case study lies in demonstrating an approach to integrating molecular observations across time and laboratory conditions. It supports further methodological development but does not establish a validated clinical decision-support system for osteosarcoma. The company’s ESMO findings should therefore be interpreted as evidence relevant to research applications rather than routine treatment recommendations.
What challenges remain before Genialis can expand biology-first AI across cancer research?
One challenge is the diversity of molecular datasets used in oncology. RNA sequencing results can vary according to tissue handling, sequencing technology, laboratory methods and data processing. Differences in tumour purity and cellular composition may also influence gene-expression patterns, making reproducibility an important consideration.
A molecular foundation model can potentially reduce some of this variability by representing biological processes rather than individual measurements. However, that capability must be assessed across datasets collected in different institutions and clinical settings. Performance demonstrated within selected retrospective cohorts may not generalise to substantially different populations.
The representativeness of training data is another consideration. Cancer biology varies with genetic background, tumour subtype, disease stage and treatment exposure. If some populations or clinical conditions are underrepresented, predictive performance may differ when a model is applied more broadly.
The interpretability of biomodules offers an advantage over purely opaque algorithms, but interpretability does not automatically establish biological correctness. A model may associate a particular pathway signature with response without fully capturing the mechanism responsible for that association. Independent experimental evidence can help determine whether model-generated explanations correspond to causal biological processes.
Prospective validation will be particularly important if Genialis intends to support patient selection in clinical trials. Developers would need to specify the model, thresholds and intended decisions before evaluating performance in new patients. This would reduce the risk that favourable retrospective results reflect repeated modelling choices or characteristics unique to the original dataset.
Regulatory requirements would also depend on the eventual intended use. Software used exclusively for exploratory pharmaceutical research may face different obligations from a clinical diagnostic or decision-support product intended to influence patient treatment. The October announcement does not establish regulatory authorisation for a clinical diagnostic application.
How could biology-first AI influence pharmaceutical partnerships and biomarker development economics?
The pharmaceutical industry increasingly relies on biomarker research to identify patients likely to benefit from targeted therapies and immunotherapies. Better patient selection can potentially improve the efficiency of clinical development, particularly when treatment response varies substantially across molecular subgroups. However, the financial value of a predictive model depends on whether it improves decisions beyond existing analytical methods.
Genialis positions its Supermodel as a platform for therapeutic intelligence rather than a single disease-specific diagnostic. This approach could allow the company to support different pharmaceutical programmes using a common molecular modelling foundation. Its potential applications include biomarker discovery, treatment-response modelling and investigation of resistance pathways.
A platform model may offer commercial scalability because the underlying representation can potentially be adapted to several therapeutic settings. Nevertheless, each application requires evidence that the model performs appropriately for the specific cancer type, treatment and patient population. A favourable result in sotorasib-treated lung cancer cannot be assumed to predict similar performance in immunotherapy-treated melanoma.
The company’s broader research activities include collaborations involving pharmaceutical and diagnostic technologies, as well as work applying molecular modelling to drug response prediction. These programmes provide context for its development strategy but should not be interpreted as evidence that the ESMO findings have already generated new commercial revenue. The October announcement did not disclose additional contract values or confirmed deployment commitments.
For pharmaceutical sponsors, the most immediate opportunity may involve improving decisions during early clinical development. If biology-first models can reliably generate useful predictions from smaller cohorts, developers might identify promising biomarkers before committing to expensive late-stage trials. However, the potential reduction in development cost remains hypothetical until demonstrated through actual project outcomes.
The company will also need to establish how its approach compares with increasingly sophisticated conventional and foundation-model-based methods. The relevant question is not whether biomodules outperform a weak baseline, but whether they provide a measurable advantage over the strongest practical alternatives available to researchers. Future validation should therefore include carefully selected benchmarks and independent testing.
What evidence will determine whether Genialis’ Supermodel becomes useful in routine precision oncology?
The ESMO 2026 findings provide an early indication that biology-first representations can improve treatment-response modelling when clinical datasets are small. The reported AUROC improvement in KRAS G12C-mutated lung cancer is a concrete result, while the melanoma and gastric cancer cohorts provide additional evidence of performance across different treatment settings. Nevertheless, all three datasets remain limited in size.
The next important requirement is independent validation using patient populations not involved in model development or earlier optimisation. Such studies would help determine whether the reported predictive performance persists across different hospitals, sequencing procedures and patient characteristics. They would also provide more reliable estimates of sensitivity, specificity and other clinically relevant performance measures.
Prospective studies would add another layer of evidence by evaluating models on newly enrolled patients. This would reduce reliance on retrospective datasets and provide an opportunity to assess how predictions perform under conditions closer to intended use. Depending on the application, researchers could also investigate whether model-informed trial decisions improve enrolment efficiency or clinical outcomes.
The longitudinal osteosarcoma analysis presents a different research opportunity. Future work could examine whether biomodule representations consistently track meaningful tumour changes and whether those measurements correspond to independently established biological or clinical events. Demonstrating reproducibility across additional patients would be necessary before drawing broader conclusions.
For Genialis, the immediate development opportunity is to convert retrospective modelling performance into robust, independently validated research tools. Its emphasis on biological structure may help address limitations associated with high-dimensional molecular data, particularly when patient numbers are constrained. However, model accuracy, biological interpretation and clinical utility are separate requirements that must each be demonstrated.
The October findings suggest that artificial intelligence can potentially extract more useful information from limited oncology datasets when modelling is informed by established biological relationships. The approach may prove valuable in biomarker discovery and the investigation of treatment resistance, but it has not yet demonstrated that AI-guided decisions improve outcomes for cancer patients. The decisive milestone will be whether the reported advantages persist in independent prospective evaluations and eventually translate into better-supported treatment development decisions.
