pith. sign in

arxiv: 2605.19677 · v1 · pith:JUGA3ZNTnew · submitted 2026-05-19 · 💻 cs.LG · q-bio.QM

Agentic Discovery of Cryomicroneedle Formulations

Hao Li , Lifu Du , Nurul Hameed , Shemonti Saha Authai , Zlata Stefanovic , Chenjie Xu This is my paper

Pith reviewed 2026-05-20 07:48 UTC · model grok-4.3

classification 💻 cs.LG q-bio.QM
keywords cryomicroneedlescryopreservationBayesian optimizationGaussian processformulation discoverymesenchymal stem cellsclosed-loop optimizationcell viability
0
0 comments X

The pith

An iterative AI workflow using literature data and wet-lab feedback discovers effective cryomicroneedle formulations with high cell viability.

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

The paper sets out to show that a closed-loop system combining a Gaussian process model trained on existing cryopreservation studies with Bayesian optimization and repeated laboratory tests can identify cryoprotectant mixtures suitable for cryomicroneedles. A reader would care because traditional formulation discovery requires extensive trial and error, and this method claims to reduce that burden by adapting predictions based on new measurements. It starts with 198 formulations from literature turned into 21 features to build an initial model, then updates it over ten rounds with 106 new observations. The process led to better predictions and a top formulation with 95.15 percent post-thaw viability using low amounts of certain protectants. This suggests computational tools can help labs find viable options more efficiently even when starting with limited specific data.

Core claim

The authors report that an uncertainty-aware model built from literature data on mesenchymal stem-cell cryopreservation initially performed poorly on cryomicroneedle tasks but improved through sequential wet-lab validation. Batch RMSE fell from 41.21 to 6.86 percentage points across iterations, rank correlations turned positive, and the overall predicted-versus-measured fit reached an R squared of 0.942. The best validated mixture delivered 95.15 percent viability after thawing while using reduced levels of DMSO, ectoin, ethylene glycol, and fetal bovine serum, though viability alone did not guarantee proper needle formation.

What carries the argument

The uncertainty-aware literature prior trained on 21 ingredient features from 198 formulations, updated iteratively via Gaussian-process surrogate modelling and Bayesian optimization with wet-lab observations.

If this is right

  • The model becomes more accurate for cryomicroneedle outcomes as more lab data is incorporated.
  • Formulation discovery can proceed with fewer initial experiments when leveraging literature priors.
  • High cell viability must be paired with physical integrity checks for successful cryomicroneedle devices.
  • Laboratories without deep data analysis expertise can access effective discovery tools through this infrastructure.

Where Pith is reading between the lines

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

  • Similar closed-loop methods could accelerate formulation work in other areas like drug delivery or tissue engineering where physical and biological constraints interact.
  • Starting with multi-objective optimization that includes needle formation metrics from the first iteration might reduce the need for later adjustments.
  • Scaling the approach to larger ingredient libraries or different cell types would test how general the adaptation process is.

Load-bearing premise

The chosen 21 features from literature data plus the iterative wet-lab corrections together account for the main influences on both cell survival and the physical properties needed for cryomicroneedle creation.

What would settle it

Measuring the viability of a formulation that the final model predicts will perform well but finding the actual result far below the predicted value would challenge the claim that the workflow has adapted successfully.

read the original abstract

Cryomicroneedles offer a route to minimally invasive intradermal delivery of living cells, but their cryogenic formulations must reconcile cell protection with constraints on toxicity and device fabrication. Here we report an AI-assisted, closed-loop workflow for cryomicroneedle cryoprotectant discovery that combines literature curation, Gaussian-process surrogate modelling, Bayesian optimization, and sequential wet-lab validation. A curated dataset of 198 mesenchymal stem-cell cryopreservation formulations from 42 studies was converted into 21 ingredient features and used to train an uncertainty-aware literature prior. This model captured moderate structure in the literature data but failed prospectively, motivating iterative wet-lab correction. Across ten validation iterations and 106 wet-lab observations, the model progressively adapted to cryomicroneedle-specific outcomes: batch RMSE decreased from 41.21 to 6.86 percentage points, later-stage rank correlations became consistently positive, and the cumulative wet-lab predicted-versus-measured summary reached $R^2 = 0.942$. The best validated formulation achieved 95.15\% post-thaw viability with low DMSO, ectoin, ethylene glycol, and fetal bovine serum. However, high viability alone did not ensure intact cryomicroneedle formation, highlighting the need for future multi-objective optimization. These results demonstrate that agent-assisted computational infrastructure can make data-efficient formulation discovery more accessible to labs with minimal data expertise in-house. Project code is available at https://github.com/baitmeister/ML-for-CryoMN.

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.

Referee Report

2 major / 1 minor

Summary. The manuscript presents an AI-assisted closed-loop workflow for cryomicroneedle formulation discovery. It curates 198 literature formulations into 21 ingredient features to train a Gaussian-process literature prior, then applies Bayesian optimization with sequential wet-lab validation across ten iterations and 106 observations. Reported outcomes include batch RMSE reduction from 41.21 to 6.86, later-stage positive rank correlations, a cumulative wet-lab predicted-versus-measured R² of 0.942, and identification of a formulation achieving 95.15% post-thaw viability with low DMSO, ectoin, ethylene glycol, and fetal bovine serum. The work notes that viability alone does not guarantee intact needle formation and releases project code.

Significance. If the adaptation metrics reflect genuine prospective accuracy on unseen formulations, the work demonstrates a practical route to data-efficient discovery in a complex, multi-constraint formulation space using modest experimental budgets. The combination of a literature-derived prior with iterative experimental feedback is a constructive approach, and the explicit release of code at https://github.com/baitmeister/ML-for-CryoMN supports reproducibility and extension by other labs. The practical outcome of a high-viability, low-DMSO formulation is relevant for reducing toxicity in intradermal cell delivery.

major comments (2)
  1. [Abstract and Results] Abstract and Results section on iterative validation: The cumulative R² = 0.942 and the batch RMSE drop from 41.21 to 6.86 are computed on the 106 observations that were acquired and incorporated into the Gaussian-process model during the ten Bayesian-optimization iterations. Because each new batch is chosen by the current surrogate and performance is summarized on the growing training set, these statistics largely reflect in-sample interpolation after data incorporation rather than prospective accuracy on formulations never seen by the adapted model. A fixed hold-out set, temporal split, or cross-validation protocol for the post-adaptation regime is required to support claims of progressive model improvement.
  2. [Methods and Results] Methods and Results on feature construction: The 21 ingredient features extracted from the 198 literature formulations form the basis of the uncertainty-aware prior, yet the manuscript provides limited justification that these features adequately encode the cryogenic temperature profiles, device fabrication constraints, and physical needle-formation mechanics specific to cryomicroneedles. The reported initial prospective failure of the literature prior is consistent with this possible mismatch, and additional multi-objective terms (viability plus mechanical integrity) should have been included from the first iteration rather than noted only after the fact.
minor comments (1)
  1. [Abstract] The abstract and main text would benefit from clearer separation between literature-based predictions and wet-lab measured outcomes when reporting rank correlations and RMSE values.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their detailed and constructive review. We address each major comment below, providing clarifications where the manuscript can be strengthened and indicating revisions accordingly.

read point-by-point responses

  1. Referee: [Abstract and Results] Abstract and Results section on iterative validation: The cumulative R² = 0.942 and the batch RMSE drop from 41.21 to 6.86 are computed on the 106 observations that were acquired and incorporated into the Gaussian-process model during the ten Bayesian-optimization iterations. Because each new batch is chosen by the current surrogate and performance is summarized on the growing training set, these statistics largely reflect in-sample interpolation after data incorporation rather than prospective accuracy on formulations never seen by the adapted model. A fixed hold-out set, temporal split, or cross-validation protocol for the post-adaptation regime is required to support claims of progressive model improvement.

    Authors: We appreciate the referee highlighting the distinction between in-sample and prospective performance. The reported batch RMSE values reflect the model's predictions on each new batch of formulations prior to their incorporation into the updated Gaussian process (i.e., using the surrogate from the previous iteration), thereby providing a measure of accuracy on data unseen by the current model at the time of selection. The cumulative R², by contrast, is computed on the full set of 106 observations after all updates. We agree that the manuscript should more explicitly distinguish these computations to support claims of progressive improvement. In the revised manuscript, we will clarify the batch-wise evaluation protocol in the Results section and add a limitations discussion addressing the practical constraints of a fixed hold-out set within a sequential, budget-limited experimental design. The observed trends in batch RMSE reduction and later-stage rank correlations nonetheless provide evidence of domain adaptation. revision: partial

  2. Referee: [Methods and Results] Methods and Results on feature construction: The 21 ingredient features extracted from the 198 literature formulations form the basis of the uncertainty-aware prior, yet the manuscript provides limited justification that these features adequately encode the cryogenic temperature profiles, device fabrication constraints, and physical needle-formation mechanics specific to cryomicroneedles. The reported initial prospective failure of the literature prior is consistent with this possible mismatch, and additional multi-objective terms (viability plus mechanical integrity) should have been included from the first iteration rather than noted only after the fact.

    Authors: The 21 features were derived directly from the compositional variables (concentrations of cryoprotectants and additives) appearing across the 198 literature formulations to enable the Gaussian process to model general viability trends. We acknowledge that these features do not explicitly capture cryomicroneedle-specific factors such as temperature profiles during freezing or mechanical integrity of the needle structure, which is consistent with the observed failure of the initial literature prior on prospective tests. This domain mismatch motivated the iterative wet-lab adaptation. Regarding multi-objective optimization, the study focused on post-thaw viability to first establish the closed-loop workflow; the manuscript already states that viability alone does not guarantee intact needle formation. We will revise the Methods section to provide further justification for the chosen features and expand the Discussion to explain the initial single-objective focus while outlining extensions to multi-objective optimization that incorporate mechanical integrity metrics from the outset in future work. revision: yes

Circularity Check

1 steps flagged

Adaptation metrics computed on sequentially incorporated wet-lab data without hold-out

specific steps
  1. fitted input called prediction [Abstract (and Results section describing cumulative wet-lab summary)]
    "Across ten validation iterations and 106 wet-lab observations, the model progressively adapted to cryomicroneedle-specific outcomes: batch RMSE decreased from 41.21 to 6.86 percentage points, later-stage rank correlations became consistently positive, and the cumulative wet-lab predicted-versus-measured summary reached R² = 0.942."

    The RMSE and R² summarize performance on the exact 106 observations acquired during the Bayesian-optimization loop and fed back into the Gaussian-process surrogate for iterative updates. Because each new batch is selected by the current model and then becomes part of the training data, the metrics largely reflect how well the updated model fits the data it has already seen rather than independent prediction on unseen formulations.

full rationale

The paper trains a Gaussian-process surrogate on literature data, then uses Bayesian optimization to select batches for wet-lab testing and incorporates those 106 observations to update the model across 10 iterations. The reported batch RMSE drop (41.21 to 6.86) and cumulative R² = 0.942 are evaluated on the growing set of observations that were chosen by the current surrogate and then added to it. This makes the performance numbers in-sample interpolation after data incorporation rather than prospective accuracy on formulations never seen by the final model. The manuscript explicitly notes the literature prior failed on first prospective tests but provides no fixed test set, temporal hold-out, or cross-validation for the adapted regime.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The central claim rests on the surrogate model's ability to capture structure in literature data and adapt via Bayesian updates to domain-specific wet-lab outcomes, using standard assumptions about response surface modeling without new physical postulates.

free parameters (1)
  • Gaussian process kernel hyperparameters
    Fitted during initial training on literature data and updated with each wet-lab batch to minimize prediction error.
axioms (1)
  • domain assumption The 21 ingredient features derived from literature studies are sufficient to represent the relevant chemical and biological factors for cryoprotection in cryomicroneedle devices.
    Invoked when converting the 198 formulations into model inputs for the uncertainty-aware prior.

pith-pipeline@v0.9.0 · 5817 in / 1795 out tokens · 73985 ms · 2026-05-20T07:48:59.607473+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

17 extracted references · 17 canonical work pages · 3 internal anchors

  1. [1]

    Hao Chang, Sharon W. T. Chew, Mengjia Zheng, Daniel Chin Shiuan Lio, Christian Wiraja, Yu Mei, Xiaoyu Ning, Mingyue Cui, Aung Than, Peng Shi, Dongan Wang, Kanyi Pu, Peng Chen, Haiyan Liu, and Chenjie Xu. Cryomicroneedles for transdermal cell deliv- ery.Nature Biomedical Engineering, 5(9):1008–1018, May 2021. doi: 10.1038/s41551-021- 00720-1

    work page doi:10.1038/s41551-021- 2021

  2. [2]

    In situ-formed cryomicroneedles for intradermal cell delivery.NPG Asia Materials, 16, February 2024

    Mengjia Zheng, Tianli Hu, Yating Yang, Xuan Qie, Huaxin Yang, Yuyue Zhang, Qizheng Zhang, Ken-Tye Yong, Wei Liu, and Chenjie Xu. In situ-formed cryomicroneedles for intradermal cell delivery.NPG Asia Materials, 16, February 2024. doi: 10.1038/s41427- 024-00531-1

    work page doi:10.1038/s41427- 2024

  3. [3]

    Mi- croneedles in the clinic.Journal of Controlled Release, 260:164–182, August 2017

    Shubhmita Bhatnagar, Kaushalkumar Dave, and Venkata Vamsi Krishna Venuganti. Mi- croneedles in the clinic.Journal of Controlled Release, 260:164–182, August 2017. doi: 10.1016/j.jconrel.2017.05.029

    work page doi:10.1016/j.jconrel.2017.05.029 2017

  4. [4]

    Witek, Alan Mendoza, Michael Alexander, and Jonathan R

    David Whaley, Kimia Damyar, Rafal P. Witek, Alan Mendoza, Michael Alexander, and Jonathan R. T. Lakey. Cryopreservation: An overview of principles and cell-specific con- siderations.Cell Transplantation, 30, 2021. doi: 10.1177/0963689721999617

    work page doi:10.1177/0963689721999617 2021

  5. [5]

    Murray and Matthew I

    Kathryn A. Murray and Matthew I. Gibson. Chemical approaches to cryopreservation. Nature Reviews Chemistry, 6(8):579–593, August 2022. doi: 10.1038/s41570-022-00407-4

    work page doi:10.1038/s41570-022-00407-4 2022

  6. [6]

    Taking the Human Out of the Loop: A Review of Bayesian Optimization.Proceedings of the IEEE

    Bobak Shahriari, Kevin Swersky, Ziyu Wang, Ryan P. Adams, and Nando de Freitas. Taking the human out of the loop: A review of bayesian optimization.Proceedings of the IEEE, 104(1):148–175, January 2016. doi: 10.1109/JPROC.2015.2494218

    work page doi:10.1109/jproc.2015.2494218 2016

  7. [7]

    Peter I. Frazier. A tutorial on bayesian optimization.CoRR, abs/1807.02811, 2018. doi: 10.48550/arXiv.1807.02811

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.1807.02811 2018

  8. [8]

    Self-Driving Laboratories for Chemistry and Materials Science.Chemical Reviews, 2024

    Gary Tom, Stefan P. Schmid, Sterling G. Baird, Yang Cao, Kourosh Darvish, Han Hao, Stanley Lo, Sergio Pablo-Garc´ ıa, Ella M. Rajaonson, Marta Skreta, Naruki Yoshikawa, Samantha Corapi, Gun Deniz Akkoc, Felix Strieth-Kalthoff, Martin Seifrid, and Al´ an Aspuru-Guzik. Self-driving laboratories for chemistry and materials science.Chemical Reviews, 124(16):9...

    work page doi:10.1021/acs.chemrev.4c00055 2024

  9. [9]

    Executable code actions elicit better llm agents,

    Xingyao Wang, Yangyi Chen, Lifan Yuan, Yizhe Zhang, Yunzhu Li, Hao Peng, and Heng Ji. Executable code actions elicit better llm agents.CoRR, abs/2402.01030, 2024. doi: 10.48550/arXiv.2402.01030

    work page doi:10.48550/arxiv.2402.01030 2024

  10. [10]

    SWE-agent: Agent-Computer Interfaces Enable Automated Software Engineering

    John Yang, Carlos E. Jimenez, Alexander Wettig, Kilian Lieret, Shunyu Yao, Karthik Narasimhan, and Ofir Press. SWE-agent: Agent-computer interfaces enable automated software engineering.Advances in Neural Information Processing Systems, 37, 2024. doi: 10.48550/arXiv.2405.15793

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2405.15793 2024

  11. [11]

    URL https: //doi.org/10.1038/s41586-026-10265-5

    Chris Lu, Cong Lu, Robert Tjarko Lange, Yutaro Yamada, Shengran Hu, Jakob Foer- ster, David Ha, and Jeff Clune. Towards end-to-end automation of AI research.Nature, 651(8107):914–919, March 2026. doi: 10.1038/s41586-026-10265-5. 17

    work page doi:10.1038/s41586-026-10265-5 2026

  12. [12]

    Semantic scholar.Journal of the Medical Library Association, 106(1):145– 147, 2018

    Suzanne Fricke. Semantic scholar.Journal of the Medical Library Association, 106(1):145– 147, 2018. doi: 10.5195/jmla.2018.280

    work page doi:10.5195/jmla.2018.280 2018

  13. [13]

    OpenAlex: A fully-open index of scholarly works, authors, venues, institutions, and concepts

    Jason Priem, Heather A. Piwowar, and Richard Orr. OpenAlex: A fully-open index of scholarly works, authors, venues, institutions, and concepts.CoRR, abs/2205.01833, 2022. doi: 10.48550/arXiv.2205.01833. 18 Supplementary Methods AI-assisted workflow construction The initial one-shot prompt and agent transcript were recorded inprompt.mdin the GitHub reposit...

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2205.01833 2022

  14. [14]

    Data Parsing✓200 unique formulations extracted

    work page

  15. [15]

    Model T raining✓GP model trained (CV R 2=0.24, Training R2=0.69)

    work page

  16. [16]

    Optimization✓20 general + 15 DMSO-free candidates generated

    work page

  17. [17]

    DMSO-free

    V alidation Loop✓Template created, awaiting wet lab results Key Results: •Best general candidate: 78.6%±21.1% viability at 0.5% DMSO •Best DMSO-free candidate: 77.9%±23.5% viability at 0% DMSO The pipeline is ready for use. Step 4 correctly indicates it’s waiting for wet lab validation results before it can update the model. Dataset parsing and feature en...

    work page