acceptodds
Under review as a conference paper at ICLR 2027

LEXA: Learning a lexicon of gene-programme states from single-cell transcriptomes

Abstract

Single-cell transcriptomic modeling faces a persistent trade-off between preserving complex cellular variation and providing biologically interpretable representations. Continuous latent models capture rich cellular information but are difficult to interpret at the level of individual dimensions, whereas factorization- and topic-based gene-programme methods provide interpretable components but retain less fine-grained variation. We propose LEXA, which couples programme-factorized vector quantization with scaffold-guided gene responsibility and gene-resolved additive decoding to organize cellular variation into programme-specific axes, where each cell selects one of multiple discrete states along each axis, forming a compositional lexicon with explicit gene-resolved molecular effects. We pretrained LEXA on CELLxGENE and evaluated on multiple held-out datasets against representative continuous, discrete, and gene-programme baselines, and found that it preserves substantially more cellular information than conventional gene-programme methods while yielding state-level molecular interpretations that are biologically coherent and reproducible. In an external COVID-19 cohort, both the atlas-pretrained and cohort-specific LEXA settings revealed severity-associated programme variation across immune lineages beyond conventional cell-subtype descriptions. These results demonstrate LEXA as a framework for reconciling information-rich representation with explicit molecular interpretation, turning cellular variation into enumerable and biologically testable molecular states.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
representation learning, single-cell transcriptomics, vector quantization, interpretable machine learning, compositional representations, gene programme modelling

Reviews

Among:
Shape:
1234this paper, Top 32%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.50 over 4 reviews, above 68% of the others.
Reviewer_syfmOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces a method that represents each single-cell transcriptome as one discrete state on each of 32 programme axes. It combines per-axis vector quantization, a cell-type/donor-derived gene-responsibility scaffold, and an additive decoder whose axis-state outputs define signed gene-effect vectors.

Critical strengths
  • The compositional discrete representation is conceptually interesting.
  • The method is described clearly.
Critical weaknesses
  • If I understand it correctly, the scaffold incorporates cell-type labels, so agreement with known programmes does not alone establish de novo discovery. Maybe it would be better to quantify how much the biological-coherence results depend on the scaffold.
  • Baseline factors are selected within donor-held-out folds, to me it is unclear when LEXA’s Axis 10 and codes C2/C12 were chosen. Can the authors clarify the selection procedure and repeat the comparison with fold-nested selection for every method.
  • Severity-associated states are identified and assessed in COVID-A, without independent-cohort replication of their association directions. Applying the frozen atlas model demonstrates representation transfer, but does not validate those clinical associations.
Questions for authors
  • Were the underlying models fitted without the held-out donors, and were LEXA’s axis, codes, and gene lists selected within each fold, as the comparator factors were? If not, how does a fully nested comparison affect the concordance ranking?
  • Can you repeat GO coverage, fragmentation, monocyte concordance, and state interpretation, not merely macro-F1 and reproducibility, for the model without the cell-type-derived scaffold?
  • Do the COVID-A severity-associated states replicate in COVID-B using a frozen lexicon, pre-specified state mappings, effect directions, and one correction family?
  • What was the prespecified primary family of COVID-A severity tests, and do the principal associations survive correction across the cell classes and representation families discussed as one result?
  • How many independent donors enter each regression, and are results robust to estimated dispersion, HC3 or bootstrap uncertainty, nonlinear severity effects, and site-stratified analysis?
  • What identifies an axis-state gene contribution uniquely, given that reconstruction can redistribute signal among soft gene responsibilities, codebooks, FiLM parameters, and the shared decoder?
  • What are the empirical code entropies, state occupancies, dead-code rates, and stability of assignments across seeds and cohorts, rather than the nominal 128-bit capacity based on all states being usable?
7 comments
AuthorsOpenReview ↗

Authors — Response (Part 1/5)

We sincerely thank the reviewer for the careful and insightful assessment. To present our additional analyses clearly within OpenReview’s character limit, we have divided our response into five consecutive comments (Parts 1–5). These address scaffold dependence, donor-level selection and independent validation. We evaluate an annotation-free scaffold, a fully nested monocyte analysis, a pooled COVID-A correction family and cross-cohort validation.

Response to W1 & Q2: Dependence of biological interpretability on the scaffold

We appreciate the reviewer's concern that the cell-type-derived scaffold may influence the biological interpretation of the learned programme states. We therefore evaluated the dependence on scaffold construction using programme-level and molecular endpoints, rather than cell-type classification alone.

We first clarify that the scaffold is used to initialize gene responsibilities, which are subsequently updated during training rather than held fixed. In addition to the original OneK1K-derived scaffold, we constructed two alternatives: (1) an atlas-derived scaffold based on CELLxGENE and (2) an annotation-free scaffold in which cell-type contrasts are replaced with unsupervised KMeans strata. We trained LEXA on OneK1K with these different scaffold initializations and evaluated fine-cell-type information retention, held-out age-associated molecular prediction, and GO programme organization.

As shown in Table 1, the annotation-free KMeans scaffold achieves comparable performance to the original model, with fine-type MLP scores of 0.614 versus 0.620, age-associated-expression Pearson correlations of 0.698 versus 0.701, and identical GO programme coverage of 0.750. Notably, its fragmentation is lower (0.156 versus 0.300). The atlas-derived scaffold similarly maintains or improves performance across these evaluations.

These results suggest that LEXA's ability to recover biologically coherent programmes is not solely driven by the original cell-type-derived scaffold. Even when cell-type annotations are excluded from scaffold construction, the model retains comparable programme coverage and held-out molecular prediction performance.

Table 1. OneK1K cohort-fitted LEXA under alternative scaffold initializations.

ConfigurationFine-type MLP ↑Age-associated-expression Pearson ↑GO coverage ↑fragmentation ↓
Original LEXA (OneK1K-derived scaffold)0.6200.7010.7500.300
Atlas-derived scaffold0.6440.7040.8540.281
KMeans scaffold0.6140.6980.7500.156
AuthorsOpenReview ↗

Authors — Response (Part 2/5)

Response to W2 & Q1: Fairness of monocyte comparisons and fully nested selection

We thank the reviewer for raising this important concern about potential selection bias. We clarify that the illustrative states shown in Figure 4B and the quantitative donor-held-out benchmark serve different purposes and follow different selection procedures.

Illustrative states versus quantitative evaluation. The A10 and related states from the reference model (seed 6101) are used as illustrative examples in Figure 4B. In contrast, the quantitative benchmark does not rely on these predefined states. Instead, each method independently selects its native programme objects within each outer donor fold, using only the remaining four folds for selection.

All representation models were fitted on 56,333 OneK1K training cells, with no donor overlap between model training and the donor-disjoint test partition used for the monocyte evaluation. Within each outer fold, candidate objects were ranked by their standardized CD16-versus-CD14 activity difference using the four discovery folds. For LEXA, candidates are its native axis-state objects; for continuous programme models, candidates are their native factors or topics. The most CD16-associated and CD14-associated objects were selected separately before evaluating their molecular concordance in the held-out donor fold.

As shown in Table 2, the selected LEXA objects are highly consistent across the five donor folds within individual model fits. For example, the reference seed 6101 consistently selects A10 for CD16 and A10 for CD14 across all five folds. Importantly, all selections are performed without access to the held-out fold.

Table 2. Native LEXA objects selected within the five donor folds.

FitCD16-associated native objectCD14-associated native object
Reference seed 6101Axis 10 C2 (5/5)Axis 10 C12 (5/5)
Seed 6102Axis 21 C3 (5/5)Axis 16 C5 (5/5)
Seed 6103Axis 21 C9 (5/5)Axis 26 C6 (5/5)
Atlas-mask scaffoldAxis 23 C10 (3/5); Axis 26 C9 (2/5)Axis 19 C15 (5/5)
KMeans scaffoldAxis 23 C14 (5/5)Axis 23 C0 (5/5)

Fully nested molecular concordance. We then evaluated all methods on the same complete 10,310-gene space, avoiding differences in gene-set selection across methods. As shown in Table 3, LEXA achieves the highest Pearson correlation for CD16 monocytes (0.238), compared with 0.142 for cNMF, 0.162 for LDA, and 0.180 for Spectra. For CD14 monocytes, LDA achieves the highest correlation (0.281), while LEXA achieves 0.200.

Table 3. Fully nested monocyte Pearson correlation over the common 10,310-gene space.

Native objectCD16 PearsonCD14 Pearson
LEXA states0.2380.200
cNMF factor0.1420.236
LDA topic0.1620.281
Spectra factor0.1800.135

Response to W3 & Q3, Q4, Q5: Independent cross-cohort validation of COVID-19 severity associations

We thank the reviewer for raising these important concerns regarding the reproducibility and statistical reliability of our COVID-19 severity associations. We address the three questions separately below.

(1) Multiple-testing correction across severity associations (Q4)

We acknowledge that the original analysis applied BH correction separately within cell-class-by-representation families. To address Q4, we repeated the analysis using a single pooled BH correction across all 5,452 eligible linear severity tests, including 39 cell-subtype tests, 1,766 atlas-state tests, and 3,647 cohort-fitted-state tests.

As shown in Table 4, 2 subtype associations, 133 atlas-state associations, and 212 cohort-fitted-state associations remain significant after this correction. These results demonstrate that substantial disease-associated molecular variation remains detectable even after controlling the false discovery rate across all cell classes and representation families in one testing family.

Table 4. COVID-A severity associations after one pooled BH correction across all 5,452 linear tests.

Representation familyTestsSignificant after pooled BH
Subtype392
Atlas LEXA1,766133
Cohort-fitted LEXA3,647212
AuthorsOpenReview ↗

Authors — Response (Part 3/5)

(2) Independent donors and robustness of statistical inference (Q5)

To address Q5, we clarify that the donor, rather than the individual cell, is the independent statistical unit. We retained only the earliest available sample per individual, resulting in 83–86 independent donors per major cell class, as detailed in Table 5.

Table 5. Independent COVID-A donors entering each major-cell-class regression.

Cell classIndependent COVID-A donors
B cell83
CD4 T85
CD8 T85
Dendritic cell86
Monocyte86
NK / ILC86
Plasma / plasmablast84
Platelet / MK86
Unconventional T85

We further examined the robustness of the 20 fixed CD4-associated states shown in Figure 5B. All 20 associations retain nominal significance (p < 0.05) under HC3 robust covariance estimation and jointly estimated negative-binomial dispersion. All 20 donor-bootstrap percentile confidence intervals exclude zero across 500 resamples. In addition, site-stratified analyses yield coefficients consistent in direction with the pooled estimates for all 20 states in both Cambridge and Newcastle (Table 6). These analyses demonstrate that the principal CD4 associations are robust to alternative uncertainty estimates, dispersion assumptions, and recruitment-site stratification.

Table 6. Robustness checks for the 20 fixed Figure 5B CD4 associations.

CheckResult
HC3 covariance, fixed Figure 5B CD4 states20/20 associations retain p<0.05
Jointly estimated negative-binomial dispersion20/20 associations retain p<0.05
500 donor bootstrap resamples20/20 percentile intervals exclude zero
Site-stratified models20/20 coefficients retain the pooled direction in Cambridge and Newcastle

(3) Nonlinear severity effects (Q5)

The reviewer also asks whether the severity associations are robust to nonlinear effects. We therefore extended the regression models by adding a centred quadratic severity term and applied one BH correction across all 5,452 quadratic-term tests.

Five quadratic associations remain significant after correction (q < 0.05), including two atlas states and three cohort-fitted states (Table 7). These associations exhibit non-monotonic patterns across the five severity levels, including inverted-U-shaped and U-shaped relationships. These results extend the original linear severity analysis by identifying additional molecular configurations associated with nonlinear disease responses.

Table 7. Quadratic severity associations significant after one BH correction across all 5,452 tests.

Cell classRepresentationStateQuadratic βRaw quadratic pGlobal quadratic BH qFitted pattern
Unconventional TAtlas LEXAA6:C11−0.38997.685e−070.001397Inverted U
Dendritic cellAtlas LEXAA6:C80.32815.203e−060.007092U-shaped
MonocyteCohort-fitted LEXAA23:C11−0.26305.550e−093.026e−05Inverted U
MonocyteCohort-fitted LEXAA30:C13−0.23382.086e−075.686e−04Inverted U
Dendritic cellCohort-fitted LEXAA2:C1−0.32894.335e−050.04727Inverted U

(4) Independent cross-cohort validation of severity-associated molecular signatures (W3 & Q3)

To address W3 and Q3, we performed independent cross-cohort validation of the molecular signatures associated with disease severity, rather than relying solely on the transferability of the pretrained representation.

In COVID-A, we identified significant severity-associated programmes using donor-level negative-binomial regression, adjusting for age, sex, site, and days from symptom onset, with a cell-count offset and the pooled multiple-testing correction described above. The resulting significant programmes were converted into direction-oriented molecular signatures using their top 50 and bottom 50 genes. These gene sets and their association directions were frozen before evaluation in COVID-B.

Importantly, COVID-A and COVID-B have different clinical designs. COVID-A supports a five-level severity analysis, whereas COVID-B contains only 14 acute cases and lacks moderate and critical acute donors. Therefore, the original five-level severity coefficient cannot be directly replicated in COVID-B. Instead, we tested whether the frozen molecular signatures retained their prespecified disease-association directions under the available acute COVID-19 versus healthy comparison.

Reviewer_3qsYOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This manuscript presents LEXA, a framework that represents single-cell transcriptomes through combinations of discrete gene-programme states with explicit gene-level contributions. The model is clearly described, and the idea of representing alternative molecular configurations within a programme is interesting. However, I have several concerns regarding the novelty, baseline comparisons, and evidence for biological utility.

Critical strengths

See above

Critical weaknesses
  1. The methodological novelty is insufficiently established. Although the architecture contains nonlinear components, once the states are selected, reconstruction is equivalent to a linear decoder based on concatenated one-hot encodings. The gene-level contributions therefore follow directly from a standard additive dictionary representation and are not novel. The potentially interesting contribution is how this dictionary is learned as programme-specific states. However, the manuscript does not convincingly establish what this adds beyond existing programme models, straightforward discretisation of their activities, or capacity-matched additive dictionary models (see also point 2).

  2. Important baselines are missing. In particular, expiMap [1] already combines nonlinear encoding with an interpretable linear decoder, while Tripso [2] directly addresses expressive, multidimensional gene-programme representations. A prior-informed sparse factor model, such as muvi [3], would also be relevant. These comparisons are needed to evaluate the trade-off between information retention and interpretability in the context of SOTA approaches. In addition, existing comparisons are not fair as is and require closer matching of training data, gene inputs, and model capacity: 32 LEXA axes containing 16 state profiles each are substantially more flexible than 32 conventional factors.

  3. The evaluation is too narrow for the central claims. The main representation benchmark has only nine major immune-cell classes. This provides limited evidence that LEXA preserves fine-grained biological variation, particularly because major cell-type annotations contribute to scaffold construction, so indirectly leaking into the model. Although the manuscript also examines GO coherence and monocyte expression contrasts, these analyses also have limitations: First, a lower fragmentation reflects (at least in parts) grouping multiple state profiles under one parent axis, and second, expression concordance is evaluated on different gene sets for each method making meaningful comparisons hard. Instead, I wold expect more complex large-scale perturbation-response benchmarks (like [4] - this is a must), recovery of withheld programmes, and evaluations within cell types.

  4. The COVID-19 application does not demonstrate biological insights that require LEXA. Showing severity-associated variation beyond conventional subtype proportions is useful, but existing bioinformatics method are specifically designed to identify variation within annotated populations. The reported findings largely concern established processes, including ER stress, cell-cycle changes, mitochondrial metabolism, and proteostasis. The relevant question is whether LEXA reveals a reproducible disease association that competing programme methods and conventional differential-expression analyses fail to recover.

  5. Robustness to imperfect prior information is insufficiently evaluated. Here, the relevant prior is primarily the scaffold derived from cell-type annotations and donor-level expression variation. Learnable gene responsibilities do not establish that incorrect annotations or misleading scaffold assignments can be corrected. Removing the scaffold entirely addresses a different question. I would like to see systematic annotation corruption, coarsening of cell labels/noise in cell type annotation (which is a common issue), and omission or corruption of scaffold memberships, assessing programme recovery rather than cell-type classification alone.

Refs:

[1] Lotfollahi, M., Rybakov, S., Hrovatin, K., Hediyeh-Zadeh, S., Talavera-López, C., Misharin, A. V., & Theis, F. J. (2023). Biologically informed deep learning to query gene programs in single-cell atlases. Nature Cell Biology, 25(2), 337-350.

[2] Moullet, M., Isobe, T., Vahidi, A., Leonardi, C., Paulas-Condori, L., Soelistyo, C., ... & Lotfollahi, M. (2026). Self-supervised learning for a gene program-centric view of cell states. bioRxiv, 2026-03.

[3] Qoku, A., & Buettner, F. (2023, April). Encoding domain knowledge in multi-view latent variable models: A bayesian approach with structured sparsity. In International Conference on Artificial Intelligence and Statistics (pp. 11545-11562). PMLR.

[4] Jiang, L., Dalgarno, C., Papalexi, E., Mascio, I., Wessels, H. H., Yun, H., ... & Satija, R. (2025). Systematic reconstruction of molecular pathway signatures using scalable single-cell perturbation screens. Nature Cell Biology, 27(3), 505-517.

Questions for authors

See above

14 comments
AuthorsOpenReview ↗

Authors — Response (Part 1/9)

We thank the reviewer for the careful assessment of the methodological contribution. We address three questions: what the finite programme-state structure contributes, whether the comparison remains meaningful under stronger and capacity-matched baselines, and whether that structure provides useful biological resolution in settings beyond coarse cell-type classification.

Due to OpenReview's character limit, we have divided our response into nine consecutive comments (Parts 1–9).

Response to W1: Methodological novelty and the value of learned discrete programme states

We thank the reviewer for highlighting the distinction between a standard additive dictionary and the programme-specific discrete states learned by LEXA. We agree that the additive decoder itself is not the primary methodological novelty. Rather, the main advantage we intend for LEXA's finite representation is that it provides a complementary view in which cellular variation is organized into combinations of discrete programme states that can be directly used for biological analysis. The axes provide a compact overview of the programmes active in a cell, while each axis is resolved into alternative molecular configurations of that programme.

This finite organization is useful because biological regulatory systems are often naturally interpreted through recurring configurations and transitions among them, and may in some settings exhibit quasi-multistable behaviors (some of which biologists may term as cell types or cell states). Making these configurations enumerable provides a direct way to inspect the structure of a programme: we can ask where the same state occurs, how its occupancy changes across populations or conditions, and what gene-expression effect accompanies that change. This gives a compact representation of programme structure while retaining a direct molecular interpretation.

We acknowledge that the original experiments did not sufficiently demonstrate the biological utility of this representation. We therefore added three complementary analyses.

(1) Biological resolution of perturbation responses. In the independent IFNγ Perturb-seq dataset from Jiang et al. (Nature Cell Biology, 2025), LEXA recovers published perturbation-response programmes and further reveals that different gene knockouts induce distinct, reproducible state-transition patterns within the same programme axis. This demonstrates how LEXA provides additional biological resolution beyond identifying responsive programmes alone. Further details and comparisons with other programme models are provided in Response to W3.

(2) Cross-dataset reproducibility and biological transferability. In the extended COVID analysis, independently trained COVID-A and COVID-B models exhibit similar state-frequency distributions after molecular alignment. Furthermore, molecular signatures derived from severity-associated LEXA states in COVID-A retain significant biological associations when transferred without modification to COVID-B. In total, 23 LEXA signatures are validated across six cell classes, compared with at most two for cNMF and none for LDA under the same validation criteria. These results support the reproducibility and biological relevance of the learned programme states across independent datasets. Further details are provided in Response to W4.

(3) Comparison with post-hoc discretization. To examine whether discretizing continuous representations after training provides comparable information retention, we performed a post-hoc discretization experiment on OneK1K. For each continuous coordinate, 15 quantile thresholds were estimated using training cells only and then applied to donor-disjoint test cells, producing 16-state one-hot representations.

As shown in Table 1, post-hoc discretization preserves substantial cell-type information, although its effects vary across models. Importantly, LEXA also retains competitive information despite its explicitly discrete representation, achieving a major-cell-type MLP macro-F1 of 0.885 and a fine-cell-type MLP macro-F1 of 0.620. These results indicate that learning a structured discrete representation does not necessarily require a severe loss of cellular information.

AuthorsOpenReview ↗

Authors — Response (Part 2/9)

We emphasize that cell-type information retention alone does not establish the biological value of learned programme states, nor does this experiment exclude the possibility that post-hoc discretization can recover some similar biological structures. The main contribution of LEXA is the joint learning of programme-specific discrete states and their molecular effects. The independent perturbation and COVID analyses provide complementary evidence that this representation supports reproducible, interpretable, and transferable analyses of biological variation beyond conventional programme-level activity measurements.

Table 1. Donor-disjoint OneK1K readout before and after 16-quantile post-hoc discretization.

MethodNative major kNNNative major MLPNative fine kNNNative fine MLP16-quantile major kNN16-quantile major MLP16-quantile fine kNN16-quantile fine MLP
LEXA0.8660.8850.5460.620––––
scVI0.9310.8280.6770.7220.8480.9290.5730.708
expiMap0.6740.5260.3920.4560.5930.6580.2620.408
Tripso0.7860.5740.3790.4040.7830.7660.3810.399
cNMF0.8620.8190.5920.6430.6060.8130.3400.622
LDA0.8680.8850.6220.7110.5930.6570.3970.603
Spectra0.7700.7300.4640.5580.7380.7600.4090.537

Response to W2: Stronger baselines and capacity-matched comparisons

We thank the reviewer for emphasizing the need for stronger baselines and fairer comparisons of representation capacity. To address these concerns, we added modern programme representation methods, matched the input genes and data splits where applicable, and extended the comparisons across a broader range of model capacities.

(1) Comparisons with modern programme representation methods. We evaluated expiMap and Tripso using the same donor-disjoint OneK1K split and the same 10,310 input genes as LEXA. For Tripso, we used its official 256-dimensional Global representation for the representation benchmark and its native Global decoder for the perturbation-response analysis. We also included MuVI, but report it separately because the implementation used here estimates sample-specific factors jointly with test expression and is therefore transductive, unlike the inductive inference used by the other methods.

As shown in Table 2, LEXA retains competitive cellular information compared with these methods. In particular, LEXA achieves a major-cell-type MLP macro-F1 of 0.885, outperforming expiMap (0.526) and Tripso (0.574), while its fine-cell-type MLP score of 0.620 also exceeds both methods (0.456 and 0.404, respectively). We acknowledge that general-purpose latent models such as scVI, as well as the transductive MuVI implementation, achieve stronger performance in some settings, particularly fine-cell-type classification.

Table 2. Same-input donor-disjoint OneK1K representation readout including modern programme baselines.

MethodTest accessMajor kNNMajor MLPFine-type kNNFine-type MLP
LEXAInductive0.8660.8850.5460.620
expiMapInductive0.6740.5260.3920.456
TripsoInductive0.7860.5740.3790.404
MuVITransductive0.9210.8960.7110.771
scVIInductive0.9310.8280.6770.722
cNMFInductive0.8620.8190.5920.643
LDAInductive0.8680.8850.6220.711
SpectraInductive0.7700.7300.4640.558

(2) Comparisons across representation capacities. We agree that comparing LEXA's 32 axes, each containing 16 states, only against 32 continuous factors is insufficient. At the same time, LEXA's 512 possible axis-state entries are not independently activated: each cell selects one state per axis, corresponding to a nominal categorical representation capacity of 32 × log2(16) = 128 bits.

AuthorsOpenReview ↗

Authors — Response (Part 3/9)

To assess whether LEXA's performance simply reflects a larger representation space, we extended the cNMF, LDA, and Spectra comparisons from 32 to 64 and 128 components. Although the number of continuous factors is not directly equivalent to the bit capacity of a discrete representation, this comparison provides a broader assessment across different model capacities.

As shown in Table 3, LEXA remains competitive in cell-type information retention across these settings. Its major-cell-type MLP macro-F1 of 0.885 is comparable to LDA-64 (0.893) and Spectra-128 (0.888), although cNMF-128 achieves a higher score (0.920). These results suggest that LEXA retains substantial cellular information despite its constrained discrete representation.

Table 3. Cell-information readout across comparator capacities.

MethodComponents / statesMajor kNNMajor MLP
LEXA32×16 states (128 bits)0.8660.885
cNMF320.8620.819
cNMF640.8330.836
cNMF1280.8060.920
LDA320.8680.885
LDA640.8820.893
LDA1280.8450.872
Spectra320.7700.730
Spectra640.6930.738
Spectra1280.8480.888

(3) Biological programme organization across capacities. We further compared programme coverage and fragmentation across the same range of model capacities. As shown in Table 4, LEXA achieves a coverage of 0.750 and fragmentation of 0.300. While some continuous models achieve higher coverage, LEXA exhibits lower fragmentation than cNMF and LDA across all tested capacities, and remains competitive with Spectra. These results indicate a favorable coverage–fragmentation trade-off, rather than a uniform advantage on every metric.

Table 4. Programme coverage and fragmentation across comparator capacities.

MethodComponentsProgramme coverage ↑Programme fragmentation ↓
LEXA32 axes × 16 states (128 bits)0.7500.300
cNMF320.8130.620
cNMF640.8130.523
cNMF1280.8540.474
LDA320.7710.766
LDA640.7920.833
LDA1280.8130.690
Spectra320.8130.578
Spectra640.8750.367
Spectra1280.7710.250

Response to W3: Fine-grained biological evaluation and perturbation-response benchmarks

We thank the reviewer for emphasizing the importance of evaluating fine-grained biological information and perturbation responses beyond coarse cell-type classification. We agree that these are more direct tests of whether LEXA captures biologically meaningful variation. We therefore extended our evaluation to include fine-cell-type information retention, an independent Perturb-seq benchmark, and a stricter within-cell-type molecular comparison.

(1) Fine-grained cellular information.

We extended the OneK1K representation benchmark from major PBMC classes to 29 fine cell types, using the same donor-disjoint split and gene inputs across methods. As shown in Table 2 of Response W2, LEXA retains competitive fine-cell-type information despite its discrete programme-state representation, achieving an MLP macro-F1 of 0.620, compared with 0.456 for expiMap and 0.404 for Tripso. While continuous models such as scVI achieve higher classification performance, these results show that LEXA preserves substantial fine-grained cellular information.

We also evaluated the dependence on scaffold construction using alternative initializations, including an atlas-derived scaffold and an annotation-free KMeans scaffold (Table 5). Both preserve comparable fine-cell-type information, age-associated molecular prediction, and programme organization. These results suggest that the learned biological structure is not solely determined by the original cell-type-derived scaffold.

Table 5. OneK1K cohort-fitted LEXA under alternative scaffold initializations.

ConfigurationFine-type MLP ↑Age-associated-expression Pearson ↑GO coverage ↑fragmentation ↓
Original LEXA (OneK1K-derived scaffold)0.6200.7010.7500.300
Atlas-derived scaffold0.6440.7040.8540.281
KMeans scaffold0.6140.6980.7500.156
Reviewer_DNwWOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper introduces a new model architecture for single-cell RNA-seq modeling, as a solution to make latent models more interpretable while maintaining representational power. The model contains three main components : a set of discrete codebooks, an additive decoder, and a loss component that pushes axis to be responsible for the reconstruction of different genes.

One model with 32 axis and 16 states is trained on cellxgene and used in several experiments to support the three main claims : (1) the model encodes cellular information, as assessed by its performance on cell type classification (2) the model is interpretable, as assessed by gene program enrichment analysis and a case study (3) the model’s results are reproducible across random seeds, as assessed by stability metrics.

Overall, the paper is well written and easy to follow, and the motivation is clear. However, the experiments that support the main claims could be stronger.

Critical strengths
  1. The main contribution of the paper is a new architecture with several components : a set of discrete codebooks, an additive decoder, and a loss component based on a scaffold. This architecture is interesting and sound, the discrete aspect is convenient for interpretability purposes.
  2. The experimental setup covers various aspects: cell type information retention, interpretability, stability and a concrete use case. The model is compared against multiple alternatives from the literature. The case study is interesting and shows how re-fitting the model on a specific dataset helps answer targeted question. The evaluation is well structured and covers the right aspects but needs a bit more effort to consolidate the claims.
Critical weaknesses
  1. Performances on cell type classification are strongly below the latent models on a quite simple task (cell type annotation of major cell types in PBMC), and scVI and scArches were trained under less favorable conditions (less genes in input and less training data). The information retention analysis is too weak.
  2. Regarding interpretability, there is no evidence for the hierarchical aspect of axis and states, which is to some extent expected. The enrichment analysis seems permissive (500 genes and q<0.05) and is compared against 32 latents in other models, which is fine if the states are indeed levels within one axis. The “cell-type-specific states within a shared LEXA axis” analysis is conducted on one axis only. Overall, the interpretability results could be strengthened, especially regarding the cost in terms of information retention.
Questions for authors

Clarifications

  1. I understand that the objective is not to beat the other models on downstream performances, but the cost seems quite high, especially since the best models (scVI and scArches-32) were trained under less favorable conditions (less genes in input and less training data). In addition, there is only one task to assess information retention (cell type classification with major classes, in PBMC). Extending this evaluation to other tissues (are there only PBMC samples in the training dataset ?) and other tasks would strengthen the results. The figures with macro-F1 scores are difficult to read, could you provide the tables in appendix ? Could you also report error ranges (standard deviation across seeds or bootstrap confidence interval)?
  2. Why didn’t you use the 4 datasets for the ablations (Figure 3)? Could you include error ranges? It seems that the results are analyzed differently in the different ablations. For example, you say “expanding the codebook to 64 states provided only limited additional benefit”, but that “default radius-normalized LEXA representation […] substantially improved MLP performance in COVID-B”, whereas the gap between k=16 and k=32 with MLP seems as important.
  3. In the gene program enrichment analysis, why did you use the top 500 genes and a threshold of q<0.05? These choices seem permissive, could you indicate how many terms were found per axis with such parameters ? How was the “functional summary” obtained in Table 5 ?
  4. In the analysis of cell-type-specific states within a shared LEXA axis, do the results observed in axis 10 generalize to other axis? An automated analysis for all axis would strengthen the result. Also, the current analysis does not show that states within an axis encode a similar biological function for various cell types. This is to some extent expected given the architecture, and what the “within a shared axis” suggests. Do you have evidence for that? Could you include the description of codes within axis ?
  5. The description of axis 10 in Table 2 is “An extracellular-matrix organization and remodelling programme involving matrix structural components and heparin binding.”, but axis 10 in Section 5 is described as “an immune associated axis linked to leukocyte-mediated cytotoxicity and TNF/NF-κB inflammatory processes”. Is it the same model ?

Suggestions for future work

  1. What is the impact of the encoder architecture ? Does increasing the number of parameters of the encoder increase the performances ? Did you try several architectures ?
14 comments
AuthorsOpenReview ↗

Authors — Response (Part 1/7)

We sincerely thank the reviewer for the detailed and constructive suggestions. Given OpenReview’s character limit, we have divided our response into seven consecutive comments (Parts 1–7). We evaluated both information retention and biological performance using matched gene inputs, 29 fine cell types, all four representation datasets, all 32 axes, and an independent perturbation-response benchmark across six non-PBMC cell lines.

Response to Q1 & W1: Information retention and broader biological evaluation

We thank the reviewer for highlighting the importance of evaluating the information-retention cost of LEXA's discrete programme-state representation. We agree that a fair comparison requires matched input conditions, finer-grained cellular annotations, and biological evaluations beyond PBMC cell-type classification. We therefore extended our experiments in three directions.

(1) Matched-input information retention at major and fine cell-type resolutions.

We retrained scVI using the same 10,310 input genes and donor-disjoint train/test split as LEXA on OneK1K, and expanded the evaluation from major cell classes to 29 fine cell types.

As shown in Table 1, LEXA retains competitive cellular information despite its discrete programme-state representation. For major cell types, LEXA achieves an MLP macro-F1 of 0.885, matching LDA and exceeding scVI (0.828) and the other programme models. For fine cell types, LEXA achieves an MLP macro-F1 of 0.620, comparable to cNMF (0.643), although below scVI (0.722) and LDA (0.711). These results indicate that LEXA preserves substantial cellular information at both resolutions, while we acknowledge that continuous latent models retain an advantage in some classification settings, particularly fine-grained kNN classification. Table 7 additionally reports mean performance and standard deviations across three independent random seeds for the final radius-4 LEXA configuration.

Table 1. Same-input donor-disjoint representation readout on OneK1K.

MethodMajor kNNMajor MLPFine-type kNNFine-type MLP
LEXA0.8660.8850.5460.620
scVI0.9310.8280.6770.722
expiMap0.6740.5260.3920.456
Tripso0.7860.5740.3790.404
cNMF0.8620.8190.5920.643
LDA0.8680.8850.6220.711
Spectra0.7700.7300.4640.558
AuthorsOpenReview ↗

Authors — Response (Part 2/7)

(2) Biological evaluation beyond PBMC cell-type annotation.

We first clarify that LEXA's CELLxGENE training atlas is not restricted to PBMCs. It contains 13,826,509 cells from 210 datasets, selected using tissue-wise round-robin sampling to provide broad tissue coverage.

More importantly, we added an independent biological evaluation using an independent perturbation dataset from Jiang et al. (Nature Cell Biology, 2025). This experiment aims to address two questions: whether LEXA can recover biologically established perturbation-response programmes, and whether its discrete states provide additional biological resolution beyond identifying responsive programmes.

Recovery of known IFNγ programmes. We first asked whether LEXA can identify biologically meaningful perturbation-response programmes, and how well the recovered programmes agree with independently established biological signatures. Given the substantial experimental noise commonly observed in Perturb-seq data, we selected seven perturbation targets whose responses showed relatively high reproducibility between Rep1 and Rep2, to minimize the influence of experimental noise on the evaluation. Specifically, all methods were fitted on Rep1, and each contributed 32 native programme objects. These objects were ranked according to the magnitude of their whole-transcriptome response to perturbation in Rep1, without using the published signature genes for ranking. We then evaluated how well the top 1 and top 2 objects recovered the genes belonging to the published IFNγ signatures, using AUPRC as the evaluation metric. This design allows us to assess not only whether a method captures known biological responses, but also whether it concentrates the relevant molecular information into a small number of programmes. We evaluated two shared IFNγ signatures (IFNG_program1_up and IFNG_program1_down) across 30 cell-line–target contexts and one MCF7-specific signature (IFNG_MCF7_up) across five MCF7 contexts.

As shown in Table 2, LEXA successfully recovers published IFNγ programmes and achieves competitive performance against other programme representation methods. For IFNG_program1_up, LEXA achieves the highest top-1 AUPRC (0.119), outperforming Tripso (0.063) and cNMF (0.026), while cNMF performs better when considering the top 2 objects (0.249 vs. 0.136). For IFNG_program1_down, Tripso achieves the highest AUPRC for both the top 1 and top 2 objects (0.600 and 0.656), followed by LEXA (0.386 and 0.537). Importantly, for the MCF7-specific signature, LEXA achieves the highest AUPRC for both the top 1 (0.091) and top 2 objects (0.103), outperforming all competing methods. Together, these results suggest that LEXA can effectively recover known perturbation-response programmes using only a small number of programme axes, with particularly better performance in identifying the MCF7-specific response.

Table 2. AUPRC for published IFNγ signatures using the top 1/2 Rep1 response-ranked native programme objects. AUPRC values were calculated over the same 10,218-gene panel.

Published signatureObjectsLEXATripsocNMFNMFSpectraexpiMap
IFNG_program1_upTop 10.1190.0630.0260.0710.0120.021
IFNG_program1_upTop 20.1360.0880.2490.0510.0120.021
IFNG_program1_downTop 10.3860.6000.2120.0160.1180.043
IFNG_program1_downTop 20.5370.6560.4180.0970.1170.062
IFNG_MCF7_upTop 10.0910.0630.0060.0310.0040.007
IFNG_MCF7_upTop 20.1030.0630.0880.0300.0210.006
AuthorsOpenReview ↗

Authors — Response (Part 3/7)

Additional biological resolution from discrete programme states. Beyond recovering known programmes, LEXA provides a more detailed characterization of how cells respond to perturbations. The MCF7-specific IFNγ response offers a particularly informative example. Without using the IFNG_MCF7_up signature for programme selection, the Rep1 whole-transcriptome response ranking identifies Axis 21. The learned gene-responsibility mask of this axis is also significantly enriched for the published MCF7-specific signature, independently supporting the biological relevance of the identified programme.

More importantly, LEXA further resolves the MCF7-specific response into distinct discrete states within Axis 21. As shown in Table 3, we found that MCF7 cells are predominantly distributed among three codes (C0, C1, and C3), with unperturbed cells primarily occupying C0. Perturbations of seven IFNγ-related genes induce distinct changes in the occupancy of these states. In particular, IFNGR1, JAK2, and STAT1 knockouts consistently reduce C0 occupancy and increase C1 occupancy across both independent experimental replicates. IFNGR1 knockout increases C1 occupancy by 10.04 and 8.71 percentage points in Rep1 and Rep2, respectively, whereas JAK2 and STAT1 knockouts produce larger reductions in C0 occupancy (approximately 15–16 percentage points) together with increases in both C1 and C3. Other perturbations, including IFNGR2, JAK1, IRF1, and IRF2, exhibit different or less consistent occupancy changes.

Interestingly, JAK2 and STAT1 perturbations show highly similar patterns of C0 depletion, consistent with their closely related functions in canonical IFNγ signaling. The discrete states also exhibit distinct molecular characteristics: C1 is associated with cell-division and insulin-response programmes, whereas C3 is associated with cell-junction-related functions and reduced oxidative-phosphorylation activity.

These findings demonstrate that LEXA can resolve reproducible and perturbation-specific molecular state-occupancy patterns within a single programme axis, providing additional biological resolution beyond identifying a responsive programme or quantifying its overall activity.

Table 3. State occupancy distributions of LEXA Axis 21 across IFNγ-related gene perturbations in MCF7 cells. The percentages of cells assigned to codes C0, C1, and C3 are reported for seven gene perturbations and non-targeting controls (NT) in two independent experimental replicates (Rep1 and Rep2).

PerturbationRep1 C0Rep1 C1Rep1 C3Rep2 C0Rep2 C1Rep2 C3
NT46.1029.7122.6745.0234.4620.32
IFNGR138.9339.7519.2635.0243.1721.81
IFNGR236.7634.9926.8137.8337.6324.44
JAK140.7329.2427.6837.3937.6524.71
JAK231.1835.4131.1829.5242.0928.06
STAT129.6134.2234.9529.5342.3427.72
IRF141.3631.7325.8239.7737.6122.24
IRF240.9123.8635.2346.2031.6522.15

Together, these findings illustrate the biological value of LEXA's discrete representation at two levels. At the programme level, LEXA identifies established IFNγ responses with competitive accuracy relative to existing programme models. At the state level, it provides additional resolution by distinguishing alternative molecular configurations within a responsive programme and quantifying their condition-specific redistribution. Although continuous programme models such as cNMF can also identify the responsive programme, LEXA directly represents its internal state structure, enabling the same programme to be examined through discrete state identities, occupancy changes, and state-specific molecular effects.

Reviewer_xRzfOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces a framework called LEXA for learning discrete and interpretable gene-program states from single-cell transcriptomic data. The method aims to address the trade-off between the expressive power of continuous latent models and the biological interpretability of traditional gene-program decomposition approaches.

Critical strengths
  1. Retaining fine-grained cellular information while providing interpretable molecular representations is a significant challenge. The proposed program-state formulation offers an intriguing alternative to both conventional continuous latent spaces and fixed gene-program decompositions.

  2. By combining program-specific codebooks with scaffold-guided gene assignments and additive decoding, the framework provides a structured mechanism for characterizing cellular variation. The distinction between program axes and discrete states is conceptually useful, as it allows for the representation of distinct molecular configurations within the same broader program.

  3. The study includes large-scale atlas pre-training, evaluations on multiple held-out datasets, comparisons with several related representation learning methods, and ablation experiments with architecture-matched baselines. These experiments provide valuable evidence regarding the trade-off between discrete representation capacity and downstream predictive performance.

Critical weaknesses
  1. The biological advantages of discrete programme states over continuous programme representations have not yet been fully substantiated. While LEXA introduces a finite set of states for each programme axis, the experimental results do not explicitly isolate the benefits that "discretization" itself provides for biological interpretation. As shown in Figure 3A, continuous representations—when aligned with the underlying architecture—perform comparably to, or even better than, the proposed method in downstream prediction tasks; furthermore, the biological comparisons presented in the paper are primarily benchmarked against traditional factorization methods. Although Figure 4 demonstrates coherent and reproducible gene programmes, it remains unclear whether similar biological structures could be recovered by simply discretizing or clustering the learned continuous programme embeddings. A direct comparison between the proposed method and "post-hoc discretized continuous representations" would provide a more compelling argument for the necessity of the discrete structures introduced here.
  2. Comparisons of "programme enrichment" may be influenced by discrepancies in the number of candidate representations considered. In Figure 4A, the biological programme enrichment analysis for cNMF, LDA, and Spectra is based on the core genes within individual factors or topics, whereas for LEXA, the evaluation considers each state within an axis, selecting the best-matching state as the representative metric for that axis. Despite the authors' use of statistical corrections, LEXA effectively benefits from having multiple candidate states available for each programme axis. This disparity could skew comparisons regarding "programme coverage" and the degree of "programme fragmentation." Further evaluation—standardizing the number of candidates or employing comparable aggregation procedures—would help clarify whether the reported performance gains stem from a superior biological organizational framework or merely from differences in the granularity of the evaluation.
  3. The interpretation of programme states remains contingent upon the additive decoder and the "scaffold" incorporated during the training process. The gene effects described in the paper are derived from the deviations of the decoder's contribution values ​​relative to the mean state within each axis. While these effects offer a useful mathematical interpretation, they do not necessarily correspond to independently existing biological regulatory programs. Furthermore, the "scaffold" incorporates primary cell-type annotations and donor-level variation, which may bias the program axes toward reflecting known cell identities or dataset-specific structures. Although the authors examine the impact of this "scaffold" on characterization quality and reproducibility, further analysis is required to determine whether the learned states capture biological variation that exists independently of the "scaffold" construction.
  4. The reproducibility of individual discrete states is moderate rather than consistently high. Figure 4D shows an average cosine similarity of 0.765 for matched "program axes" and 0.598 for corresponding "state-level" representations, with an average overlap of 46 genes among the top 100 genes. These results support the conclusion of "partial reproducibility" but also indicate significant differences in state-level molecular effects across models initialized independently. Given that a core contribution of this study is the development of an enumerable and reusable "lexicon" of molecular states, the stability of individual "codes" is particularly critical. Further analysis—examining the consistency of code assignments across different cells and the stability of rare states—would help clarify the reliability of the learned "lexicon" for practical applications.
  5. The generalizability of this "program-state lexicon" across different biological contexts remains unclear. Although the experiments demonstrate "atlas transfer" and "cohort-specific refinement," the cohort-specific models are trained independently, resulting in distinct program axes and states. While the authors link axes from the atlas to those within specific cohorts using "gene-responsibility profiles," the extent to which individual molecular states can be consistently transferred, aligned, or reused across different datasets has not yet been fully validated. Conducting a more systematic analysis of state correspondences and stability across datasets will help clarify whether the representations learned by LEXA constitute a broadly reusable "vocabulary" of molecular states, or if they are merely state decompositions specific to a particular dataset.
Questions for authors

Weaknesses

7 comments
AuthorsOpenReview ↗

Authors — Response (Part 1/7)

We sincerely thank the reviewer for the thoughtful and constructive feedback. To address all concerns thoroughly while respecting OpenReview’s character limit, we have organized our response into seven consecutive comments (Parts 1–7). We performed analyses that directly test LEXA's programme-state representation at several levels: post-hoc discretization of continuous representations, search-corrected programme enrichment, state recurrence across independent fits, and molecular validation across independent COVID cohorts and perturbation experiments.

Response to W1: biological advantages of discrete programme states, and comparison with post-hoc discretized continuous representations

Biological advantages of discrete programme states

Thank you for this insightful question. We appreciate the point, as it goes directly to the motivation for representing each programme through a finite set of molecular states. The main advantage we intend for LEXA's finite representation is that it provides a complementary view in which cellular variation is organized into combinations of discrete programme states that can be directly used for biological analysis. The axes provide a compact overview of the programmes active in a cell, while each axis is resolved into alternative molecular configurations of that programme.

This finite organization is useful because biological regulatory systems are often naturally interpreted through recurring configurations and transitions among them, and may in some settings exhibit quasi-multistable behaviors (some of which biologists may term as cell types or cell states). Making these configurations enumerable provides a direct way to inspect the structure of a programme: we can ask where the same state occurs, how its occupancy changes across populations or conditions, and what gene-expression effect accompanies that change. This gives a compact representation of programme structure while retaining a direct molecular interpretation.

We realized that this motivation was not sufficiently demonstrated by the original experiments, and therefore added analyses that directly test the value of this representation:

(1) Additional analysis of the COVID case study

In our extended analysis of the original COVID-A case study, we found that many clinically associated axis-state objects recur across cell classes, allowing shared and cell-context-specific deployment to be represented through the same finite vocabulary. Furthermore, independently fitted COVID-A and COVID-B models also show stable state-frequency correspondence, and frozen COVID-A state-derived molecular signatures transfer to COVID-B. These findings suggest that LEXA's discrete programme states are relatively stable across datasets and may capture reproducible biological structures rather than dataset-specific patterns.

Moreover, we compared LEXA with other programme representation methods in a cross-cohort validation task, where molecular signatures associated with disease severity in COVID-A were evaluated in COVID-B. LEXA achieved the strongest validation results, further supporting the biological relevance and practical utility of its learned discrete states. Please see our response to W5 for details.

AuthorsOpenReview ↗

Authors — Response (Part 2/7)

(2) Additional analysis of an independent perturbation dataset

We further evaluated the biological utility of LEXA's discrete programme-state representation using an independent perturbation dataset from Jiang et al. (Nature Cell Biology, 2025). This experiment aims to address two questions: whether LEXA can recover biologically established perturbation-response programmes, and whether its discrete states provide additional biological resolution beyond identifying responsive programmes.

Recovery of known IFNγ programmes. We first asked whether LEXA can identify biologically meaningful perturbation-response programmes, and how well the recovered programmes agree with independently established biological signatures. Given the substantial experimental noise commonly observed in Perturb-seq data, we selected seven perturbation targets whose responses showed relatively high reproducibility between Rep1 and Rep2, to minimize the influence of experimental noise on the evaluation. Specifically, all methods were fitted on Rep1, and each contributed 32 native programme objects. These objects were ranked according to the magnitude of their whole-transcriptome response to perturbation in Rep1, without using the published signature genes for ranking. We then evaluated how well the top 1 and top 2 objects recovered the genes belonging to the published IFNγ signatures, using AUPRC as the evaluation metric. This design allows us to assess not only whether a method captures known biological responses, but also whether it concentrates the relevant molecular information into a small number of programmes. We evaluated two shared IFNγ signatures (IFNG_program1_up and IFNG_program1_down) across 30 cell-line–target contexts and one MCF7-specific signature (IFNG_MCF7_up) across five MCF7 contexts.

As shown in Table 1, LEXA successfully recovers published IFNγ programmes and achieves competitive performance against other programme representation methods. For IFNG_program1_up, LEXA achieves the highest top-1 AUPRC (0.119), outperforming Tripso (0.063) and cNMF (0.026), while cNMF performs better when considering the top 2 objects (0.249 vs. 0.136). For IFNG_program1_down, Tripso achieves the highest AUPRC for both the top 1 and top 2 objects (0.600 and 0.656), followed by LEXA (0.386 and 0.537). Importantly, for the MCF7-specific signature, LEXA achieves the highest AUPRC for both the top 1 (0.091) and top 2 objects (0.103), outperforming all competing methods. Together, these results suggest that LEXA can effectively recover known perturbation-response programmes using only a small number of programme axes, with particularly better performance in identifying the MCF7-specific response.

Table 1. AUPRC for published IFNγ signatures using the top 1/2 Rep1 response-ranked native programme objects. AUPRC values were calculated over the same 10,218-gene panel.

Published signatureObjectsLEXATripsocNMFNMFSpectraexpiMap
IFNG_program1_upTop 10.1190.0630.0260.0710.0120.021
IFNG_program1_upTop 20.1360.0880.2490.0510.0120.021
IFNG_program1_downTop 10.3860.6000.2120.0160.1180.043
IFNG_program1_downTop 20.5370.6560.4180.0970.1170.062
IFNG_MCF7_upTop 10.0910.0630.0060.0310.0040.007
IFNG_MCF7_upTop 20.1030.0630.0880.0300.0210.006
AuthorsOpenReview ↗

Authors — Response (Part 3/7)

Additional biological resolution from discrete programme states. Beyond recovering known programmes, LEXA provides a more detailed characterization of how cells respond to perturbations. The MCF7-specific IFNγ response offers a particularly informative example. Without using the IFNG_MCF7_up signature for programme selection, the Rep1 whole-transcriptome response ranking identifies Axis 21. The learned gene-responsibility mask of this axis is also significantly enriched for the published MCF7-specific signature, independently supporting the biological relevance of the identified programme.

More importantly, LEXA further resolves the MCF7-specific response into distinct discrete states within Axis 21. As shown in Table 2, we found that MCF7 cells are predominantly distributed among three codes (C0, C1, and C3), with unperturbed cells primarily occupying C0. Perturbations of seven IFNγ-related genes induce distinct changes in the occupancy of these states. In particular, IFNGR1, JAK2, and STAT1 knockouts consistently reduce C0 occupancy and increase C1 occupancy across both independent experimental replicates. IFNGR1 knockout increases C1 occupancy by 10.04 and 8.71 percentage points in Rep1 and Rep2, respectively, whereas JAK2 and STAT1 knockouts produce larger reductions in C0 occupancy (approximately 15–16 percentage points) together with increases in both C1 and C3. Other perturbations, including IFNGR2, JAK1, IRF1, and IRF2, exhibit different or less consistent occupancy changes.

Interestingly, JAK2 and STAT1 perturbations show highly similar patterns of C0 depletion, consistent with their closely related functions in canonical IFNγ signaling. The discrete states also exhibit distinct molecular characteristics: C1 is associated with cell-division and insulin-response programmes, whereas C3 is associated with cell-junction-related functions and reduced oxidative-phosphorylation activity.

These findings demonstrate that LEXA can resolve reproducible and perturbation-specific molecular state-occupancy patterns within a single programme axis, providing additional biological resolution beyond identifying a responsive programme or quantifying its overall activity.

Table 2. State occupancy distributions of LEXA Axis 21 across IFNγ-related gene perturbations in MCF7 cells. The percentages of cells assigned to codes C0, C1, and C3 are reported for seven gene perturbations and non-targeting controls (NT) in two independent experimental replicates (Rep1 and Rep2).

PerturbationRep1 C0Rep1 C1Rep1 C3Rep2 C0Rep2 C1Rep2 C3
NT46.1029.7122.6745.0234.4620.32
IFNGR138.9339.7519.2635.0243.1721.81
IFNGR236.7634.9926.8137.8337.6324.44
JAK140.7329.2427.6837.3937.6524.71
JAK231.1835.4131.1829.5242.0928.06
STAT129.6134.2234.9529.5342.3427.72
IRF141.3631.7325.8239.7737.6122.24
IRF240.9123.8635.2346.2031.6522.15

Together, these findings illustrate the biological value of LEXA's discrete representation at two levels. At the programme level, LEXA identifies established IFNγ responses with competitive accuracy relative to existing programme models. At the state level, it provides additional resolution by distinguishing alternative molecular configurations within a responsive programme and quantifying their condition-specific redistribution. Although continuous programme models such as cNMF can also identify the responsive programme, LEXA directly represents its internal state structure, enabling the same programme to be examined through discrete state identities, occupancy changes, and state-specific molecular effects.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.