Unravelling Causal Genetic Biomarkers of Alzheimer's Disease via Neuron to Gene-token Backtracking in Neural Architecture: A Groundbreaking Reverse-Gene-Finder Approach

Jacqueline CK Lam; Victor OK Li; Yang Han

REVIEW 5 major objections 5 minor 1 cited by

Unravelling Causal Genetic Biomarkers of Alzheimer's Disease via Neuron to Gene-token Backtracking in Neural Architecture: A Groundbreaking Reverse-Gene-Finder Approach

T0 review · 5 major / 5 minor · reviewed 2026-08-09 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Backtracking from causal neurons to input tokens yields ten novel candidate causal genes for Alzheimer's.

desk verdict A clear, transparent interpretability pipeline for Geneformer, but the 'causal biomarker' claim rests on an unproven equivalence between attention backtracking and gene-to-neuron causality; worth serious review as hypothesis generation, not as validated discovery. read the letter →

arxiv 2502.03938 v1 pith:SI6LKGWI submitted 2025-02-06 cs.LG

Victor OK Li , Yang Han , Jacqueline CK Lam This is my paper · ORCID

classification cs.LG

keywords Alzheimer'sdiseasegeneticbiomarkerscausaltracingbacktrackinggenomicfoundationmodelsingle-celltranscriptomicsinterpretabilitygenetokens

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper aims to extract candidate causal genetic biomarkers for Alzheimer's disease (AD) from a fine-tuned genomic foundation model by working backwards from the neurons most responsible for AD prediction to the input gene tokens that activate them. The authors claim that genes most likely to cause AD are the genes most likely to activate the neurons most likely to cause AD, and they operationalize that idea by masking known AD genes, locating the most causal neurons via causal tracing, and backtracking through self-attention weights to score every input gene token. The top ten genes produced by this procedure are PLXDC2, MT-CO3, DOCK4, ARHGAP24, MEF2A, RUNX1, ITPR2, FOXN3, MT-CO1, and SORL1, none of which are in the known AD gene list used for masking. If these candidates hold up, the method would provide an interpretable route from a black-box model to prioritized, testable genetic hypotheses about disease onset.

What carries the argument

The central machinery is the Reverse-Gene-Finder backtracking algorithm, which computes a score for the $i$-th unit at layer $l$ as $s_i^l = \sum_j W_{i,j}^l \left( IE(h_j^{l+1}) + s_j^{l+1} \right)$, where $IE$ is the indirect causal effect of a most causal neuron measured by causal tracing and $W$ is the self-attention weight between units in adjacent layers. The recurrence is evaluated from the last layer to the input layer, so each input gene token accumulates the weighted causal influence of all MCNs it can reach. The load-bearing identity is the assumed equivalence behind the method: the most causal genes are exactly those whose tokens most strongly activate the most causal neurons. That equivalence turns a neural-network interpretability question into a biological causation claim.

What would settle it

Run the same Reverse-Gene-Finder procedure on an independent AD single-cell dataset or on a different microglial cluster from the same cohort and check whether the top-10 most causal genes reproduce; if the list changes entirely, the top-10 result is an artifact of the specific model and data subset rather than a robust causal signal. A complementary check is to mask the identified top-10 genes and test whether the fine-tuned model's AD prediction shifts significantly beyond what random gene masking produces.

Watch

Extended reading notes

Core claim

The paper's central discovery is a procedure that turns a fine-tuned genomic foundation model for AD classification into a reverse gene finder. The authors posit that the genes with the highest probability of causing AD, which they call the most causal genes (MCGs), must have the highest probability of activating the neurons with the highest probability of causing AD, the most causal neurons (MCNs). Using causal tracing, they perturb the embeddings of ten known AD-associated genes, measure the indirect effect of each hidden state on the AD prediction, and select the top hidden states as MCNs. They then compute, for each input position, a cumulative influence score by summing MCN effects weighted by self-attention weights layer by layer back to the input, and the input gene tokens with the highest averaged scores are designated the most causal tokens and their genes the MCGs. The resulting top-10 list is presented as novel causal genetic biomarkers for AD, supported by literature references and by pathway enrichment that flags Alzheimer's, oxidative phosphorylation, and related neurodegenerative and metabolic pathways.

Load-bearing premise

The load-bearing premise is that the strength with which an input gene token activates the model's most influential neurons is a faithful ranking of that gene's real-world causal role in Alzheimer's disease.

Editorial extensions

If this is right

The top-10 MCGs (PLXDC2, MT-CO3, DOCK4, ARHGAP24, MEF2A, RUNX1, ITPR2, FOXN3, MT-CO1, SORL1) become prioritized candidates for experimental validation in AD models.
The same neuron-to-token backtracking can be applied to other diseases by fine-tuning a genomic foundation model on a different phenotype and backtracking from that phenotype's most causal neurons.
Because most causal neurons concentrate in the early and middle layers of the model, the method suggests that causal signal for AD is distributed across low- and mid-level genetic interaction features rather than only in the final prediction layer.
The enriched pathways (Alzheimer's, Parkinson's, Huntington's, oxidative phosphorylation, thermogenesis) indicate that the candidate genes point to shared bioenergetic and neuroinflammatory mechanisms across neurodegenerative diseases.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

If the equivalence between model activation and biological causation is only approximate, the top-10 list is better read as a ranked set of hypotheses than as established causal genes; the paper's own definition of 'causal' as 'most probable from a statistical perspective' already leans toward this weaker reading.
The procedure uses a single microglial cluster from one public cohort, so a natural stress test is to run the same backtracking on other clusters and other AD cohorts; a stable top-10 list across those settings would substantially strengthen the case that the genes are disease-relevant rather than cluster-specific.
The backtracking score depends on a 95th-percentile cutoff for selecting most causal neurons and on the noise level used for perturbations; probing how sensitive the top-10 list is to those hyperparameters would tell future users how much weight to place on the exact ranking.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

You should know this before reading it: the paper is a well-written, transparent application of causal tracing (Meng et al.) to a genomic foundation model for Alzheimer's disease, but the headline claim—that it discovers novel causal genetic biomarkers—is not supported by the evidence. The method is a plausible hypothesis generator, not a validated discovery.

What's actually new: the gene-token input representation and the backtracking from MCNs to input tokens via self-attention weights. Adapting causal tracing from LLMs to single-cell gene expression is a legitimate extension, and the authors are honest about their data, code, and hyperparameters. The fine-tuned Geneformer reaches 74.67% AUC, which is modest but reasonable for genetic-only data.

The soft spots are serious. The load-bearing postulate in §2.2.1—that MCGs must have the highest probability of activating MCNs—is asserted, not demonstrated. Self-attention weights measure learned associations, not causal effects. The causal tracing step intervenes only on known AD genes to locate MCNs; it never intervenes on the candidate MCGs to test whether perturbing them changes MCN activation or AD prediction. So the final gene ranking is essentially a correlation-based artifact of the model and the chosen data subset (one microglial cluster). The validation is a post-hoc literature review, and several top-10 genes (SORL1, RUNX1, MT-CO1) already appear in the AD literature cited elsewhere in the paper. There are no negative controls, no error bars on gene scores, and the cross-validation may leak subject-level information because single cells from the same subject can appear in both train and test folds.

The paper itself, in §4.4, redefines 'causal' as 'most probable'—an honest but significant downgrade. The reversal test suggested in the stress-test would make the causal claim empirically checkable, and its absence is the core gap.

Who this is for: readers working on interpretability of genomic foundation models. It's a useful demonstration that causal tracing can be adapted to this setting, and the code and experimental detail support reproduction. But it should not be read as a credible causal gene discovery without functional validation or at least a reversal test.

My recommendation: I would send this to peer review despite my skepticism. The methodological transparency and the potential value as a hypothesis-generation tool justify referee time. I expect reviewers to push for interventional validation or a reframing away from 'causal' language.

Referee Report

5 major / 5 minor

Summary. The manuscript proposes Reverse-Gene-Finder, an interpretability pipeline applied to Geneformer for Alzheimer's disease (AD) gene discovery. The authors fine-tune a 12-layer Geneformer on single-cell microglial RNA-seq data from ROSMAP for early-AD versus non-AD classification, use causal tracing (masking ten known AD genes) to identify 'most causal neurons' (MCNs), and then backtrack through self-attention weights to score input gene tokens. The top-10 scored genes (PLXDC2, MT-CO3, DOCK4, ARHGAP24, MEF2A, RUNX1, ITPR2, FOXN3, MT-CO1, SORL1) are presented as novel causal genetic biomarkers of AD, with validation claimed through literature review and KEGG pathway enrichment. The paper positions the 'must' postulate connecting MCGs to MCNs as the central methodological novelty.

Significance. If the causal claim were supported, the approach would be a useful hypothesis-generation tool for AD and other diseases. The paper has concrete strengths: it ships a public code repository, builds on a publicly available foundation model (Geneformer), uses a public single-cell dataset, and reports cross-validated classification performance. However, the evidence presented does not establish the causal interpretation. The core postulate equating biological causation with neuron activation is unproven; the candidate genes are never intervened on; the ranking is derived from model-internal attention; and the reported gene scores come without uncertainty quantification or negative controls. The computational pipeline is coherent, but the central claim of 'novel causal genetic biomarkers' is not supported by the current experiments.

major comments (5)

[§2.2.1] The load-bearing postulate—that MCGs 'must have the highest probability of activating' MCNs—is asserted as an observation but is neither derived nor empirically tested. The manuscript provides no link between biological causation of AD and activation of a hidden state in a fine-tuned transformer. Since Algorithm 1's scores and the resulting top-10 list are constructed entirely on this equivalence, the central claim collapses if the postulate fails. A reversal test (e.g., masking the top-ranked candidate and measuring whether MCN activation or AD prediction changes) is needed to make the claim empirically checkable.
[§3.2 and Algorithm 1] The causal tracing step intervenes only on ten known AD genes to locate MCNs; the candidate MCGs identified in §4.1 are never perturbed. The backtracking recurrence uses self-attention weights as 'model weights,' but in a transformer, attention weights are per-layer, per-head, normalized coefficients over tokens, not causal connections from a neuron at one layer to the input. The recurrence therefore propagates a model-internal association measure, not a counterfactual effect of a gene on an MCN. Without an intervention on the candidates, labeling them 'causal' is not justified.
[§3.1.1, §3.2.1, §4.1] The pipeline contains several under-analyzed free choices: the single microglial cluster (cluster 10 from the prefrontal cortex), the early-AD-only split, the 95th percentile MCN cutoff, the Gaussian noise level (set to 1), the 256-token truncation, and the top-10 cutoff. No sensitivity or robustness analysis is reported for any of these choices, and the gene scores in §4.1 are presented without standard deviations or confidence intervals across the five cross-validation folds. The top-10 list may therefore be an artifact of these choices rather than a stable discovery.
[§4.2–§4.3] The validation consists of citing prior literature for each gene and reporting enriched KEGG pathways. A literature review cannot distinguish a true causal gene from a well-known correlated one, and the pathway analysis on ten genes is post-hoc and vulnerable to selection bias. There are no negative controls (e.g., genes not believed to be AD-related), no permutation test on the ranking, no independent cohort replication, and no functional or interventional follow-up. Consequently, the abstract and §4.1 claim of 'novel causal genetic biomarkers' is unsupported by the evidence presented.
[§4.4] The paper introduces its definition of causality only in the discussion: 'This study defines "causal" as "the most probable" genetic biomarkers from a statistical perspective.' This is not the sense of causality used in the title and abstract, and it is considerably weaker than the counterfactual interpretation that the causal-tracing and 'intervention' language suggests. The manuscript should either operationalize causal claims with actual interventions or explicitly limit the contribution to hypothesis ranking, which would avoid overclaiming.

minor comments (5)

[Abstract] The abstract contains a typo: 'ave-nue' should be 'avenue'.
[§4.1] The text reads 'fine-turned model performance'; this should be 'fine-tuned model performance'.
[§3.2.1] Section number 3.2.1 is used twice: once for 'Known Genes Related to AD' and once for 'Use of Causal Tracing Analysis'. The second subsection should be renumbered (e.g., 3.2.2).
[Algorithm 1] The phrase 'interconnected MCNs in the next layer' is ambiguous because self-attention operates within a layer, not between layers. Please clarify whether MCNs are selected across all layers and how the recurrence maps from layer l+1 to layer l.
[Figure 2] Figure 2 shows the locations of MCNs, but the axes and the marker encoding are not defined. Please specify what the x- and y-axes represent and how the 95th percentile cutoff is reflected in the plot.

Circularity Check

2 steps flagged · score 6.0 of 10

The paper defines MCGs as the genes most likely to activate MCNs, then Algorithm 1 returns exactly that attention-weighted activation score, so the top-10 'causal genes' are the definition's output rather than an independently tested discovery.

self definitional [Section 2.2.1 (Discovery of the Most Causal Genes and Neurons); Algorithm 1; Section 3.3 (Identification of MCTs and MCGs)]
"We exploit the observation that genes with the highest probability of causing AD, defined as the MCGs, must have the highest probability of activating those neurons with the highest probability of causing AD, defined as the MCNs."

The definition makes MCG equivalent to 'token that most activates MCNs.' Algorithm 1 computes s_i^l = sum_j W_{i,j}^l * (IE(h_j^{l+1}) + s_j^{l+1}), which is exactly an attention-weighted activation score propagated back from MCNs. Section 3.3 then states: 'The gene tokens with the highest scores are identified as MCTs... The MCGs corresponding to the identified MCTs are selected as the putative causal genetic biomarkers for AD.' Thus the reported top-10 MCGs (PLXDC2, MT-CO3, DOCK4, etc.) are produced by evaluating the definitional criterion, not by an independent causal test; no intervention is performed on these candidate genes to verify that perturbing them changes MCN activity or AD prediction.
other [Section 4.4 (Insights into Causal Biomarker Discovery)]
"This study defines 'causal' as 'the most probable' genetic biomarkers from a statistical perspective."

The paper explicitly redefines 'causal' as 'the most probable' and then uses that redefinition to call the attention/backtracking scores causal. Since Section 2.2.1 already equated MCGs with genes most probably activating MCNs, the phrase 'novel causal genetic biomarkers' is a terminological relabeling of the highest-scoring model tokens, not a causal relationship established by intervening on the candidate genes. The subsequent literature review and pathway enrichment are post hoc and do not test the causal claim.

full rationale

The derivation chain has a genuine empirical component at the MCN-selection step: known AD genes are corrupted, hidden states are restored, and indirect effects are measured with causal tracing, so the MCNs are not purely definitional. The circularity enters at the MCG step. The paper defines MCGs as the genes with the highest probability of activating MCNs, and its backtracking Algorithm 1 computes exactly that quantity as a weighted sum of MCN indirect effects using the model's self-attention weights. Selecting the top-scoring tokens and announcing them as 'novel causal genetic biomarkers' restates the definitional postulate rather than testing it. The paper itself confirms this by defining 'causal' as 'the most probable' in Section 4.4, a statistical-importance notion rather than an interventional causal effect on the candidate genes. A reversal test (perturbing the top MCGs and checking MCN activation or AD prediction) would be needed to break the circle; none is reported. The post-hoc literature validation and STRING pathway enrichment are external support but are after-the-fact associations, not independent predictions from the model. There is no load-bearing self-citation chain; citations to Geneformer, causal tracing, and prior AD genetics are appropriate external sources. Overall, the central discovery claim reduces to the MCG definition plus the model's learned attention weights, so a moderate-to-high circularity score is warranted.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

The method's output depends on several hand-chosen thresholds and on the unproven equivalence between model attention and biological causation. The most important free parameters are the MCN cutoff, noise level, token limit, top-k count, and the selected microglial cluster. The axioms are the asserted equivalences and redefinitions that make the causal claim unfalsifiable without external validation.

free parameters (6)

MCN selection threshold = 95th percentile of indirect effects
Chosen by hand; determines which neurons count as most causal and therefore drives backtracking results.
Gaussian noise level for perturbation = 1
Chosen by hand for causal tracing; no sensitivity analysis is reported.
Maximum input gene tokens = 256
Set to reduce computational cost; truncation of top and bottom tokens may affect which genes can be discovered.
Number of top MCT/MCG candidates = 10
Arbitrary cutoff defining the final gene list; no stability analysis across other cutoffs.
Fine-tuning hyperparameters = selected via HyperOpt (values in code)
Fitted to validation F1; the paper states they are in the code but does not list them.
Microglial cluster used for analysis = cluster 10
Selected because it showed more significant transcriptional changes in early AD; a post-hoc data selection that shapes the gene list.

assumptions (5)

ad hoc to paper Genes with the highest probability of causing AD must have the highest probability of activating the neurons with the highest probability of causing AD.
This is the paper's central postulate (Section 2.2.1) and is asserted without proof; it equates biological causation with model neuron activation.
domain assumption Perturbing known AD gene embeddings and measuring indirect effects on the model output identifies neurons causally related to AD.
Causal tracing (Section 3.2.2) is applied to a fine-tuned black-box model; model-internal causation is assumed to reflect disease causation.
domain assumption Self-attention weights measure causal interconnection strength between neurons and input tokens.
Algorithm 1 uses attention weights W as causal weights; attention is correlation, not causation.
domain assumption Single-cell gene expression rank values represent causal gene activity for AD.
The input is expression from one microglial cluster at one time point; expression level is not necessarily causal for AD onset.
ad hoc to paper Causal can be redefined as 'most probable from a statistical perspective'.
Section 4.4 explicitly narrows 'causal' to 'most probable', which is weaker than the causal claim in the title.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Unravelling Causal Genetic Biomarkers of Alzheimer's Disease via Neuron to Gene-token Backtracking in Neural Architecture: A Groundbreaking Reverse-Gene-Finder Approach." pith.science (2026). https://pith.science/paper/SI6LKGWI

@misc{pith2026250203938,
  author       = {Pith},
  title        = {Pith review of: Unravelling Causal Genetic Biomarkers of Alzheimer's Disease via Neuron to Gene-token Backtracking in Neural Architecture: A Groundbreaking Reverse-Gene-Finder Approach},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SI6LKGWI}},
  note         = {Machine review of arXiv:2502.03938}
}

read the original abstract

Alzheimer's Disease (AD) affects over 55 million people globally, yet the key genetic contributors remain poorly understood. Leveraging recent advancements in genomic foundation models, we present the innovative Reverse-Gene-Finder technology, a ground-breaking neuron-to-gene-token backtracking approach in a neural network architecture to elucidate the novel causal genetic biomarkers driving AD onset. Reverse-Gene-Finder comprises three key innovations. Firstly, we exploit the observation that genes with the highest probability of causing AD, defined as the most causal genes (MCGs), must have the highest probability of activating those neurons with the highest probability of causing AD, defined as the most causal neurons (MCNs). Secondly, we utilize a gene token representation at the input layer to allow each gene (known or novel to AD) to be represented as a discrete and unique entity in the input space. Lastly, in contrast to the existing neural network architectures, which track neuron activations from the input layer to the output layer in a feed-forward manner, we develop an innovative backtracking method to track backwards from the MCNs to the input layer, identifying the Most Causal Tokens (MCTs) and the corresponding MCGs. Reverse-Gene-Finder is highly interpretable, generalizable, and adaptable, providing a promising avenue for application in other disease scenarios.

Figures

Figures reproduced from arXiv: 2502.03938 by the authors.

**Figure 1.** summarizes the overarching Reverse-Gene-Finder mechanism. Firstly, we fine-tune a genomic foundation model specifically for AD classification, utilizing singlecell gene expression data from AD patients and healthy control subjects. Secondly, by modifying the input data by masking out known genes strongly associated with AD, we systematically identify MCNs related to AD, employing causal tracing techniques to obser… view at source ↗

**Figure 2.** Locations of the Most Causal Neurons (Based on [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

Reverse-Speech-Finder: A Neural Network Backtracking Architecture for Generating Alzheimer's Disease Speech Samples and Improving Diagnosis Performance
cs.LG 2025-05 conditional novelty 4.0 of 10

RSF uses causal tracing and backtracking in a fine-tuned LLM to identify 'most probable' AD speech markers and generates synthetic transcripts that improve AD classification by about 3.5% accuracy.

Reference graph

Works this paper leans on

3 extracted references · 2 canonical work pages · cited by 1 Pith paper

[3]

2023) and a large genomic foundation model (namely, Geneformer (Theodoris et al

Experimental Setting This section details the experimental setup, using a single-cell genetic dataset from a comprehensive AD case/control study (Sun et al. 2023) and a large genomic foundation model (namely, Geneformer (Theodoris et al. 2023)) to demonstrate the proposed Reverse-Gene-Finder technology. Algorithm 1: Reverse-Gene-Finder Backtracking Method...

work page 2023
[4]

causal” as “the most probable

Results and Discussion 4.1 Discovery of Previously Unknown Genetic Candidates The AUC score of the fine-tuned Geneformer model based on five-fold validation is 74.67% ± 6.74%. The sensitivity and specificity scores are 62.68% ± 11.43% and 73.22% ± 16.01%, respectively. It should be noted that the focus of our study is more about demonstrating a new approa...

work page 2017
[5]

Future work can also extend the proposed Reverse-Gene-Finder technology to other AD biomarker identification sce-narios

Future Work While the current work focuses on identifying novel causal biomarkers, integrating these biomarkers into predictive models to enhance AD prediction can be done in the future. Future work can also extend the proposed Reverse-Gene-Finder technology to other AD biomarker identification sce-narios. For instance, by backtracking language models, th...

arXiv 2020

Pith tools

Reviewed August 9, 2026 · model on record in the stance chip above.

[1] [3]

2023) and a large genomic foundation model (namely, Geneformer (Theodoris et al

Experimental Setting This section details the experimental setup, using a single-cell genetic dataset from a comprehensive AD case/control study (Sun et al. 2023) and a large genomic foundation model (namely, Geneformer (Theodoris et al. 2023)) to demonstrate the proposed Reverse-Gene-Finder technology. Algorithm 1: Reverse-Gene-Finder Backtracking Method...

work page 2023

[2] [4]

causal” as “the most probable

Results and Discussion 4.1 Discovery of Previously Unknown Genetic Candidates The AUC score of the fine-tuned Geneformer model based on five-fold validation is 74.67% ± 6.74%. The sensitivity and specificity scores are 62.68% ± 11.43% and 73.22% ± 16.01%, respectively. It should be noted that the focus of our study is more about demonstrating a new approa...

work page 2017

[3] [5]

Future work can also extend the proposed Reverse-Gene-Finder technology to other AD biomarker identification sce-narios

Future Work While the current work focuses on identifying novel causal biomarkers, integrating these biomarkers into predictive models to enhance AD prediction can be done in the future. Future work can also extend the proposed Reverse-Gene-Finder technology to other AD biomarker identification sce-narios. For instance, by backtracking language models, th...

arXiv 2020