acceptodds
Withdrawn from ICLR 2027

Active Molecular Query Synthesis by Influence-guided Latent Search

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Molecular screening relies on accurate property predictions, but obtaining the data to train these predictors can require costly assays or computations. Active learning reduces this cost by selecting molecules for evaluation, but its choices are often restricted to an existing library. Molecular generation allows the learner to construct additional queries and choose them for their estimated effect on prediction. We propose Influence-guided Molecular Query Synthesis (iMQS) to computationally generate molecular structures for their value as training data under a limited evaluation budget. Its acquisition objective is the estimated reduction in prediction loss on a validation set. An influence function approximates this improvement without retraining for every candidate and guides search in a continuous molecular representation. Our analysis bounds errors in comparing candidate improvements and the validation improvement lost through query selection when the predictor's final layer is refitted. Across nine molecular endpoints, iMQS improves predictive learning over several pool and generation baselines. In two AutoDock Vina applications, mean recovery of measured screening actives increases over the acquisition and refitting procedure. These results support molecular query generation as a way to acquire useful training data under limited evaluation budgets.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Active learning, molecular generation, influence function

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_1eevOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes Influence-guided Molecular Query Synthesis (iMQS), which generates molecules for their expected value as training data under a limited labeling budget. Experiments cover 9 endpoints and 2 screening tasks. iMQS generally outperforms generation baselines, but its advantages over pool-based methods are less consistent.

Critical strengths

The method turns influence estimates into a tractable generation objective. Each acquisition cycle requires one linear solve involving the prediction-head Hessian. So candidate representations can be optimized without repeatedly retraining the predictor.

Critical weaknesses
  1. The main benchmarks in table 2 measure teacher agreement. Both training and test labels are from TxGemma. However, it cannot show that being closer to the teacher's answer leads to more accurate predictions of molecular properties. This limits the significance of the main results.
  2. The improvements over strong simple baselines, such as random-pool, are often small or statistically inconclusive.
Questions for authors
  1. Can the authors evaluate the models against the experimental labels of the test molecules to determine whether the improvements extend beyond agreement with TxGemma?
  2. The theoretical guarantee concerns head-only refitting with the predictor body fixed, whereas the implemented algorithm retrains the full MLP after each cycle. How would performance change if the implementation followed this head-only refitting procedure?
Reviewer_wjWvOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Influence-guided Molecular Query Synthesis (iMQS), an active learning method that generates new molecular queries to improve property prediction under a limited oracle budget. The method uses an influence function to estimate the reference loss reduction from adding a new training sample, then optimizes this objective in a pretrained VAE's latent space. The generated molecule is labeled by an oracle, and then used to retrain the property predictor. The authors evaluate iMQS on nine molecular property prediction endpoints using TxGemma as the oracle, and on two LIT-PCBA virtual screening tasks using AutoDock Vina scores. The paper also provides a theoretical analysis of the influence approximation.

Critical strengths
  1. The idea of synthesizing molecular queries in the VAE latent space based on their estimated contribution to predictive learning is interesting. iMQS uses influence-guided latent search to construct potentially useful training samples beyond the fixed pool. The theoretical analysis also provides some justification for the proposed acquisition score under final-layer refitting.

  2. The experimental evaluation covers nine molecular property prediction endpoints, with comparisons against multiple acquisition methods across ten random splits. The additional docking-based virtual screening experiments provide a useful attempt to evaluate whether the acquired training data can improve downstream screening performance.

Critical weaknesses
  1. The theoretical analysis considers refitting only the final prediction head with the earlier layers fixed, while the actual acquisition workflow retrains the entire MLP. In addition, the generated molecule (m∗=Dec(z^)m^* = \mathrm{Dec}(\hat{z})) is added to the training set using the optimized latent vector (z^\hat{z}), rather than its re-encoded representation (Enc(m∗)\mathrm{Enc}(m^*)). While the authors acknowledge these differences, their empirical effects are not sufficiently investigated.

  2. The practical advantages of the proposed workflow are not fully established. The molecular property prediction experiments mainly compare acquisition strategies using a fixed VAE representation and a shallow MLP, but the authors did not compare the final predictive performance against strong molecular property prediction baselines under the same labeling budget. The virtual screening evaluation is limited to two targets, where the improvements are relatively modest and iMQS does not outperform all acquisition baselines. Comparisons with existing active-learning virtual screening approaches would help clarify its practical benefits.

Questions for authors
  1. According to the weakness 1, have the authors evaluated how well the influence scores correlate with the actual validation improvements after full MLP retraining? Such an analysis would help assess whether the approximation provides a reliable acquisition objective.

  2. Section 3.3 mentions that decoding and re-encoding an optimized latent vector may not recover the same representation. Have the authors considered using the re-encoded molecular representation instead of the optimized latent vector when updating the training set? How sensitive is the method to this difference?

  3. For molecular property prediction, how would the final predictor perform against established baselines, such as Chemprop or pretrained molecular models like MoLFormer and CheMeleon, under the same labeling budget?

  4. In the LIT-PCBA experiments, the predictor is trained on docking scores but evaluated using experimentally measured active/inactive labels. Could the authors further discuss how this difference between training and evaluation objectives affects the results? Besides, would it be possible to compare with established docking-based active-learning virtual screening approaches, such as MolPAL and Cao et al. (2024). This would help assess the method's value for practical virtual screening.

  5. Algorithm 1 and the method description "score a proposed training example" appear to acquire one molecule per cycle and then retrain the predictor. Could the authors clarify whether the queries are acquired one by one or in batches? Could the authors report the running time of the proposed method?

Reviewer_6TAtOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The author propose Influence-guided Molecular Query Synthesis (iMQS) to computationally generate molecular structures for their value as training data under a limited evaluation budget, which is different from the previous work to select candidates from a fixed library. The acquisition objective is the estimated reduction in prediction loss on a validation set. The experiments show that across nine molecular endpoints, iMQS improves predictive learning over several pool and generation baseline.

Critical strengths

The paper makes a useful distinction between molecules with desirable properties and molecules that would actually help train a better model. The proposed score is efficient to optimize, without repeatedly refitting the model during search. The experiments use well-matched controls, and I appreciate that the authors report cases where the method only matches or falls behind strong baselines. The detailed run records also makes the results easier to check.

Critical weaknesses

My main concern is that the benchmark uses TxGemma labels for both training and testing, so it mainly measures agreement with the teacher rather than prediction of measured properties. This is especially concerning for generated molecules, where the teacher’s reliability is unknown. The gains are also small and variable, with little clear improvement over the strongest pool methods and weak docking results. Another issue is that generated molecules are represented by the optimized latent vector rather than the re-encoded decoded molecule, potentially attaching a label to a mismatched feature vector. An ablation would help establish whether this matters. The theory also assumes head-only refits, while the implementation retrains the full MLP, making its relevance to the actual method unclear. The small-scale experiments and lack of results on synthesizability, diversity, and computational cost further limit what we can conclude about practical usefulness.

Questions for authors
  1. How far are the generated molecules from TxGemma's training distribution (and from the source datasets)? Can you validate a sample of teacher labels on generated molecules with a higher-fidelity oracle (docking, physics-based, or experimental)?
  2. How often does the oracle label differ from the requested label y*? How does iMQS compare with an expected-influence acquisition over the predictive distribution?
  3. Can you report validity, uniqueness, novelty, diversity, synthesizability (e.g., SA score), and wall-clock cost for each method?

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

Related papers

Discussion (0)

Sign in to comment.