acceptodds
Under review as a conference paper at ICLR 2027

Local Influence Inference: Learning Shared Feature Rankings from Spatial Neighborhoods

Abstract

Spatial analysis often seeks to identify which features of a local environment influence an outcome at a focal point. The target is a single feature ranking shared across locations, with influence that decays over distance. Predictive accuracy alone does not ensure recovery of this ranking. In a distance-weighted linear model, we show that bag-averaged permutation importance can reverse two positive influence coefficients even at the true predictor when neighborhood size co-varies with within-neighborhood feature variation. We formulate Local Influence Inference and propose LION, which represents the ranking as a shared model parameter. LION combines a learnable distance kernel, a feature-preserving normalizer, and a global influence vector through bilinear aggregation, using G + 4 trainable parameters. Under correct specification and identification conditions, the reduced maximum-likelihood estimator consistently recovers all strict feature comparisons. Across seven ranking benchmarks spanning synthetic data and real tissue geometry, LION leads or ties distance-weighted logistic regression in mean Spearman on every benchmark, with the largest gains on morphology-interaction (+0.22) and Dense (+0.11). LION ranks first in AUROC on five of seven benchmarks among the evaluated deep and interpretable MIL baselines, with 12 to 64 parameters against thousands for the deep baselines. On three Xenium/TCR-profiled tumors, LION increases the mean reactive fraction among recovered T cells and the mean bystander fraction among missed T cells over every baseline at matched predicted-positive rates. These results support a shared feature ranking as a useful model output for spatial inference, alongside bag-level prediction.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Interpretable Machine Learning, Feature Ranking, Identifiability, Spatial Transcriptomics

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_GK34OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The paper proposed LION (Local Influence Inference Network), a model for spatial data that aims to make predictions based on spatial bags of features that represent sets of spatial observations, and provide a ranking of the most influential observations that most influenced the predictions. Performance is evaluated on 3 variations of MNIST generated using LION itself, 3 modified H&E pathology datasets, and one generated spatial transcriptomic datasets, each generated or modified to provide exact ground-truth rankings of local influence. Further results on real biological data, including Visium as well as Xenium and TCR sequencing data are provided.

Critical strengths

Fewer parameters: As stated in the results, LION performs similarly or competitively with baseline methods, such as GNNs and DeepSet models, including substantially more parameters. The benefits of this less parameterised model are convincing in terms of model complexity.

Additional ablations: The paper reports a wide range of additional ablations that clarify the main results, including the effects of different model components, and further exploratory and descriptive results on real spatial transcriptomics datasets.

Biological relevance: The main setup, aiming to predict spatial labels and evaluate the influence of local observation is relevant for the chosen domain of spatial transcriptomics data analysis.

Critical weaknesses

Readability: The abstract uses very technical language, for example there would likely be a way to describe the proposed methods in a manner more understandable to a lay person. It introduces variables, such as G + 4, the number of trainable parameters, without defining G.. Further, the abstract describes results on “Dense” or “morphology-interaction” tasks, without specifying what these setups refer to. Similarly, technical terms such as reactive or bystander fraction, amongst other concepts, could benefit from further clarification. Similarly, when first reading the manuscript and reaching line 58, I remain unsure what the mentioned coefficients and sign reversal refers to as it only becomes clearer much later in the text. Overall, by providing further clarifications, defining all introduced variables, and explaining technical terminology whenever possible there is great potential to improve the readability of the text for a general ML audience.

Mathematical definitions: The presentation of the mathematical results has many imprecise parts and could be improved. Def 1(i) does not detail for which i the prediction should hold, likely for each i \in I, but it would be nice to specify. Similarly, it would be useful to clearly define Sg and Sh in line 117, to define s^ in line 128 and so on.Def 1 (ii) does not clarify how the learnt vector relates to the prediction function or if it needs to relate at all. Given how much of the other concepts in the following section are not given as definitions, I wonder if this problem statement should be a definition itself without further formalisation. Overall, the theoretical results, specifically the proofs and text in the appendix, could benefit from clearer formalisation to distinguish new theorems and proofs from definitions or general text.

Appendices: While it is useful further materials are included, the appendices are referenced frequently in the main text, including when describing the main empirical results. Overall, this appears as if the main text is not self-contained in itself, but rather requires substantial additional reading, making it difficult to follow and review. For example, the main results on the real biological data are only briefly discussed in Section 6.5, but almost exclusively contained in the appendices. Clarifying and presenting the results to be more concise and self-contained in the main text and focusing the appendix on the most relevant additional content would greatly improve the presentation.

LLM-generated text: Having read the full manuscript, I believe that substantial parts of the text read as if they have been LLM-generated, including large parts of the appendices, and potentially some of the theoretical results, making it difficult to follow the full text. I believe the presentation could benefit from further manual edits to the text to clarify and explain the key results.

Uncertainty evaluation: Due to the lack of uncertainty evaluation reported in the main text it is difficult to assess whether there are critical differences in performance between models, or clear differences in results, say for the main results in Figure 3, 4, 5, or 8. Including relevant confidence intervals for all of the results would greatly strengthen the evaluation. For example, because the results in Figure 4 are averaged after cross validation, it would be to at least report standard deviation estimates as detailed in the full results in the Appendix in Table 5.

Baseline comparisons: LION is compared to only the best rather than all evaluated baseline methods across the results in the main text. For many evaluated datasets there does not seem to be notable differences between LION and this best baseline. Overall, based on the relatively high number of ties reported, it is difficult to assess how significant or relevant an advancement the proposed model is making over the most competitive baseline methods, such as GNN+dist. The full results in Table 5 seem to support that overall there are no significant performance differences between LION and GNN+dist on many of the biological datasets.

Limited relevance of the MNIST evaluation: The biggest performance differences are seen on sparse, dense, or correlated variations of MNIST, which model digits, and are not directly relevant to biological applications. Hence, while LION seems to perform well on this simple image dataset, this does not provide clear evidence that LION would necessarily outperform baseline methods at applications to spatial transcriptomics or histopathology data.

Biological interpretation: The scope, relevance, and impact of the biological results in Section 6.5 is only briefly summarised, and likely deserves further clarification. To address a general ML audience, or even a specialised audience interested in spatial transcriptomics, more detailed explanations and clarifications would be beneficial in the main text.

Missing experimental details: Despite the extensive further materials, certain data processing steps, such as the choice of parameters for Leiden clustering (line 1442), the gene sets used to annotate the cell types, a reference to exactly how the datasets were generated, and the number of generated spatial samples are not explained in appendix K.

Questions for authors

General questions:

  • Does influence always decay over distance? It seems to be worth discussing this spatial assumption as a potential limitation of the proposed setup.
  • Does the learned ranking as defined in Def1 not also need to relate to the learned predictions?
  • Is it not a bit circular to evaluate LION on LION-generated MNIST examples?

Main suggestions:

  • Could you provide a more self-contained and concise summary of your main results in the main text rather than relying heavily on additional appendices?
  • Could you report the full results in the main text with suitable uncertainty estimates?
  • How confident are you that LION will generally outperform alternative predictive models at tasks in spatial biology in particular? Can you comment on for which types of biological datasets and tasks you expect LION to perform well and for which applications it might fail?
3 comments
AuthorsOpenReview ↗

Response to Reviewer GK34 (1/2): ranked object, presentation and uncertainty

We thank the reviewer for the detailed feedback. We first clarify the object being ranked. LION ranks features, such as genes, rather than observations: each feature gg receives one learned coefficient SgS_g, and the ranking is the ordering of these coefficients. The observations in a bag, such as the cells around a focal location, contribute to the bag prediction through a distance kernel, and the same coefficients that define the ranking also define that prediction, f(B)=f(B;S^,φ)f(B)=f(B;\hat{\mathbf S},\varphi). Our primary goal is therefore recovery of the true feature ranking, which we evaluate on benchmarks with known ground truth. Predictive accuracy is a secondary requirement, and we now describe LION's prediction as competitive rather than superior.

Readability and definitions. We revised the abstract and introduction for a general ML audience. The abstract no longer uses undefined symbols or benchmark names and defines reactive and bystander T cells; the introduction now states what the coefficients are before the reversal result; and Section 6.1 defines each benchmark name. Definition 1 now states the index range, defines Sg∗S^*_g, and makes S^\hat{\mathbf S} the feature score vector of the predictor, f(B;S^,φ)f(B;\hat{\mathbf S},\varphi); s^\hat s is defined where it first appears.

Self-containment. The length of the appendix reflects the scope of the paper. We introduce a new problem, recovering a shared feature ranking from bag-level labels, together with a model whose ranking is identifiable. Supporting this requires both theory, namely identifiability, interval estimation, and their proofs, and an evaluation broad enough to show that LION works beyond its own model form: benchmarks with known rankings on synthetic, histology, and spatial transcriptomics data, component ablations, robustness and capacity analyses, and two biological applications. We placed this material in the appendix so that the main text could present the problem, the method, and the main results. The revision adds four analyses in the same way. In Appendix P, a four-step ladder fitted on identical folds attributes most of LION's ranking advantage over DW-logistic to the learned feature normalization and distance kernel. In Appendix D.5, simulated coverage of nominal 95% intervals for the feature contrasts is 0.93 to 0.95 under correct specification. In Appendix L, under four-fold spatial block cross-validation of the Xenium study, LION reaches AUROC 0.80 versus 0.74 for the strongest baseline. In Appendix W, on a semi-synthetic benchmark with real Xenium expression, LION recovers the feature ranking best under four label rules outside its model form. To make the main text readable on its own, Section 6.1 now contains a protocol table that lists, for the main analyses, the data, splits, seeds, selection rule, and where the results are reported, and it defines each benchmark shown in the figures. Section 6.5 now ends with a paragraph explaining what the biological results show.

Writing. In preparing the submission, we used an LLM only to polish the language. The problem formulation, method, theory, experiments, and analyses are our own, and the length of the paper comes from the experimental details and results we report, not from LLM-generated text. In this revision, we have also revised the language to make it clearer and more direct, most substantially in the abstract and introduction.

Uncertainty and baselines. Figure 4 now shows error bars (±1 standard deviation over the 25 runs, 5 folds × 5 seeds): for AUROC and AUPRC, on LION and on the best distance-equipped baseline, and for Spearman correlation, on LION. Table 8 now reports the standard deviation of LION's Spearman correlation on every benchmark:

BenchmarkLION Spearman (mean ± SD)
Sparse0.648 ± 0.001
Correlated0.588 ± 0.002
Dense0.863 ± 0.020
Spatial transcriptomics0.577 ± 0.000
Immune-proximity0.590 ± 0.000
Morphology-interaction0.620 ± 0.069
Multi-type0.698 ± 0.000

A standard deviation of 0.000 denotes a value below 0.0005. Table 7 reports every distance-equipped baseline with its standard deviation, and the caption of Figure 4 now points to it.

AuthorsOpenReview ↗

Response to Reviewer GK34 (2/2): evidence beyond the model form, biology and missing details

Spatial-MNIST. This benchmark is intentionally LION-form: it tests Theorem 1 under correct specification. Evidence beyond the model's own form comes from three sources. First, the spatial transcriptomics rule and the three H&E rules are not LION-form, and LION's Spearman correlation on them is 0.577 to 0.698 (Table 8). Second, in the interaction-strength sweep of Appendix V, as within-instance interactions grow to 83% of the instance-score variance, LION's Spearman changes by only −0.003, whereas each distance-equipped baseline declines by 0.12 to 0.15. Third, a new semi-synthetic benchmark on real Xenium expression and cell geometry (Appendix W) has 360 features, 20 of them active, and four label rules outside LION's generative family: linear, saturating, pairwise interaction, and an annular distance profile. Under every rule, LION's signed Spearman correlation is 0.24 to 0.27, compared with at most 0.085 for six baselines, and its AUROC for separating active from null features is 0.70 to 0.76, compared with at most 0.64.

Biological interpretation. Section 6.5 now ends with a paragraph that explains what the two tissue studies show. In both, the models are trained only on the presence of a cell type, and the biological readouts use information the models never see. In the tumors, LION recovers more TCR-defined reactive T cells, the subset most relevant to antitumor immunity, than models with similar accuracy, and its top-ranked tumor-cell genes are enriched for predicted HLA-binding peptides. In myocarditis, the learned ordering associates mitochondrial respiration in cardiomyocytes with the absence of T cells and macrophages but not of B cells.

Does influence always decay with distance? No, and LION's kernel is monotone non-increasing, so it cannot represent an influence that peaks away from the focal cell. We tested this directly with an annular rule whose influence peaks at 30 µm (Appendix W). LION's ranking recovery under this rule (Spearman 0.273) matched that under the monotone linear rule (0.259), so the misspecified distance profile did not degrade the feature ranking in this setting, although the learned kernel does not reproduce the peak. We discuss this, together with interaction-dominated targets, region-dependent rankings, and overlapping neighborhoods, in a new paragraph of Section 6.4, "When LION is appropriate."

Missing details. Appendix K now states the protocol used to generate the VisiumHD data (10x Genomics user guide CG000685), the marker genes for the cell types used in our analyses, the Leiden clustering parameters (resolution 1.0, 15 nearest neighbors), and the number of spatial samples: two mice, each profiled on one VisiumHD tissue section.

Dear Reviewer GK34,

Thank you again for your thoughtful review. As the discussion period is drawing to a close, we wanted to gently follow up on our response and revision, which rewrites the abstract and introduction and adds Definition 1, Table 1, and the new discussion in Section 6.4. We would greatly appreciate it if you could let us know whether these changes address your concerns, and we are happy to clarify anything further.

Best regards, The Authors

Reviewer_igQLOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies recovery of one feature-coefficient ordering shared across spatial neighborhoods, alongside prediction of a binary outcome at each focal location. LION forms a neighborhood profile by applying a learned exponential distance kernel and a learned-temperature softmax over each instance's features. A shared feature-score vector and logistic output then predict the bag label. The implementation has G+4 trainable parameters; removing an affine score ambiguity gives a G+2-dimensional reduced model.

Under correct specification, global identification, and regularity assumptions, the paper establishes consistency of the reduced global maximum-likelihood estimator and recovery of strict feature comparisons. A separate oracle example shows that within-bag permutation sensitivity can reverse two positive coefficients when bag size and within-bag variation co-vary. Evaluation comprises seven constructed-label ranking benchmarks, including real H&E geometry, and observed-label VisiumHD and Xenium/TCR case studies. Results support a compact predictive model and promising shared rankings, while leaving questions about comparative attribution of the gains and the reproducibility of the biological evidence.

Critical strengths

1. A precise, compact model with unusually careful treatment of the interpretation target.

LION exposes a shared ordering with G+4 trainable parameters and makes the affine ambiguity explicit. The reduced contrasts have a coherent statistical interpretation, and the analysis includes uncertainty in the learned distance and normalization parameters. Theorem 2 is particularly useful because it demonstrates a reversal with two positive coefficients at the true predictor and matched per-instance marginals; the result is therefore more than the familiar signed-versus-absolute-importance mismatch. The manuscript also correctly limits this construction to its stated linear oracle and within-bag perturbation rule.

2. Substantial empirical controls support the usefulness of the specification.

The evaluation includes shared distance inputs and paired folds for the main deep-model comparison, tuning diagnostics, interpretable MIL comparators, and a signed DW-logistic comparator. The synthetic transcriptomics and H&E rules test departures from the exact LION generator. In the separate coefficient comparison, LION improves mean Spearman from 0.74 to 0.85 on Dense and from 0.40 to 0.62 on morphology-interaction. The authors also report cases where Additive MIL or alternative LION configurations perform better, and explain tied-target metric ceilings. The Xenium study reserves TCR-defined groups for evaluation and excludes identity/reactivity-defining inputs, offering an interesting additional test beyond prediction accuracy.

Critical weaknesses

1. The evidence does not yet isolate or fully document the central advantage of the proposed ranking model.

The large main ranking gaps compare native coefficients with absolute perturbation sensitivity, which the paper correctly acknowledges are different estimands. The more relevant signed DW-logistic comparison ties at reported precision on two benchmarks and differs by only 0.01–0.04 on three others; the two substantial gains occur on a LION-form generator and a morphology-interaction rule for which representation misspecification matters. Appendix P states requirements for matching units, folds, and tuning, but does not give the actual feature representation, decay/regularization search, or compositional identification convention used for the reported full-ranking comparison. Paired ranking uncertainty is also unavailable. Both LION and DW-logistic learn coefficients directly, so their difference cannot itself establish a benefit of direct coefficient estimation over attribution.

The component evidence also needs clarification. Appendix T describes both sum pooling and removal of distance normalization as unnormalized exponential weighting, although Table 10 reports substantially different results. The replacement used when “removing the influence vector” is unspecified. With both normalizers retained, replacing S by a shared constant vector would give a constant predictor; an alternative readout must therefore be defined before the reported nontrivial results can support the component interpretation. These are reporting concerns, not evidence that the implementation used the constant replacement.

2. The observed-tissue evidence is promising but insufficiently specified to validate the biological interpretation.

All seven exact-ranking benchmarks use constructed labels, making the observed-label case studies especially important. Appendices K–L do not adequately specify tissue sampling and train/test partitions, spatial separation of neighborhoods, or the operational rule assigning TCR clonotypes to reactive/bystander groups. The matched predicted-positive rate and its selection procedure are also unspecified, and absolute per-tumor group counts/results are absent. The exclusions of defining genes and the three-tumor pairing are useful controls, but do not resolve these questions. It is therefore unclear how much the results establish spatial interpolation, independent-region generalization, or subgroup enrichment at a particular operating point; I do not infer that leakage occurred.

There is also an unresolved discrepancy in the gene-level evidence: Appendix K specifies 500 input genes, whereas the GSEA plots in Figures 10(c) and 11 extend to approximately 3,000 ranked genes. The derivation of that longer ordering is not explained. The TCR composition and NetMHC analyses do not independently establish the correctness of the shared gene ordering, as the authors acknowledge. Clear provenance of the ranked lists and the biological evaluation protocol is needed before these results can substantiate the interpretation claim.

Questions for authors

1. Coefficient comparison — clarification and targeted analysis. What exact feature representation, scaling, contrast-identification convention, outer splits, and distance/regularization selection produced Figure 3? Please report paired per-run ranking differences, especially for Dense and morphology-interaction. Can the contribution of learning the normalization temperature be separated from choosing the nonlinear representation, using the same folds and fitting criteria? A logistic fit with LION's full learned profile is the same reduced prediction family, so the purpose is to isolate representation and fitting choices, not to demand a redundantly named baseline.

2. Component evidence — clarification. What are the exact forward equations for LION-sum-pool, LION-unnorm-dist, and the model without the influence vector? How is the last model's feature ranking obtained? Please explain which computation or training setting distinguishes the first two variants and accounts for the different Table 10 results.

3. Xenium evaluation — clarification and targeted validation. Please give the reactive/bystander assignment rule, excluded input features, per-tumor group counts, train/test split unit, and matched-positive-rate value/selection procedure. Do the main effects persist when evaluation neighborhoods are spatially separated from training neighborhoods, with overlapping input cells excluded? Please provide tumor-level absolute metrics and paired changes, including DW-logistic; a held-out-tumor analysis would be useful if feasible, but is not a request for additional patients.

4. Gene-ranking provenance — clarification. Which gene list and scores generated Figures 10(c) and 11, and how were approximately 3,000 ranks obtained from the stated 500-gene input? Was an additional model or ranking procedure used? Resolving this would materially affect my interpretation of the pathway evidence.

5 comments
AuthorsOpenReview ↗

Response to Reviewer igQL (1/3): source of the ranking advantage over DW-logistic

We thank the reviewer for the careful and detailed review. We agree that several reporting gaps made the evidence difficult to assess. Questions 2 to 4 concern how existing analyses were specified, and we answer them with exact definitions and corrected text. Question 1 required a new analysis, which we now report in Appendix P. All corresponding changes are in the revised PDF.

Q1 (coefficient comparison). We agree that Figure 3 did not isolate the source of LION's advantage over DW-logistic. Appendix P now states the full DW-logistic protocol: the outer folds and seeds of the main cross-validation, the shared base representation with any scaling estimated on training data only, the distance-decay and L2-regularization grids selected jointly by validation binary cross-entropy on an inner label-stratified split, refitting on the full outer training set, conversion of coefficients to ground-truth units, and signed Spearman correlation.

Following the reviewer's suggestion, we fitted a four-step ladder with identical outer folds and inner selection criteria, changing one component at each step: (L0) DW-logistic on the base representation; (L1) DW-logistic on softmax-normalized features with the temperature fixed at b=1b=1; (L2) unpenalized logistic regression on LION's learned profile, that is, with the normalization temperature and distance kernel learned by LION on the same training fold; and (L3) LION. For each step we averaged the Spearman difference over the five seeds within each fold and computed a 95% t interval across the five folds:

StepDenseMorphology-interaction
L1 − L0 (fixed-temperature normalization)+0.030 [0.011, 0.049]+0.060 [0.023, 0.097]
L2 − L1 (learned temperature and kernel)+0.065 [0.034, 0.096]+0.135 [0.073, 0.197]
L3 − L2 (sigmoid constraint and regularized training)+0.012 [−0.005, 0.029]+0.022 [−0.007, 0.051]

The three steps sum to 0.107 on Dense and 0.217 on morphology-interaction. The fixed-temperature normalization accounts for 28% of this total on both benchmarks, and learning the temperature and distance kernel for 61% on Dense and 62% on morphology-interaction; the interval for the last step includes zero. LION's advantage over DW-logistic therefore comes from its representation, mainly the learned feature normalization and distance kernel, and not from its training procedure: LION's estimates are comparable to the profile maximum-likelihood fit in the same model family. The ladder and these intervals are now reported in Appendix P.

AuthorsOpenReview ↗

Response to Reviewer igQL (2/3): ablation definitions and component evidence

Q2 (component evidence). We agree that Appendix T did not distinguish LION-sum-pool from LION-unnorm-dist. Both use the unnormalized kernel kij=exp⁡(−eadij)k_{ij}=\exp(-e^{a}d_{ij}), but they pool it differently. Writing Ki=∑j=1nikijK_i=\sum_{j=1}^{n_i}k_{ij} for the total kernel mass of bag ii, the default model uses oi=∑jkijzij/Kio_i=\sum_j k_{ij}z_{ij}/K_i; LION-unnorm-dist averages over the bag, oi=ni−1∑jkijzijo_i=n_i^{-1}\sum_j k_{ij}z_{ij}; and LION-sum-pool sums, oi=∑jkijzijo_i=\sum_j k_{ij}z_{ij}. Equivalently, the two variants multiply the default bag score by Ki/niK_i/n_i and by KiK_i, respectively: in LION-unnorm-dist the bag score also depends on how close the instances lie to the bag center, and in LION-sum-pool it depends in addition on how many instances the bag contains. Our description of LION-unnorm-dist as removing the distance softmax normalization omitted the 1/ni1/n_i factor, which is why the two descriptions read as the same model. Appendix T now states the exact forward equation of every variant, including LION-no-dist (wij=1/niw_{ij}=1/n_i) and LION-minimal.

The variant described as removing the influence vector did not replace S\mathbf S by a constant vector. As the reviewer notes, with both normalizers retained a constant S\mathbf S would give a constant predictor, because ∑gEij,g=1\sum_g E_{ij,g}=1 and ∑jwij=1\sum_j w_{ij}=1. The variant, now named LION-instance-MLP, replaces the linear instance score zij=S⊤Eijz_{ij}=\mathbf S^\top\mathbf E_{ij} by an instance-level MLP, zij=hθ(Eij)=v⊤ReLU(WEij+c)+c0z_{ij}=h_\theta(\mathbf E_{ij})=\mathbf v^\top\mathrm{ReLU}(\mathbf W\mathbf E_{ij}+\mathbf c)+c_0, with one hidden layer of width 32, biases in both layers, and no dropout, batch normalization, or output sigmoid. The distance weights, feature normalizer, pooling, and output are unchanged: ηi=eα∑jwijzij+β\eta_i=e^\alpha\sum_j w_{ij}z_{ij}+\beta and y^i=σ(ηi)\hat y_i=\sigma(\eta_i).

Because hθh_\theta has no per-feature coefficient, we read out its ranking from the gradient of the bag logit with respect to the normalized features, qig=∑j∂ηi/∂Eij,g=eα∑jwij ∂hθ(Eij)/∂Eij,gq_{ig}=\sum_j\partial\eta_i/\partial E_{ij,g}=e^\alpha\sum_j w_{ij}\,\partial h_\theta(\mathbf E_{ij})/\partial E_{ij,g}, summed over valid (non-padded) instances. Because the normalized features lie on the simplex, the gradient component along the all-ones direction is not constrained by the data, so we center qigq_{ig} across features and average over all training bags Itr\mathcal I_{\mathrm{tr}} of the run with equal weight: rg=∣Itr∣−1∑i∈Itr(qig−G−1∑hqih)r_g=|\mathcal I_{\mathrm{tr}}|^{-1}\sum_{i\in\mathcal I_{\mathrm{tr}}}(q_{ig}-G^{-1}\sum_h q_{ih}). Features are ordered by signed rgr_g, without absolute values, and compared with the signed target. Gradients are computed at the checkpoint selected by validation prediction loss, not by ground-truth Spearman. The readout is consistent with LION: for the linear score it gives qig=eαSgq_{ig}=e^\alpha S_g and rg=eα(Sg−Sˉ)r_g=e^\alpha(S_g-\bar S), which preserves LION's ordering because eα>0e^\alpha>0. Under the three-seed ablation protocol of Table 13, LION-instance-MLP reaches a Dense Spearman of 0.275 versus 0.844 for the default configuration, a change of −0.569-0.569; LION-no-dist reaches 0.836, a change of −0.008-0.008. Appendix T now states this forward equation and readout.

The Xenium component comparison in Section 6.5 uses the same two variants, and we have narrowed its interpretation. LION-instance-MLP has 0.0210.021 higher mean AUROC than LION-no-dist, but at a matched predicted-positive rate within each tumor, its reactive fraction among true-positive T cells is 9.0 percentage points lower and its bystander fraction among false-negative T cells is 8.2 points lower, with the same direction in all three tumors. T-cell-presence discrimination and TCR-defined composition therefore need not agree between these two variants, and this comparison alone does not establish that the gene ranking is correct.

AuthorsOpenReview ↗

Response to Reviewer igQL (3/3): Xenium protocol and gene-ranking provenance

Q3 (Xenium). We have completed the protocol description in Appendix L. Cells annotated as T, CD4+ Th, CD8+ T, or Treg are T cells, and the training target is T-cell presence. For evaluation only, a T cell is reactive if it has a TCR clonotype-probe hit and bystander otherwise; non-T cells belong to neither group, even when they carry a clone hit. Every model's input excludes 14 T-cell identity markers, 20 activation- and reactivity-associated markers removed as a precaution, the TCR constant-region genes, and all clonotype probes, and features absent from any of the three panels are also removed, so all models use the same 360 features in every tumor. The matched predicted-positive rate is 5% within each tumor: each model's predicted positives are its highest-scoring 5% of test cells, so all models select the same number of cells and differ only in which cells they select.

To exclude leakage through shared neighborhoods, we repeated the evaluation with four-fold spatial block cross-validation. Training focal cells lie more than 300 µm (twice the 150 µm neighborhood radius) from the held-out test block, and cells inside this buffer are excluded from training. We also checked every test bag for input cells shared with any training bag; after the buffer, none remained, so no additional test bag was removed and all models use the same test set. On this test set, LION reaches AUROC 0.80 versus 0.74 for the strongest baseline, and at the 5% predicted-positive rate its reactive fraction among TP T cells and bystander fraction among FN T cells are 10 and 8 percentage points higher, respectively. The preferential recovery of reactive T cells therefore does not depend on overlap between training and test neighborhoods.

Q4 (gene-ranking provenance). We thank the reviewer for catching this inconsistency. The statement of 500 input genes in Appendix K was a typo in the text; the analysis itself is unchanged. The myocarditis data comprise two VisiumHD sections. For each section, we selected the 2,500 genes with the highest mean VCM expression, using expression alone and no infiltration labels; the union of the two lists contains 3,243 unique genes. For each infiltration target, a single LION model was trained on bags pooled from both sections with these 3,243 genes as shared features (G=3,243G=3{,}243), and the GO and preranked GSEA analyses in Figures 10(c) and 11 use the full ordering of all 3,243 genes. No other gene list or ranking procedure was involved. We have corrected Appendix K and the captions of Figures 10 and 11 accordingly.

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

The authors propose a MIL model which is interpretable by design with directly learned feature ranking with statistical guarantees. the feature ranking is a parameter of the model itself. Although very small model, it perform superior than the deep MIL models. The estimated ranking comes with confidence interval, which is important in real world medical cases. Moreover, it recovers the true feature ranking on experiments with known ranking.

Critical strengths

The experiments are extensive; each answers a specific, clear question and supports the method. The number of baselines compared is substantial, and they don’t evaluate just one aspect of their method. the ranking They care about theoretical guarantees, although it is hard to provide confidence intervals, which are very important in real-world problems, for a method that works on real-world data, so I really appreciate their work.

Critical weaknesses

This is minor but the benchmark with known ranking is simple and low dimension. It's not clear to me how the method performs (not for the final tasks but in terms of ranking) on more complex data.

Questions for authors

Can you show an example where the label rules involve feature interactions or non-linearity? Can you provide the confidence intervals explained in section 5 in one of the datasets?

3 comments
AuthorsOpenReview ↗

Response to Reviewer JmBR: nonlinear rules and confidence intervals

We thank the reviewer for the positive assessment and for two concrete suggestions, both of which we have implemented.

Nonlinear and interaction rules. Section 6 and Appendix V already include an interaction-strength sweep: as within-instance interactions grow to 83% of the instance-score variance, LION's Spearman changes by only −0.003 (0.647 to 0.644), whereas each distance-equipped baseline declines by 0.12 to 0.15. To address higher dimension, we added a semi-synthetic benchmark on real Xenium expression and cell geometry (Appendix W): 360 features, 20 of them active, with labels injected by four rules outside LION's generative family (linear, saturating, pairwise interaction, and an annular distance profile peaking at 30 µm). Across the four rules, LION's signed Spearman correlation is 0.24 to 0.27, compared with at most 0.085 for six baselines, and its AUROC for separating active from null features is 0.70 to 0.76, compared with at most 0.64. The interaction rule gives LION's lowest values, and its predictive AUROC is comparable to the strongest baseline under every rule.

Confidence intervals. Theorem 1 gives Wald intervals for the identifiable contrasts Dg−DhD_g-D_h (Appendix D.3), and we now assess their coverage by simulation in a new Appendix D.5. Under correct specification with G=30G=30 features, pointwise coverage of nominal 95% intervals for all 435 pairwise contrasts is 0.931, 0.945, and 0.948 for I=I= 500, 2,000, and 5,000 bags, and simultaneous Bonferroni coverage is 0.990, 0.945, and 0.980. At I=500I=500, the intervals for contrasts between active features and the null-feature mean are anticonservative (0.872), mainly because the MLE overstates these contrasts in small samples; their coverage is 0.925 at I=2,000I=2{,}000 and 0.939 at I=5,000I=5{,}000. In every simulated data set, the interval for each of the 125 active-null contrasts excluded zero with the correct sign. We also examined a misspecified distance kernel, the unnormalized hard-boundary Gaussian kernel of our spatial transcriptomics rule. There, model-based intervals undercover the contrasts among active features (0.90 to 0.91), and sandwich intervals restore their coverage to 0.94 to 0.96. We therefore recommend sandwich intervals when the kernel may be misspecified, and Appendix D.5 gives both forms. On real tissue geometry, Appendix D.6 applies these intervals to the H&E multi-type rule, in which blood has a negative injected effect. The ratio of the blood contrast to the immune-cell contrast, each measured against the zero-influence features, is −0.40 with a 95% model-based interval of [−0.60, −0.20]. The interval excludes zero and contains the injected ratio of −0.5. Because this generator lies outside LION's model family, this interval is model-based.

Thanks for the response. I find the the current state of the paper satisfactory and recommend acceptance.

Thank you for taking the time to read our response and for recommending acceptance. We are glad that the new semi-synthetic benchmark and the coverage analysis of the confidence intervals addressed your questions, and we sincerely appreciate your comments, which helped us improve the manuscript.

Comments on the paper

3 comments

Thank you for the follow-up. We believe this comment may have been intended for another submission. We would be grateful if you could check the thread, and we welcome any remaining questions about our responses.

I apologize for the mistake. I will post my comments later.

AuthorsOpenReview ↗

General response

We thank the reviewers and the AC. After the discussion, Reviewer JmBR recommends acceptance, and Reviewer igQL confirmed that each of their four questions was resolved and raised the rating. The main questions across the reviews are each answered by a new analysis in the revised PDF (changes marked in blue):

  1. Does ranking recovery extend beyond LION's own model form? (JmBR, GK34) A new benchmark on real Xenium expression and cell geometry (Appendix W) has 360 features and four label rules outside LION's generative family, including pairwise interactions and a non-monotone distance profile. Under every rule, LION's signed Spearman correlation is 0.24 to 0.27, compared with at most 0.085 for six baselines.
  2. Uncertainty of the estimates and results (JmBR, GK34). Nominal 95% intervals for the feature contrasts reach simulated coverage of 0.93 to 0.95 (Appendix D.5), and Figure 4 and Table 8 now report standard deviations over 25 runs.
  3. Source of the advantage over DW-logistic (igQL). A four-step ladder on identical folds (Appendix P) attributes about 90% of LION's ranking gain to its learned feature normalization and distance kernel.
  4. Possible leakage through overlapping neighborhoods in the Xenium study (igQL). Under spatial block cross-validation with a 300 µm buffer and no shared input cells (Appendix L), LION's advantage holds (AUROC 0.80 versus 0.74 for the strongest baseline).

The remaining points were clarifications, now in the paper: exact forward equations for the ablations (Appendix T), a corrected typo in Appendix K, and clearer definitions and protocol details in the abstract, introduction, and Section 6.

We thank the reviewers again for questions that made the paper clearer and better supported.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.