{"id":"c582ba66-6ca9-48de-9498-7fcef253e017","arxiv_id":"2412.16247","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"Dictionary learning with ICFL and control-data whitening extracts sparse features from microscopy foundation models that correlate with cell types and genetic perturbations.","lead":"This paper applies sparse dictionary learning to the internal representations of a microscopy foundation model, and finds features that align with known cell types and genetic perturbations. It introduces a new algorithm, ICFL, with control-data PCA whitening, and reports that the extracted features are about as informative as hand-crafted biological image features.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Layer selection for the functional-group task is reused in evaluating that same task, so the reported preservation of genetic-perturbation signal is partly selection-optimistic.","rationale":"The reader's verdict is CONDITIONAL, and I agree that conditional is appropriate. The reader's weakest-assumption point about PCA whitening is a genuine generalizability concern, but the paper already shows ablations with and without whitening, and the central claim does not require whitening to be universally valid for all biological perturbations. The more directly load-bearing flaw is the layer-selection leakage: Section 5 discloses that the layer was chosen by maximizing linear probing performance on the functional-group task, and the same task is then used to quantify how much biological signal is preserved. This makes the quantitative support for 'genetic perturbations' optimistic in an unknown amount. The qualitative validations (OPA1, ALG3, TSC2, adherens junctions) and the CellProfiler correlation (0.71) provide independent evidence that some extracted features are biologically meaningful, so the concern does not refute the paper's core direction. But the specific numbers for functional groups should be trusted only after a held-out layer-selection test. Since the reader already flagged selection bias in the rationale and issued CONDITIONAL, my analysis does not move the verdict; it sharpens the reason. The proposed concrete test is straightforward and does not require retraining the foundation models, only re-running layer selection and probing on disjoint splits.","tokens_in":18595,"tokens_out":8656,"duration_ms":69184,"concrete_test":"Hold out a disjoint subset of functional-group data (e.g., a separate set of experiments) for layer selection only. Select the layer using linear probing on that subset, then train ICFL and the downstream linear probes on the remaining held-out data, and recompute the Task 5 BTA in Figure 4A and the selectivity counts in Table 2. If the Task 5 BTA drops materially below the reported value or the number of features with selectivity above 0.2 falls substantially, the reported preservation is partly an artifact of layer-selection leakage.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 5 states that the token-extraction layer (16 for MAE-L, 33 for MAE-G) was selected by maximizing linear probing performance on the functional-group task. This is the same task used in Section 6.1 and Figure 4A to demonstrate that ICFL reconstructions preserve biological signal: Task 5 (functional gene groups) is the hardest genetic-perturbation benchmark, with an original BTA of 32.1%. Because the dictionary is learned from representations at that layer, all subsequent selectivity counts (Table 2) and reconstruction-probing accuracies are conditioned on a layer chosen to make exactly those labels linearly separable. No nested or held-out layer selection is reported, so the headline evidence for 'genetic perturbations' is an optimistically biased estimate of how well the method preserves functional-group signal. This is an internal evaluation leak rather than merely a domain assumption: the same labels are used for model selection and for measuring the method's success. The central claim that dictionary learning retrieves genetic perturbations rests partly on these Task 5 numbers, so the bias is load-bearing. The ICFL algorithm itself remains unsupervised, which limits the leakage to layer choice, but it does not eliminate the problem.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Iterative Codebook Feature Learning (ICFL), a dictionary-learning method that combines batched matching pursuit with gradient updates of the decoder, and couples it with a PCA whitening transform estimated on unperturbed control cells. The authors apply this pipeline to two masked-autoencoder microscopy foundation models (MAE-L and MAE-G) and evaluate whether the resulting sparse features correspond to known biological concepts. They report that ICFL produces fewer dead features and better reconstruction cosine similarity than TopK sparse autoencoders, that features with high selectivity exist for cell types, batches, and genetic perturbations, that linear probes on ICFL reconstructions preserve a large fraction of the original representation's balanced test accuracy, and that domain experts can interpret selected features in terms of morphology and single-cell identity. The work is framed as a step toward mechanistic interpretability for scientific imaging data.","tokens_in":18765,"tokens_out":10358,"duration_ms":84639,"significance":"If the results are robust, this is a useful and timely contribution: it extends sparse dictionary learning from text and natural-image models to unsupervised scientific foundation models, proposes a practical control-data whitening preprocessing step, and introduces a five-task benchmark for measuring biological signal retention. The use of public datasets (RxRx1/RxRx3), comparison with the standard TopK SAE baseline, and expert-based case studies are strengths. However, the central quantitative claims are currently supported only by point estimates, and the functional-group preservation result is affected by an evaluation leak in layer selection. The method is not fully unsupervised, since the whitening step uses knowledge of control samples; this should be stated prominently.","major_comments":[{"comment":"The token-extraction layer is selected by maximizing linear probing performance on the functional-group task (§5), and the same task (Task 5) is then used in Fig. 4A and Table 2 to demonstrate that ICFL reconstructions preserve genetic-perturbation signal. This makes the reported Task 5 BTA and selectivity counts selection-optimistic: the layer is a hyperparameter of the full pipeline, and the evaluation labels have been used to set it. I request a nested/held-out evaluation (layer selection on a validation split, reporting on a disjoint test split) or, failing that, a plot of the Task 5 BTA/selectivity across all candidate layers. This is a correctable evaluation leak, but it directly affects the headline claim that dictionary learning retrieves genetic perturbations.","section":"§5, Fig. 4A, Table 2"},{"comment":"The main quantitative comparisons (BTA in Fig. 4A, selectivity counts in Table 2, Pearson r=0.71 in Fig. 2C) are reported without confidence intervals or error bars. The differences that support claims such as 'ICFL retains more biological signal than TopK SAE' are often a few BTA points, and some counts are small (Task 5, threshold 0.2: 37 features). Please provide confidence intervals (e.g., bootstrap over wells/experiments or multiple training seeds) for the key numbers, so the reader can assess whether the reported improvements are within sampling noise.","section":"§6.1–6.3, Fig. 4A, Table 2"},{"comment":"The PCA whitening transform is estimated on unperturbed HUVEC control cells and applied to all other cell types and perturbations. This assumes that the control covariance captures nuisance variation that is shared across conditions and cell types, and that downweighting these directions does not remove signal needed to distinguish perturbations or non-HUVEC cell types. The assumption is plausible but untested; the paper should report sensitivity, e.g., by comparing whitening transforms estimated on per-cell-type controls (for Task 1) or by showing that the top control variance directions are not informative for the perturbation tasks. The with/without-whitening ablation partially addresses this, but not the transfer across cell types.","section":"§4.1"}],"minor_comments":[{"comment":"The sentence 'We selected this layer by finding the maximized linear probing performance on the functional group task from the original embeddings (§ 5)' refers to the section in which it appears; please specify the number of layers tried, the split used, and whether the selection was repeated for each model.","section":"§5"},{"comment":"The procedure for selecting the feature directions displayed in Figure 3 is not described; please state whether these were chosen post hoc because they separate the two groups, and report how many features show separation of comparable quality.","section":"§6.1, Fig. 3"},{"comment":"The single-cell recall analysis is based on five images selected among the strongest-correlated images and a single expert annotator; please state the image-selection rule and consider reporting inter-annotator agreement.","section":"§7, Table 3"},{"comment":"The symbol W denotes both the PCA whitening matrix in Section 4.1 and the dictionary matrix in Section 3; renaming one of these would remove ambiguity.","section":"§4.1"},{"comment":"The dead-feature criterion 'activated less than a fraction of 10^-5× during the last 1000 training steps' is ambiguous; please define the denominator and the activation threshold explicitly.","section":"Table 1"},{"comment":"For the CellProfiler comparison, please specify how the quantile thresholds α are chosen and whether the same α is applied to all labels; matching the sparsity only on average may not make the features equally comparable across labels.","section":"§6.2"},{"comment":"Please clarify whether the 'original representation' baseline is the PCA-whitened or non-whitened representation, so that the with- and without-whitening panels can be interpreted correctly.","section":"Fig. 4A"}],"recommendation":"major_revision","confidential_remarks":"The layer-selection leak in §5 is the issue most likely to be raised by readers; it is fixable and should be addressed before acceptance. I would also ask the editor to encourage the authors to release code and trained dictionaries, because the current experiments (300k iterations, 40M tokens) are expensive to reproduce. The paper is a good fit for the journal's scope despite the exploratory nature of the results."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Bottom line: this paper deserves serious referee time, but it needs a revision to address an evaluation leak. The core demonstration — that sparse dictionary learning can extract features from a microscopy MAE that align with known biology — is real and useful. ICFL is a modest algorithmic variant (batched OMP encoding with gradient descent on the decoder) that does what it claims: fewer dead features, better reconstruction at the same sparsity than TopK SAE. The PCA whitening on control data is a nicely motivated domain-specific preprocessing step, and the selectivity/probing benchmark is carefully constructed (mutually exclusive experiments, class-balanced accuracy). The case studies on OPA1, ALG3, TSC2, and adherens junctions are genuinely compelling qualitative evidence.\n\nThe soft spots are real but mostly fixable. The stress-test concern is correct: Section 5 says the token-extraction layer (16/33) was selected by maximizing linear probing on the functional-group task, and that same task is used in Figure 4A to claim ICFL preserves genetic-perturbation signal. So the Task 5 numbers are selection-optimistic. It is not fatal — the dictionary learning itself is unsupervised and the leak is confined to layer choice — but the authors should either use nested or held-out layer selection, or show robustness across a range of layers. The paper also lacks confidence intervals for BTA and selectivity, which matters when the differences between methods are not huge. The single-cell validation is 124 cells from 5 images and one feature; that is a nice illustration, not a quantitative claim. The PCA whitening assumption — that control covariance captures all nuisance variation and generalizes across cell types and perturbations — is plausible but untested; a cross-cell-type check would help. Related work is well covered, and the new-result claim is fair: the domain application and the ICFL/whitening recipe are not in prior work.\n\nWho this is for: computational biologists using foundation models, and the mechanistic interpretability community looking for applications outside text. It is a solid demonstration, not a breakthrough. I would cite it for the domain application and the practical recipe. Recommendation: send to peer review with requests to fix the layer-selection leak, add uncertainty quantification, and release code/checkpoints.","headline":"A useful demonstration that dictionary learning can extract biologically meaningful features from microscopy MAEs; the main evaluation has a layer-selection leak that should be fixed before publication.","tokens_in":19377,"tokens_out":2970,"would_cite":true,"duration_ms":25078,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Sparse dictionary learning can pull biologically meaningful concepts—cell types, genetic perturbations, and subtle morphological changes—out of an unsupervised microscopy foundation model, without text or labels, if representations are…","keywords":["dictionary learning","mechanistic interpretability","microscopy foundation models","masked autoencoders","sparse coding","cell painting","PCA whitening","biological concepts"],"falsifier":"Run the same ICFL pipeline on a different microscopy dataset in which the control images come from one cell line and the perturbed images from another; if linear-probe accuracy on the non-control cell line's perturbations falls well below the unwhitened baseline while the control line's accuracy stays high, the control-derived whitening is removing cross-cell-type biological signal, and the reported generality would not transfer.","tokens_in":18322,"feed_emoji":"🧫","tokens_out":6378,"duration_ms":50614,"temperature":0.7,"pith_summary":"This paper asks whether sparse dictionary learning—the technique used to find interpretable features inside large language models—can also expose meaningful concepts inside an unsupervised vision model trained on cell images, where no text or labels guide the model. The authors propose combining a new dictionary-learning algorithm, Iterative Codebook Feature Learning (ICFL), with a PCA whitening step derived from unperturbed control cells, and report that the resulting features align with cell types, genetic perturbations, and functional gene groups. They find that ICFL features match the interpretability of expert-designed CellProfiler features, that linear probes on the sparse reconstructions retain most of the biological signal of the original representations, and that token-level heatmaps can separate perturbed from unperturbed cells in a single image. If these results hold, interpretability tools originally built for text models can become a discovery engine in a domain where experts do not yet know which concepts to look for.","feed_headline":"Sparse features expose cell types and gene edits inside microscopy AI","feed_subtitle":"A dictionary-learning step surfaces interpretable biology from a masked autoencoder, matching hand-crafted cell features","key_machinery":"The central object is the learned dictionary, a matrix $W_{\\text{dec}}$ whose columns are feature directions in the embedding space, together with sparse coefficient vectors $z$ such that a representation $x$ is approximately $W_{\\text{dec}} z + b_{\\text{pre}}$. The algorithm that carries the argument is ICFL, a batched variant of matching pursuit: at each iteration it selects the codebook columns most aligned with the current residual, solves for their coefficients, subtracts the reconstruction, and repeats, while the codebook itself is updated by gradient descent on the reconstruction loss. The other load-bearing component is a PCA whitening map computed from unperturbed control cells, which reweights the representation space so that dominant control-derived variance directions are downweighted and perturbation-relevant differences are amplified before ICFL is applied. This combination is what lets a decoder maximize reconstruction fidelity without being dominated by nuisance variation.","core_discovery":"On its own terms, the paper's central claim is that a masked autoencoder trained on microscopy images stores biologically meaningful concepts as sparse, linear feature directions, and that those directions can be recovered by dictionary learning if the encoder is paired with a control-derived whitening transform. The authors report retrieving features that correlate with specific cell types, individual siRNA and CRISPR perturbations, and functional gene groups such as adherens junctions, and they show that these features reveal morphology at a single-cell resolution in token-level heatmaps. They furthermore claim that the recovered features preserve a substantial fraction of the linear-probing signal of the original representations on five biological tasks, and that they attain selectivity comparable to hand-crafted CellProfiler features. The argument is empirical: the evidence is a set of selectivity scores, linear-probing accuracies, expert visualizations, and a small expert-annotated single-cell recall table.","pith_inferences":["Beyond the paper: if feature directions genuinely correspond to morphological concepts, they could serve as automated hypothesis generators for drug mechanism-of-action, by linking a new perturbation's feature profile to known pathways without manual image review.","Beyond the paper: the reliance on a single control cell type suggests a testable extension: derive the whitening transform from multiple control conditions or cell types and measure whether feature selectivity on non-HUVEC lines improves or degrades, which would delimit how universal the nuisance subspace is.","Beyond the paper: ICFL's matching-pursuit structure could transfer to other scientific modalities where a 'control condition' defines nuisance variation, such as genomics or hyperspectral imaging, using the same control-whitening trick.","Beyond the paper: because the paper shows linear directions encode biology, one could attempt causal intervention experiments—steering the feature direction for adherens junctions and checking whether reconstructed images shift toward the expected morphology—to test whether the directions are causally used by the model, not merely correlated."],"forward_implications":["Linear probes trained on ICFL reconstructions retain almost all of the biological signal of the original MAE representations for simple concepts like cell types and batch, and a substantial fraction for harder tasks like distinguishing hundreds of genetic perturbations.","ICFL features approach the selectivity of CellProfiler hand-crafted features and correlate with them strongly (Pearson coefficient 0.71), suggesting sparse directions capture morphology that experts already measure.","Token-level heatmaps aligned with a feature direction can identify which individual cells in a mixed image are perturbed versus control-like, with 87% recall on the expert-annotated set, in a supervision-free way.","ICFL avoids dead features far better than TopK SAEs and achieves higher reconstruction similarity at equal sparsity, particularly when combined with PCA whitening.","Larger models (MAE-G over MAE-L) improve ICFL feature selectivity on complex tasks, so the dictionary-learning route benefits from scaling."],"supporting_citations":[{"why":"Introduces dictionary learning and the superposition hypothesis for LLMs, the framing this paper transfers to microscopy.","marker":"Bricken et al., 2023"},{"why":"Supplies the matching-pursuit algorithm that ICFL is a batched variant of.","marker":"Mallat & Zhang, 1993"},{"why":"Provides the TopK SAE baseline method and training scheme that ICFL is compared against.","marker":"Gao et al., 2024"},{"why":"Trained the masked autoencoders on Cell Painting microscopy that supply the representations interpreted here.","marker":"Kraus et al., 2024"},{"why":"Provides the CellProfiler hand-crafted features used as the biology-informed comparison.","marker":"Carpenter et al., 2006"},{"why":"Defines the CRISPR single-gene perturbation benchmark tasks and the well-level aggregation procedure.","marker":"Celik et al., 2024"},{"why":"Supplies the RxRx1 dataset used for the siRNA perturbation task and the CellProfiler comparison.","marker":"Sypetkowski et al., 2023"},{"why":"Supplies the RxRx3 dataset used for the genetic perturbations and functional gene groups.","marker":"Fay et al., 2023"}],"fun_headline_variants":["Sparse dictionary learning exposes biology in microscopy AI","Microscopy AI's hidden concepts: cell types and edits found","Dictionary learning decodes cell types and gene edits in microscopy","Sparse features reveal biology from microscopy foundation models"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the variation present in unperturbed control cells is purely nuisance variation, and that this nuisance subspace is shared by all other cell types and perturbations, so that downweighting those directions amplifies only perturbation-relevant signal without removing any biology that distinguishes perturbations.","fun_headline_variants_meta":{"raw":{"variants":["Sparse dictionary learning exposes biology in microscopy AI","Microscopy AI's hidden concepts: cell types and edits found","Dictionary learning decodes cell types and gene edits in microscopy","Sparse features reveal biology from microscopy foundation models"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000169,"raw_usage":{"total_tokens":1221,"prompt_tokens":861,"completion_tokens":360,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":477,"completion_tokens_details":{"reasoning_tokens":295}},"tokens_in":477,"tokens_out":360,"duration_ms":4475,"temperature":1.0,"reasoning_tokens":295,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T11:24:04.630369+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same ICFL pipeline on a different microscopy dataset in which the control images come from one cell line and the perturbed images from another; if linear-probe accuracy on the non-control cell line's perturbations falls well below the unwhitened baseline while the control line's accuracy stays high, the control-derived whitening is removing cross-cell-type biological signal, and the reported generality would not transfer.","supporting_citations":[{"cited_title":"M., Kraus, O., Victors, M., Arumugam, L., Vuggu- mudi, K., Urbanik, J., Hansen, K., Celik, S., Cernek, N., Jagannathan, G., et al","cited_arxiv_id":null,"evidence_quote":"Supplies the RxRx3 dataset used for the genetic perturbations and functional gene groups."}],"review_version":1}