iLoRA: Bayesian Low-Rank Adaptation with Latent Interaction Graphs for Microbiome Diagnosis

Haizhou Shi; Hao Wang; Hengguan Huang; Lingfa Meng; Samir Bhatt; Tongyuan Hu; Yang Song; Yixuan Zhang

arxiv: 2605.30179 · v1 · pith:BEDJJ5VNnew · submitted 2026-05-28 · 💻 cs.LG · cs.AI

iLoRA: Bayesian Low-Rank Adaptation with Latent Interaction Graphs for Microbiome Diagnosis

Yang Song , Yixuan Zhang , Lingfa Meng , Tongyuan Hu , Haizhou Shi , Hao Wang , Samir Bhatt , Hengguan Huang This is my paper

Pith reviewed 2026-06-29 08:25 UTC · model grok-4.3

classification 💻 cs.LG cs.AI

keywords LoRABayesian adaptationlatent interaction graphsmicrobiome diagnosisparameter-efficient fine-tuninggraph-conditioned updatesIBD prediction

0 comments

The pith

iLoRA infers a latent interaction graph from microbiome inputs to condition Bayesian LoRA updates and learn predictions jointly with structure.

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

The paper presents iLoRA as a framework that infers a latent interaction graph directly from input data and uses it to produce input-specific low-rank adaptation updates in a Bayesian setting. This setup lets the model optimize both the diagnosis task and the recovery of interaction structure in one process instead of separating them. In microbiome applications, where labels depend on species abundances and cross-species interactions, the joint training yields higher accuracy on IBD diagnosis across cohorts and produces graphs that match human annotations in QA evaluations. The Bayesian component also supplies uncertainty estimates at modest extra cost.

Core claim

iLoRA is the first Bayesian graph-conditioned LoRA framework that infers a latent interaction graph from the input and uses it to generate input-conditioned LoRA updates, learning prediction and latent interaction structure jointly rather than training a predictor and applying interaction analysis only post hoc.

What carries the argument

The latent interaction graph inferred from input data, which conditions the generation of Bayesian LoRA updates for the downstream task.

If this is right

iLoRA improves diagnosis performance over strong LoRA and Bayesian adaptation baselines on multi-cohort IBD tasks.
Recovered graphs align with human annotations in interactive QA settings and with known cohort-level microbiome associations.
The model supplies calibrated uncertainty estimates while adding only moderate overhead from the graph branch.
Joint optimization removes the need for separate post-hoc interaction analysis after predictor training.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The same graph-conditioning idea could be tested on other high-dimensional biological data where interactions are hypothesized to drive outcomes, such as single-cell or metabolomic profiles.
Calibrated uncertainty from the Bayesian component might support downstream uses like selective prediction or active learning in diagnostic pipelines.
If the inferred graphs capture causal cross-talk, they could serve as hypotheses for targeted follow-up experiments in microbial ecology.

Load-bearing premise

A meaningful latent interaction graph can be inferred directly from input microbiome data in a way that reliably conditions and improves the LoRA updates for the diagnosis task.

What would settle it

Apply iLoRA to held-out microbiome cohorts and observe no gain in diagnosis accuracy over plain LoRA baselines or no alignment between recovered graphs and independent human or cohort-level annotations.

Figures

Figures reproduced from arXiv: 2605.30179 by Haizhou Shi, Hao Wang, Hengguan Huang, Lingfa Meng, Samir Bhatt, Tongyuan Hu, Yang Song, Yixuan Zhang.

**Figure 1.** Figure 1: Framework overview of iLoRA. Given a microbiome abundance profile X ∈ R M, our model predicts diagnosis with an LLM while, in parallel, selecting K < M key taxa to infer a latent interaction graph. We first infer a Poisson edge graph, then transform it into a sparse graph with Laplace-distributed edge weights to encourage sparsity, embed the graph with a GNN, and use the embedding to generate the LoRA matr… view at source ↗

**Figure 2.** Figure 2: Visualization of latent interaction graphs. Left: Groundtruth adjacency matrix. Right: Inferred graph by iLoRA. The top example (Sample 1371) shows thread disentanglement, while the bottom (Sample 2568) shows discourse chain recovery [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗

**Figure 3.** Figure 3: Chord diagram visualizations for two representative samples from the Kumbhari 2024 cohort. For iLoRA’s sample-level graphs, we symmetrize directed scores and select the top 10% of edges per sample (K = 19) ranked by descending weight. We then evaluate these edges against the global significant set of 41 taxon pairs. As shown in [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗

**Figure 4.** Figure 4: Heatmap of microbe–microbe association (ii) Cohort-level pair → y association: log-ratio + logistic + BH–FDR (Fig. 6b). To characterize compositional contrasts associated with diagnosis, we construct, for each taxon pair (a, b), a log-ratio feature sab,i = log(xia + ϵ) − log(xib + ϵ), (25) 22 [PITH_FULL_IMAGE:figures/full_fig_p022_4.png] view at source ↗

**Figure 5.** Figure 5: Pair-Y (CD vs UC): conditional log-ratio network (q¡0.05) (iii) Definition and role of the intersection reference set EGT. The two screening procedures yield edge sets ESpear and Eratio, respectively. We define their intersection as the cohort-level statistical reference set: EGT = ESpear ∩ Eratio. (29) In our dataset, |EGT| = 41. At the chosen FDR threshold, this set collects taxon pairs that have both co… view at source ↗

read the original abstract

Parameter-efficient adaptation has made LLMs practical for domain prediction, but standard LoRA still relies on a static low-rank update and does not expose the latent interactions that often drive scientific labels. We introduce iLoRA. To our knowledge, it is the first Bayesian graph-conditioned LoRA framework. It infers a latent interaction graph from the input and uses it to generate input-conditioned LoRA updates. As a result, iLoRA learns prediction and latent interaction structure jointly, rather than training a predictor and applying interaction analysis only post hoc. We instantiate this idea for microbiome diagnosis, where disease state can depend on both species-level abundance and microbe-microbe cross-talk, and evaluate it in two complementary settings: interactive QA with human-annotated graphs, which tests latent structure recovery, and multi-cohort IBD diagnosis, which tests biomedical utility. Across both settings, iLoRA improves over strong LoRA and Bayesian adaptation baselines, recovers graphs aligned with human annotations and cohort-level microbiome associations, and provides calibrated uncertainty with moderate graph-branch overhead.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

iLoRA adds a latent graph inference step to Bayesian LoRA for microbiome tasks, but the abstract alone gives no evidence the joint learning works or improves anything.

read the letter

The core claim is that iLoRA infers a latent interaction graph from microbiome inputs and uses it to produce input-conditioned LoRA updates inside a Bayesian setup, so that prediction and graph structure are learned together rather than in sequence. That is the only concrete novelty on offer.

The framing makes sense for the domain. Microbiome diagnosis can depend on species interactions, and standard LoRA is static, so conditioning the adapter on a recovered graph is a logical next step. The two test settings—one checking recovery against human-annotated graphs, one checking IBD diagnosis across cohorts—are also reasonable ways to separate structure recovery from task performance.

Everything else is thin. The description stops at the abstract level, with no equations, no training details, no numbers on improvement size, and no discussion of how the graph is actually inferred or how it modulates the low-rank matrices. The central assumption—that the inferred graph supplies useful conditioning that beats plain Bayesian LoRA—remains unexamined. Without the methods or results, there is no way to tell whether the graph recovery is meaningful or whether the reported gains are real.

This is for people already working on graph-augmented adapters or microbiome ML who want to see one possible direction. A reader looking for a worked-out method or reproducible results will not find enough here. The paper should go to peer review only if the full manuscript supplies the missing technical sections and experiments; the abstract by itself does not justify referee time.

Referee Report

2 major / 1 minor

Summary. The manuscript introduces iLoRA as the first Bayesian graph-conditioned LoRA framework. It infers a latent interaction graph from microbiome inputs to generate input-conditioned LoRA updates, enabling joint learning of prediction and latent structure rather than post-hoc analysis. The approach is instantiated for microbiome diagnosis and evaluated in two settings: interactive QA with human-annotated graphs (for structure recovery) and multi-cohort IBD diagnosis (for biomedical utility). Claims include improvements over strong LoRA and Bayesian adaptation baselines, recovery of graphs aligned with human annotations and cohort associations, and provision of calibrated uncertainty at moderate overhead.

Significance. If the central claims hold under detailed scrutiny, the work could meaningfully extend parameter-efficient adaptation methods to scientific domains where input-dependent latent structures (such as microbe-microbe interactions) matter. The joint Bayesian inference of graph and predictor, along with explicit uncertainty calibration, would distinguish it from standard LoRA if empirically supported; the dual evaluation settings (annotation alignment and clinical utility) are a constructive design choice.

major comments (2)

[Abstract] Abstract: The central claim that inferring and conditioning on a latent interaction graph improves both prediction and graph recovery rests on an unelaborated mechanism; without the specific form of the graph inference (e.g., variational posterior, prior, or conditioning operator on the LoRA factors), it is impossible to determine whether the graph step supplies non-redundant structure or reduces to standard input-dependent adaptation.
[Abstract] Abstract: The assertion of joint learning 'rather than training a predictor and applying interaction analysis only post hoc' is load-bearing for novelty, yet the abstract provides no indication of how the graph posterior is optimized jointly with the diagnosis loss or whether the graph variables are marginalized in a way that avoids circular dependence on the predictor.

minor comments (1)

The phrase 'to our knowledge, it is the first' requires a dedicated related-work paragraph with explicit comparisons to prior graph-augmented adaptation or Bayesian LoRA variants to be convincing.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the detailed comments on the abstract. We address each point below. The abstract is intentionally concise, but we agree it can be improved to better indicate the inference mechanism and joint optimization; we will revise it accordingly while preserving brevity.

read point-by-point responses

Referee: [Abstract] Abstract: The central claim that inferring and conditioning on a latent interaction graph improves both prediction and graph recovery rests on an unelaborated mechanism; without the specific form of the graph inference (e.g., variational posterior, prior, or conditioning operator on the LoRA factors), it is impossible to determine whether the graph step supplies non-redundant structure or reduces to standard input-dependent adaptation.

Authors: The abstract provides a high-level summary of the approach. The full manuscript specifies the form of the graph inference, including the variational posterior, prior, and conditioning operator on the LoRA factors. This mechanism ensures the inferred graph supplies non-redundant structure for the adaptation. We will revise the abstract to briefly elaborate on the graph inference mechanism. revision: yes
Referee: [Abstract] Abstract: The assertion of joint learning 'rather than training a predictor and applying interaction analysis only post hoc' is load-bearing for novelty, yet the abstract provides no indication of how the graph posterior is optimized jointly with the diagnosis loss or whether the graph variables are marginalized in a way that avoids circular dependence on the predictor.

Authors: The full manuscript indicates how the graph posterior is optimized jointly with the diagnosis loss through a unified variational objective that marginalizes the graph variables, avoiding circular dependence. We will revise the abstract to provide an indication of this joint optimization process. revision: yes

Circularity Check

0 steps flagged

No significant circularity

full rationale

The provided abstract and description present iLoRA as a novel joint Bayesian graph-conditioned LoRA framework that infers latent interactions from inputs to produce conditioned updates, with evaluation on graph recovery and diagnosis tasks. No equations, self-citations, or derivation steps are visible that reduce any claimed prediction or result to a fitted parameter or prior self-referential definition by construction. The central claims rest on the joint learning setup and empirical improvements over baselines, which remain independent of the inputs described.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 1 invented entities

Based on abstract only, the central claim rests on standard Bayesian inference assumptions and the existence of recoverable latent interaction structure in microbiome data; no specific free parameters or invented entities beyond the latent graph itself can be identified.

axioms (1)

standard math Standard assumptions of Bayesian modeling and low-rank adaptation hold for the microbiome domain
The framework builds directly on Bayesian methods and LoRA without stating deviations.

invented entities (1)

Latent interaction graph no independent evidence
purpose: To condition input-dependent LoRA updates by capturing microbe-microbe interactions
Introduced as a core component of iLoRA; no independent evidence of its recoverability is provided in the abstract.

pith-pipeline@v0.9.1-grok · 5733 in / 1313 out tokens · 27161 ms · 2026-06-29T08:25:22.571678+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

8 extracted references · 8 canonical work pages · 2 internal anchors

[1]

LERD: Latent Event-Relational Dynamics for Neurodegenerative Classification

doi: 10.1136/gutjnl-2012-302578. URL https: //doi.org/10.1136/gutjnl-2012-302578. Faust, K. and Raes, J. Microbial interactions: from networks to models.Nature Reviews Microbiology, 10(8):538– 550, 2012. doi: 10.1038/nrmicro2832. URL https: //doi.org/10.1038/nrmicro2832. Feng, Y ., Chen, H., Jia, Z., Bhatt, S., and Huang, H. LERD: Latent event-relational ...

work page internal anchor Pith review Pith/arXiv arXiv doi:10.1136/gutjnl-2012-302578 2012
[2]

URL https: //doi.org/10.1038/nature23273

doi: 10.1038/nature23273. URL https: //doi.org/10.1038/nature23273. Houlsby, N., Giurgiu, A., Jastrzebski, S., Morrone, B., de Laroussilhe, Q., Gesmundo, A., Attariyan, M., and Gelly, S. Parameter-efficient transfer learning for NLP. InProceedings of the 36th International Conference on Machine Learning, volume 97 ofProceedings of Machine Learning Researc...

work page doi:10.1038/nature23273 2019
[3]

Huang, H., Shen, X., Hao, G.-Y ., Wang, S., Meng, L., Liu, D., Duchene, D

URL https://papers.nips.cc/paper _files/paper/2022/hash/e9e1a0abc1a5b 19a4aeb80dab19c82ae-Abstract-Confere nce.html. Huang, H., Shen, X., Hao, G.-Y ., Wang, S., Meng, L., Liu, D., Duchene, D. A., Wang, H., and Bhatt, S. Bayesagent: Bayesian agentic reasoning under uncertainty via verbal- ized probabilistic graphical modeling. InProceedings of the AAAI Con...

work page doi:10.1186/s12866-0 2022
[4]

doi: 10.1109/TKDE.2016.2606

ISSN 1041-4347. doi: 10.1109/TKDE.2016.2606

work page doi:10.1109/tkde.2016.2606 2016
[5]

URL https://doi.org/10.1109/TKDE.2 016.2606428. Wang, H. and Yeung, D.-Y . A survey on bayesian deep learning.ACM Comput. Surv., 53(5), September 2020. ISSN 0360-0300. doi: 10.1145/3409383. URL https: //doi.org/10.1145/3409383. Wang, H., Shi, X., and Yeung, D.-Y . Natural-parameter networks: A class of probabilistic neural networks. In Proceedings of the ...

work page doi:10.1109/tkde.2 2020
[6]

Qwen3 Technical Report

doi: 10.1038/ismej.2015.235. URL https: //doi.org/10.1038/ismej.2015.235. Yang, A., Li, A., Yang, B., Zhang, B., Hui, B., Zheng, B., Yu, B., Gao, C., Huang, C., Lv, C., Zheng, C., Liu, D., Zhou, F., Huang, F., Hu, F., Ge, H., Wei, H., Lin, H., Tang, J., Yang, J., Tu, J., Zhang, J., Yang, J., Yang, J., Zhou, J., Zhou, J., Lin, J., Dang, K., Bao, K., Yang, ...

work page internal anchor Pith review Pith/arXiv arXiv doi:10.1038/ismej.2015.235 2015
[7]

+” root is strictly positive while the “−

URL https://iclr.cc/virtual/2024 /poster/19071. Zheng, J., Sun, Q., Zhang, M., Liu, C., Su, Q., Zhang, L., Xu, Z., Lu, W., Ching, J., Tang, W., Cheung, C. P., Hamilton, A. L., O’Brien, A. L. W., Wei, S. C., Bernstein, C. N., Rubin, D. T., Chang, E. B., Morrison, M., Kamm, M. A., Chan, F. K. L., Zhang, J., and Ng, S. C. Non- invasive, microbiome-based diag...

work page doi:10.1038/s41591-024-03280-4 2024
[8]

yes”/“no

as the backbone. For theIBD Diagnosistask, we utilizedQwen3-8B(Yang et al., 2025). •LoRA Configuration: – Molweni:We applied LoRA to projection layers ( q proj, k proj, v proj, o proj) with rank r= 8 , scaling factorα= 16, and dropout0.05. – IBD Diagnosis:We applied LoRA to the same projection layers but with rank r= 16 and α= 32 . Additionally, we set th...

work page arXiv 2025

[1] [1]

LERD: Latent Event-Relational Dynamics for Neurodegenerative Classification

doi: 10.1136/gutjnl-2012-302578. URL https: //doi.org/10.1136/gutjnl-2012-302578. Faust, K. and Raes, J. Microbial interactions: from networks to models.Nature Reviews Microbiology, 10(8):538– 550, 2012. doi: 10.1038/nrmicro2832. URL https: //doi.org/10.1038/nrmicro2832. Feng, Y ., Chen, H., Jia, Z., Bhatt, S., and Huang, H. LERD: Latent event-relational ...

work page internal anchor Pith review Pith/arXiv arXiv doi:10.1136/gutjnl-2012-302578 2012

[2] [2]

URL https: //doi.org/10.1038/nature23273

doi: 10.1038/nature23273. URL https: //doi.org/10.1038/nature23273. Houlsby, N., Giurgiu, A., Jastrzebski, S., Morrone, B., de Laroussilhe, Q., Gesmundo, A., Attariyan, M., and Gelly, S. Parameter-efficient transfer learning for NLP. InProceedings of the 36th International Conference on Machine Learning, volume 97 ofProceedings of Machine Learning Researc...

work page doi:10.1038/nature23273 2019

[3] [3]

Huang, H., Shen, X., Hao, G.-Y ., Wang, S., Meng, L., Liu, D., Duchene, D

URL https://papers.nips.cc/paper _files/paper/2022/hash/e9e1a0abc1a5b 19a4aeb80dab19c82ae-Abstract-Confere nce.html. Huang, H., Shen, X., Hao, G.-Y ., Wang, S., Meng, L., Liu, D., Duchene, D. A., Wang, H., and Bhatt, S. Bayesagent: Bayesian agentic reasoning under uncertainty via verbal- ized probabilistic graphical modeling. InProceedings of the AAAI Con...

work page doi:10.1186/s12866-0 2022

[4] [4]

doi: 10.1109/TKDE.2016.2606

ISSN 1041-4347. doi: 10.1109/TKDE.2016.2606

work page doi:10.1109/tkde.2016.2606 2016

[5] [5]

URL https://doi.org/10.1109/TKDE.2 016.2606428. Wang, H. and Yeung, D.-Y . A survey on bayesian deep learning.ACM Comput. Surv., 53(5), September 2020. ISSN 0360-0300. doi: 10.1145/3409383. URL https: //doi.org/10.1145/3409383. Wang, H., Shi, X., and Yeung, D.-Y . Natural-parameter networks: A class of probabilistic neural networks. In Proceedings of the ...

work page doi:10.1109/tkde.2 2020

[6] [6]

Qwen3 Technical Report

doi: 10.1038/ismej.2015.235. URL https: //doi.org/10.1038/ismej.2015.235. Yang, A., Li, A., Yang, B., Zhang, B., Hui, B., Zheng, B., Yu, B., Gao, C., Huang, C., Lv, C., Zheng, C., Liu, D., Zhou, F., Huang, F., Hu, F., Ge, H., Wei, H., Lin, H., Tang, J., Yang, J., Tu, J., Zhang, J., Yang, J., Yang, J., Zhou, J., Zhou, J., Lin, J., Dang, K., Bao, K., Yang, ...

work page internal anchor Pith review Pith/arXiv arXiv doi:10.1038/ismej.2015.235 2015

[7] [7]

+” root is strictly positive while the “−

URL https://iclr.cc/virtual/2024 /poster/19071. Zheng, J., Sun, Q., Zhang, M., Liu, C., Su, Q., Zhang, L., Xu, Z., Lu, W., Ching, J., Tang, W., Cheung, C. P., Hamilton, A. L., O’Brien, A. L. W., Wei, S. C., Bernstein, C. N., Rubin, D. T., Chang, E. B., Morrison, M., Kamm, M. A., Chan, F. K. L., Zhang, J., and Ng, S. C. Non- invasive, microbiome-based diag...

work page doi:10.1038/s41591-024-03280-4 2024

[8] [8]

yes”/“no

as the backbone. For theIBD Diagnosistask, we utilizedQwen3-8B(Yang et al., 2025). •LoRA Configuration: – Molweni:We applied LoRA to projection layers ( q proj, k proj, v proj, o proj) with rank r= 8 , scaling factorα= 16, and dropout0.05. – IBD Diagnosis:We applied LoRA to the same projection layers but with rank r= 16 and α= 32 . Additionally, we set th...

work page arXiv 2025