acceptodds
Withdrawn from ICLR 2027

Multi-task Prediction of Solid-phase Extraction Protocol Components from Molecular Descriptors

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Solid-phase extraction (SPE) is widely used to enrich trace pollutants before instrumental analysis, yet selecting cartridge chemistries, solvents and mixture ratios still relies on literature search and expert adaptation. Learning these choices from published methods is difficult because a pollutant can have several reported alternatives, solvent labels are highly imbalanced, and ratio annotations are incomplete. We propose SPEGen (Solid-Phase Extraction Protocol Generator), a multi-task framework that predicts SPE components from molecular descriptors. We aggregate published methods by pollutant to construct set-valued supervision that retains alternative reported choices. A shared encoder with task-specific heads predicts cartridge sets, workflow-step indicators, step-specific solvent sets and their ratios, with ratio training restricted to observed annotations. Rank-Gauss normalisation handles heavy-tailed descriptors, while a validation-tuned decoder selects label sets and suppresses solvents for absent steps. Bootstrap aggregation combines predictions from resampled training data. Experiments against six tabular baselines on a held-out pollutant split show that SPEGen-Bag achieves the highest cartridge and solvent F1 and the lowest ratio mean absolute error. Cumulative ablations show gains from normalisation and bagging, while decoding analyses reveal how predicted set size affects the trade-off between recovering reported alternatives and exact-match accuracy.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Solid-phase extraction, Structured prediction

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_Aiu4OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies the prediction of solid-phase extraction (SPE) cartridges, workflow steps, step-specific solvents, and associated ratios from molecular descriptors. The authors aggregate literature methods by pollutant, using set-valued labels to retain alternative reported components and masking missing ratio annotations. The resulting dataset contains 11,441 pollutants, of which 10,335 have protocol annotations.

The proposed framework, SPEGen, combines a shared encoder and task-specific heads with Rank-Gauss preprocessing, imbalance-aware losses, step-guided solvent prediction, and validation-calibrated decoding. Its bagged variant averages predictions from eight bootstrap-trained models.

On a fixed random pollutant split, the bagged variant achieves the highest cartridge and solvent micro-F1 and the lowest ratio mean absolute error among the reported methods. However, random forest and a standard MLP perform best on component-triplet accuracy and strict ratio-verified Hit@1, respectively. Cumulative ablations and cardinality diagnostics examine the effects of normalization, ensembling, and predicted set size. The outputs should be interpreted as candidate components rather than complete executable protocols with verified joint compatibility.

Critical strengths
  1. A meaningful task formulation with supervision tailored to the data.

SPE component selection is a practically relevant analytical chemistry problem. Rather than assigning a single answer to each pollutant, the authors retain multiple reported components and mask missing ratio annotations. This accommodates the multiple alternatives and partial annotations present in literature-derived data and constitutes the paper’s most substantive contribution.

  1. Informative and transparent analysis of decoding behavior and limitations.

The paper examines the trade-offs between predicted set size, precision, recall, and exact match. The fixed-ranking cardinality diagnostic identifies output size as an important bottleneck. The authors also acknowledge the loss of joint assignments, the effect of missing ratios on strict evaluation, and the stronger baseline performance on some metrics. These analyses help establish an appropriately bounded interpretation of the method.

Critical weaknesses
  1. The current comparisons do not adequately establish the advantage of the proposed method itself.

Baselines use a fixed threshold of 0.5, while SPEGen receives per-label threshold calibration and architecture search; the best-performing variant additionally averages eight models. The paper does not provide comparisons with matched calibration, search budgets, or outer-loop ensemble sizes. It also lacks a simple multi-task control matched in preprocessing, capacity, and decoding.

This directly affects the main conclusion: the observed improvements may partly reflect more extensive optimization and ensembling rather than distinctive framework components. The solvent F1 and ratio MAE gains are relatively small, and baselines remain strongest on the two stricter component metrics. 2. The task is meaningful, but methodological novelty and evidence of practical recommendation validity remain limited.

The framework primarily combines established multi-task networks, imbalance-aware losses, rank-based transformations, threshold calibration, and bagging. Its value therefore depends substantially on whether the task and data contributions support reliable and generalizable recommendations.

Pollutant-level aggregation discards the original joint relationships between cartridges, solvents, and ratios. Triplet only checks separate membership in aggregated label sets, rather than whether the components occur together in a source method or are compatible. The random split does not rule out scaffold or publication overlap, and no laboratory extraction validation is reported. Although these limitations are acknowledged, the evidence does not yet establish the system’s practical value beyond retrospective recovery of literature labels.

Questions for authors
  1. Could the strongest baselines be evaluated with the same validation-based per-label threshold calibration and step gating? Do the main gains persist when search budgets and outer-loop ensemble sizes are reasonably matched?

  2. Is there a simple multi-task MLP control matched in Rank-Gauss preprocessing, model capacity, and decoding, but without step guidance and the auxiliary training terms? This would help determine whether the task-specific components provide independent benefits.

  3. Using the original method records, what fraction of predicted cartridge–step–solvent triples actually co-occur within a single source method? Would this method-level evaluation change the model ranking relative to the current marginal-membership evaluation?

Reviewer_7gPzOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes SPEGen, a model that predicts parts of a solid-phase extraction (SPE) protocol (cartridges, workflow steps, solvents per step and their ratios) from 21 PubChem descriptors of a pollutant. Published methods (36,776 records) are aggregated per pollutant (11,441 pollutants), so the targets are unions of all reported alternatives, and missing ratios are masked. The model is a shared MLP with task towers, rank-Gauss inputs and per-label thresholds tuned on validation. The main variant (Bag) averages 8 bootstrap members. On one random split Bag gets the best cartridge F1 (0.491 vs 0.418), solvent F1 (0.475 vs 0.457) and ratio MAE (0.146 vs 0.152) compared with 6 tabular baselines. Random forest is still better on triplet accuracy, and the MLP baseline on strict Hit@1. The oracle-size experiment shows that the number of labels returned has a large effect on set-recovery scores.

Critical strengths
  1. The task is new and practically motivated, and the set-valued, partially observed targets are handled carefully (unions over methods, masked ratios, availability masks). The limits of aggregation are discussed openly in App. F.
  2. Transparent reporting of metric definitions and limitations. The paper shows where SPEGen loses (Triplet, strict Hit@1), gives exact counts for the triplet decomposition and points out weak spots of its own metrics. The code matches the described model, losses and split.
Critical weaknesses
  1. The advantage over comparably tuned baselines is not established. SPEGen uses per-label thresholds tuned on validation, while the baselines use a fixed threshold of 0.5, so the current comparison does not isolate the contribution of the learned model from that of the decoder. Since this is the main claim of the paper, the baselines should be run through the same decoder and compared with simple frequency-prior and kNN baselines.
  2. The data are not documented, and the results cannot be independently reproduced from the provided package. The model is built from standard parts, so the dataset is the main contribution, but the paper does not say where the 36,776 records come from or how they were extracted. The provided supplementary package is insufficient to independently reproduce the reported experimental results: it contains code only, without data, configs or the code for the Bag ensemble.
Questions for authors
  1. Where do the 36,776 records come from, how were they extracted, and will the processed data be released?
  2. How do the baselines perform with the same validation-tuned decoder, and how do frequency-prior and kNN baselines perform?
  3. Was the test split used when choosing configurations, or only the validation split?
  4. Can you provide the missing materials (data, configs, Bag aggregation code, environment) and the script that produced the baseline rows of Table 2? The supplied compute_baseline_overall.py does not compute C-top1 or Triplet and averages SSS/HC over all test rows, and the README command for Chain differs from Table A2.
  5. Are “Unknown”/”Null cartridge” and the UNK solvent among the targets, and how many ratio values were clipped to 1.0?
Reviewer_v1NXOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes SPEGen, a model to predict solid-phase extraction (SPE) components from 21 molecular descriptors. For each pollutant, the authors merge all published methods into one set of labels, including cartridges, workflow steps, step-wise solvents and solvent ratios. The model uses a shared MLP encoder with separate heads for each task. It also uses rank-Gauss normalisation, validation-tuned thresholds and an 8-member bagging ensemble. On a random held-out pollutant split, SPEGen-Bag gets the best cartridge F1, solvent F1 and ratio MAE compared with six tabular baselines.

Critical strengths
  1. The task is new and has practical value. Now, choosing SPE conditions still depends on manual literature search. The curated dataset (11,441 pollutants and 36,776 method records) is useful for this area.
  2. The paper discusses its limitations honestly, such as the effect of aggregation, the missing ratios and the low Hit@1. It also reports the metrics where SPEGen does not win.
Critical weaknesses
  1. The baselines are too old and too weak. The paper only compares with classic methods, such as decision tree, random forest and XGBoost. Recent tabular models the authors cited (e.g., TabPFN, TabR) and molecular models (e.g., Chemprop) are missing. Besides, the baselines use a fixed threshold of 0.5, while SPEGen tunes thresholds on the validation set and uses an 8-model ensemble. So the comparison is not fair.
  2. The data split may overestimate the generalisation. The split is random by pollutant, and the authors admit that scaffolds and source papers may overlap between training and test sets. A scaffold split or source-disjoint split is needed.
  3. The output is not a usable protocol, and the novelty in the machine learning aspect is very limited. After aggregation, the model only predicts a list of candidate components, not a method that can be used in the lab. Strict Hit@1 is lower than 0.11. Most parts of the method are standard techniques.
Questions for authors
  1. Can you compare with recent baselines, such as TabPFN, FT-Transformer, TabR, classifier chains or Chemprop? Can you also tune the thresholds of the baselines on the validation set in the same way as SPEGen?
  2. What are the results on a scaffold split or a source-paper-disjoint split?
  3. Can structural features, such as Morgan fingerprints or a GNN encoder, give better results than the 21 descriptors?
  4. After bagging, Ratio≤0.10 drops from 0.6395 (Chain) to 0.5916, and Hit@1 drops from 0.0952 to 0.0872. Is this because the bucket centres of different members are averaged?

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

Related papers

Discussion (0)

Sign in to comment.