acceptodds
Under review as a conference paper at ICLR 2027

Autonomous Discovery of Training Distribution for Controllable Molecular Generation

Abstract

Molecular generative models are typically pretrained on fixed drug-like datasets, implicitly constraining the structural support of the learned representation and limiting downstream adaptability. We investigate whether pretraining distribution support acts as a controllable lever on generative expressivity, and whether expanding this support through structurally distant molecules improves downstream diversity–activity trade-offs. We instantiate this hypothesis using a large language model (LLM) as an autonomous research agent and a lightweight autoregressive Transformer as the molecular generator. The agent iteratively proposes molecular pretraining mixtures, the generator is retrained under fixed model capacity and compute budget, and performance is evaluated on the task of spleen tyrosine kinase (SYK) inhibitor generation and optimization using a joint objective combining scaffold entropy and predicted pIC after fine-tuning on known SYK inhibitors. Across multiple iterations, the search identifies boundary configurations composed of molecules maximally distant from canonical drug-like datasets, yielding higher diversity–activity scores than conventional baselines. The best-performing configuration improves the joint diversity–activity objective by 1.6% relative to the baseline, driven primarily by increased scaffold entropy while maintaining competitive predicted affinity. These results suggest that training distribution discovery can serve as a data-centric mechanism for improving controllable molecular generation under constrained model capacity.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
generative models, chemistry, drug design, drug discovery, training distribution optimization, autonomous agents, data-centric AI

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_iBUHOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper asks whether pretraining-corpus composition is a lever on the diversity–activity trade-off of a molecular generator. An LLM agent (Qwen-3.5-27B, autoresearch-style loop plus a "Grill-me" clarification step) proposes mixtures over ChEMBL/ZINC/COCONUT/HMDB and sampling operators. A 2.8M-parameter SMILES GPT is pretrained on 500K molecules, fine-tuned on SYK inhibitors, and scored by S = scaffold entropy × mean predicted pIC50 of the top-50 generated molecules (QSAR). Over ~22 iterations the best mixture is "ZINC molecules farthest from ChEMBL". It is reported as +1.6% over a 500K-ChEMBL baseline, attributed to higher scaffold entropy, and said to rival architectural scaling.

Critical strengths
  1. Code and agent instructions are released, agent failure modes are disclosed (out-of-space proposals, wrong epoch selection), and the selected and baseline configurations are re-run over five seeds. Seeds 1–4 are independent of the selection run.
  2. The question is reasonable, and the observation that moderate ChEMBL/ZINC/natural-product mixes do not help is a useful negative data point. AI involvement is disclosed; my concerns are verifiability and scope, not authorship.
Critical weaknesses
  1. The headline is not robust and the conclusions exceed the evidence.
    • "+1.6%" is seed 0 vs seed 0 (9.093 vs 8.949), the run used for selection. Five-seed means are 9.035 vs 8.956 (+0.9%), and about 9.021 vs 8.957 (+0.7%) on seeds 1–4. Seed 0 is an outlier (entropy 0.970 vs 0.947–0.951; validity 0.506 vs 0.635–0.652).
    • Fig. 5 shows the reported scores are per-run peaks. The "BEST" run spikes at epoch 2 (≈9.09), but at the authors' own recommended stop (epoch 3) it is ≈8.88, below the ChEMBL baseline (≈8.95) and, by my reading, the other mixtures shown. Taking the maximum over epochs on the objective itself is optimistic and favours spiky curves. The epoch used in Tables 3–4 is not stated.
    • Under the authors' Bonferroni correction only the composite remains significant (p=0.0026; entropy p=0.048, pIC50 p=0.27, validity p=0.16). The text nevertheless asserts "diversity-driven" and quotes p<0.05 for entropy.
    • The QSAR model is both objective and evaluator. Validity drops (0.62 vs 0.67; 0.51 in the selected run). There is no uniqueness, novelty, SA, or distance-to-training-ligands analysis, and the N behind the top-50 is unspecified.
    • No scaling experiment supports "rival/more impactful than architectural scaling".
  2. The agent's contribution is untested, and figures/tables are internally inconsistent.
    • The search space is tiny (4 corpora, mixture weights, 7 operators). The best configuration (found ~iteration 6, Fig. 3) amounts to "take the molecules farthest from ChEMBL", a one-line human hypothesis.
    • There is no random/grid/BO baseline and no repeated agent runs or LLM variants. Roughly half of the ~22 trials sit at 8.97–9.03, close to seed noise (baseline SD ≈0.02). Picking the best of ~22 noisy trials invites winner's curse.
    • Fig. 2: n=23 vs 16 mixtures in Table 2. R²=0.86 is largely arithmetic (S=H×pIC50, with pIC50 only 9.29–9.62). Entropy and pIC50 trade off (best point 9.375 vs baseline 9.482). The SYK reference point lies on the fitted line and, if included in n=23, adds leverage.
    • Fig. 4: the farthest-ZINC subset is a compact island in UMAP, possibly homogeneous. Internal diversity and size/charge distributions are not reported, and under-adaptation to the target is not excluded.
    • Table 2 lists three different mixtures with identical pIC50/scaffold-count/validity (9.455/31/0.633), and two rows with the same label but different results (8.637 vs 9.021).
    • Text vs appendix: 9.025 vs 9.0354; entropy SD 0.001 vs 0.0095; top-50 vs top-1000; vocabulary 165 vs 512.
Questions for authors
  1. Which epoch is reported in Tables 2–4 for each arm and seed? Provide per-seed trajectories and a fixed-epoch comparison.

  2. Explain the identical Table 2 rows and the duplicate label, and release raw logs for all runs.

  3. Add random/grid/BO baselines with the same trial budget, and repeat the agent search ≥3 times or with other LLMs.

  4. Define scaffold entropy (which set, normalization) and N generated. Does the QSAR training set overlap the SYK fine-tuning set? What fraction of the top-50 are near-duplicates of SYK ligands?

  5. (suggestion) Add a tuned ChEMBL-only baseline (epochs, temperature), more targets, and a real model-size scaling comparison if parity is claimed.

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

This manuscript investigates whether optimizing the composition of pretraining data can improve downstream molecular generation under a fixed model architecture and computational budget. The authors propose an LLM-driven framework that iteratively selects molecular datasets and sampling strategies, pretrains a lightweight Transformer, fine-tunes it on known SYK inhibitors, and evaluates the resulting molecules using a composite objective combining scaffold entropy and predicted pIC50. The experimental results explore different mixtures of ChEMBL, ZINC, COCONUT, and HMDB. The best configuration uses ZINC molecules selected to be structurally distant from ChEMBL, achieving a modest improvement over the ChEMBL baseline. The authors find that this improvement is primarily driven by increased scaffold diversity rather than improved predicted activity.

Critical strengths
  • I like the idea of treating the pretraining distribution as an optimization variable rather than always focusing on model architecture or training objectives. The experiments are reasonably controlled, with a fixed generator and training budget across different dataset configurations. This makes the effect of pretraining data composition easier to interpret.

  • Rather than claiming that the proposed approach improves every aspect of molecular generation, the paper shows that the performance gain mainly comes from scaffold diversity. I appreciate this distinction, especially since the predicted activity does not significantly improve. The additional multi-seed comparison also provides some evidence that the diversity improvement is reproducible.

Critical weaknesses
  • The best-performing configuration is simply a ZINC subset selected to be maximally distant from ChEMBL. This is a reasonable data selection strategy, but it does not require an LLM to identify or implement. The paper does not compare the autonomous search with random search, a predefined distance-based selection strategy, or other conventional optimization approaches under the same experimental budget. As a result, it is difficult to determine whether the main contribution is the autonomous discovery framework or the observation that structurally diverse pretraining data can improve scaffold diversity.

  • The evaluation focuses entirely on SYK inhibitor generation, and the reported improvement in the composite objective is relatively small. More importantly, the gain comes primarily from scaffold entropy, while predicted pIC50 does not significantly improve. The best configuration also has lower molecular validity than the ChEMBL baseline. I think these results support a narrower conclusion about the diversity–validity trade-off, rather than demonstrating a general improvement in molecular generation or target-specific adaptability.

Questions for authors
  • Could the authors compare the LLM-driven search with a simple random search or a predefined distance-based data selection strategy using the same number of evaluations? Since the best configuration corresponds to selecting ZINC molecules farthest from ChEMBL, I would like to understand what additional benefit the autonomous agent provides.

  • How sensitive are the conclusions to the choice of the composite objective and biological target? In particular, would the selected pretraining distribution remain preferable if molecular validity were included in the objective, or if the evaluation were repeated on another protein target? This seems important because the current improvement is driven mainly by scaffold entropy rather than predicted activity.

Reviewer_LHQPOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper searches for the pretraining mixture of a molecular generator. The mixture is not fixed in advance. An LLM agent proposes mixtures from four molecular databases. A SMILES transformer with 2.8M parameters is retrained for each mixture. Each run is scored on SYK inhibitor generation. The score is scaffold entropy times predicted pIC50. The winning mixture is 100% ZINC, and these molecules are selected to be far from ChEMBL. The winner is 1.6% better than a ChEMBL baseline on the combined score. The improvement comes from diversity. Predicted potency does not improve.

Critical strengths
  • The experimental control is clean. Only one variable moves. Architecture, tokenizer, optimizer, schedules, seed and dataset size are fixed. .
  • The authors use 5 seeds and a Welch t-test. They also apply a Bonferroni correction.
  • The artifact works. I downloaded it. The agent prompt, the mixture code and the full search log are there.
  • The limitations section is open about context window limits and about agent proposals that went out of scope.
Critical weaknesses
  • §2 says that optimization of pretraining distributions is "largely underexplored". This is not correct. Data mixture optimization is an active topic. DoReMi (Xie et al., NeurIPS 2023) optimizes mixture weights for language model pretraining, and there is a large data selection literature next to it. The paper cites none of it. The new part is smaller: mixture search for a SMILES generator. The paper should say this.
  • The objective does not measure a bottleneck in drug discovery. It is scaffold entropy times predicted pIC50. It says nothing about synthesizability. It says nothing about the reliability of the scoring function. It says nothing about transfer to new chemotypes. A chemist cannot use a 0.01 change in normalized scaffold entropy. §1 speaks about drug discovery, but the measured quantity does not support any decision there.
  • The winning configuration selects molecules that are far from ChEMBL in fingerprint space. A QSAR model then scores these molecules. That model was trained on drug-like actives, so the winning molecules are outside its applicability domain. QSAR predictions are weakest there. The paper reports "competitive predicted affinity" (9.515 to 9.478, p = 0.29). It is possible that the predictor cannot separate good molecules from bad molecules out of domain.
  • Each reported score is the maximum over 10 fine-tuning epochs. The best epoch is not the same for each run. The log gives "best epoch 3", "best epoch 2", "best epoch 1", and in one run "best epoch 9". So each run had 10 chances, and the number was read at different points. Please report a fixed epoch.
  • The search space is small: a few databases, some proportions and seven fixed strategies. There is no random search baseline and no grid baseline. §5.2 says the search is "heuristic and LLM-driven rather than formally optimized". The title and abstract say "autonomous research agent".
  • The released program.md gives the hypothesis to the agent: "structural diversity increases the model's adaptability". It also tells the agent to "try more radical mixtures — 100% non-druglike, extreme strategies". The paper then says the search identifies maximal structural distance, and that the agent finds this direction from scalar reward alone. To define the search space in the prompt is normal. But to write the answer in the prompt is improper, and then the discovery claim is not valid. Please show the prompt in the paper and make the claim weaker. Or run the search again without this direction.
Questions for authors

See weaknesses.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.