acceptodds
Under review as a conference paper at ICLR 2027

MolLingo: Molecule-Native Representations for LLM-Powered Scientific Agents

10 views

Abstract

Large language models (LLMs) have shown strong general reasoning capabilities, yet molecular design remains challenging because it requires precise structural reasoning, specialized computational tools, and iterative search over a large chemical space. We present MolLingo, a multi-agent framework that addresses these challenges through chemically informed representations, tool-grounded reasoning, and coordinated optimization. We introduce BRICS-based Fragment Enumeration (BFE), which combines chemically valid BRICS fragmentation with corpus-level fragment frequencies to represent molecules as recurring chemical blocks. Encoding these blocks with both SMILES and common chemical names exposes meaningful structural units while preserving molecular connectivity. Building on this molecular representation, MolLingo integrates a Chemist Agent, a Literature Agent, an Orchestrator, and shared memory, together with tools for ADMET prediction, molecular docking, and binding retrieval. For property optimization, across GPT-5.4, Claude-4.6-Sonnet, and Gemini-3-Pro, block-based input improves the joint objective of property improvement and structural similarity over raw SMILES under matched LLM backbones. For binding optimization, incorporating docking-derived structural context increases mean docking-score improvement from 2.4% with direct prompting to 10.4% with GPT-5.4. Iterative experiments further show that memory improves structural preservation and reduces redundant edits, while adaptive search allocation improves optimization as the candidate budget increases. Together, these results show that molecular representation, structural context, and agentic coordination provide complementary signals for LLM-guided molecular design. Code is available at https://anonymous.4open.science/status/MolLingo-7450.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Multi-Agent Systems, Molecular Design, Drug Discovery

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_7QPKOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2(+1)
Read review
Paper summary

MolLingo is a multi-agent framework for molecular design with three parts. BFE splits molecules into frequent BRICS blocks, each written as SMILES plus a common name. Chemist, Literature and Orchestrator agents share a memory and call docking and ADMET tools. A docking-guided hit-to-lead stage conditions fragment growth on per-atom hotspots. Experiments cover ADMET optimization on Withdrawn 2.0 (67 DILI and 33 hERG molecules), hit-to-lead on 30 targets, TOMG-Bench, and hit identification against DrugGEN. The paper claims that block input beats raw SMILES under the same LLM (GPT-5.4 Imprv×Sim goes from 266.37 to 476.93) and that docking context raises mean docking improvement from 2.4% to 10.4%.

Critical strengths
  1. The paper is upfront about what its evidence does not show. It concedes that the BFE vs. full-fragment gap is only 0.569 (§3.2), that no significance test was run (App. E), and that the InChI and IUPAC variants failed (App. M.2). This made the paper easier for me to assess.
  2. The ablations say something useful about agentic search on their own. Memory brings repeated edits from 16.7% to 0%, and five iterations trace a clear trade-off between property improvement and similarity (Tables 12, 15, 16). I found this more convincing than the representation claim.
Critical weaknesses
  1. The representation claim does not hold up at matched budget. Table 16 is the only matched comparison (16 oracle calls). There, block SMILES scores 1031.94 against 1003.9 for raw SMILES, about 2.8% apart, with a per-molecule std of 12 to 17 pp on n=67, and Graph-GA reaches 1044.1 under the same budget. Table 14 also shows that k=40, the setting used for the headline Table 1, is the worst of the four candidate counts (476.9, vs. 794.9 at k=20). The paper does not say why it was chosen.
  2. There are no statistical tests, and the samples are small. DILI has 67 molecules and hERG has 33, with no paired tests, CIs or repeated runs. Several gains sit inside one std (hERG, Gemini-3-Pro: −0.56 ± 1.40 to 0.68 ± 3.22). The Orchestrator gain in Table 13 also comes with a larger generation budget.
Questions for authors
  1. [Clarification] Is the block vs. raw gap in Table 16 significant under a paired test or bootstrap over the 67 molecules? The same question applies to Tables 1 and 8.
  2. [Clarification] Why was Table 1 run at k=40? Does the block vs. raw gap survive at k=20?
  3. [Clarification] How much do the evaluation molecules and their optimized analogs overlap with the TDC data used to train the ADMET oracles?
  4. [Clarification] Improvement is measured as relative change in the oracle logit (App. L.1). Do the rankings change under probability or a thresholded success rate?
  5. [Suggestion] Graph-GA is competitive at equal budget. What does the LLM pipeline add for property optimization? If the answer is little, the abstract should say less.
Reviewer_ftJnOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper presents MolLingo, a multi-agent framework (Orchestrator, Literature Agent, Chemist Agent + shared memory) for LLM-guided molecular design across early drug discovery. Its core technical contribution is BFE (BRICS-based Fragment Enumeration): a single-pass, corpus-frequency-based construction of a ~1.2M-block molecular vocabulary from 20M drug-like molecules, plus a hierarchical decomposition-selection rule (fewest blocks - frequency threshold f_min=20 - minimum std of block frequencies). Molecules are then fed to frozen frontier LLMs as ordered sequences of wildcard-anchored block SMILES paired with common chemical names. A second contribution is docking-guided fragment growing: per-atom growth hotspots (neighboring residues within 7.0 Å + available volume from a clearance grid) derived from AutoDock Vina poses condition the LLM's fragment proposals. Validation: (1) DILI/hERG ADMET optimization on Withdrawn 2.0 (67/33 oracle-TP molecules) where block-based input improves the joint improvement x similarity score over raw SMILES for all three frontier backbones (GPT-5.4, Claude-4.6-Sonnet, Gemini-3-Pro); (2) hit-to-lead on 30 targets where the docking-guided workflow improves mean docking-score improvement from 2.4% (direct prompting) to 10.4%; (3) TOMG-Bench, where MolLingo reports the best SR and SR x Sim on LogP / QED / MR; (4) hit identification vs DrugGEN on CDK2/AKT1. Ablations cover memory, adaptive orchestration, candidate/iteration budgets, BFE vs naive BRICS, block names, and matched-budget non-LLM optimizers (REINVENT, Graph-GA).

Critical strengths
  1. BFE is a novel, well-specified representation with consistent within-backbone gains. The single-pass enumeration of two-BRICS-bond cuts plus corpus-frequency-guided decomposition selection (fewest blocks - f_min threshold - min-std criterion) is, to my knowledge after checking FragGPT / Trio / GraphBPE / mCLM / FragBERTa, not previously published; the complexity argument (Lemma 1 + Proposition 1) is a real efficiency case vs iterative graph BPE. The representation result — block-based input improving the joint improvement x similarity score over raw SMILES for all three frontier backbones under a matched generation budget — is the cleanest and most reproducible-in-principle result in the paper, and it directly supports the central hypothesis that LLMs edit molecules more effectively at chemically meaningful granularity.

  2. Honest evaluation. The paper names its own confounds in-text (DrugCLIP fallback unevaluated on novel targets; no significance tests reported; orchestration comparison not compute-matched; attention probe explicitly non-causal), collects limitations in Appendix O, and reports non-superiority where it exists (Graph-GA matches the joint score; REINVENT/SELFIES improve property more at the cost of similarity). This honesty materially increases the credibility of the positive results.

  3. The paper references and discussed not only classical approaches, but also the most recent relevant works in the field which adds to credibility and paper positioning.

Critical weaknesses
  1. The quantitative evidence is not statistically established and is self-referential. All ADMET claims rest on the authors' own oracle models as both guide and metric, on small sets (67 DILI / 33 hERG molecules), with large overlapping standard deviations and no paired tests or seed reporting. The primary metric is a relative change in the oracle's raw logits. Without significance testing, an external validation (different predictor family or at least one experimentally validated candidate), the central claim — that the representation improves optimization — is plausible but unproven. In my view, this point is fixable.

  2. The headline hit-to-lead result is a confounded workflow comparison, and the code is inaccessible at review time. The increase from 2.4% to 10.4% in docking conflates docking pose, residue context, volume estimates, and attachment-site information with the representation; no ablation isolates these (the authors concede this). Separately, the code link did not work when I checked it, so BFE vocabulary construction, the full prompt set, and oracle training cannot be verified at all. Together these mean neither of the two headline numbers is currently independently checkable.

Questions for authors
  1. Can you report paired significance tests (or ≥3 seeds with per-seed results) for the Table 1 and Table 2 comparisons? Given the std magnitudes, which within-backbone differences survive a paired test on 67/33 molecules?

  2. How do you rule out that the DILI "improvement" reflects overfitting to your own oracle?

  3. Would you commit to evaluating the top generated candidates with an independently trained predictor (different architecture / data) or, ideally, a small experimental panel? In the hit-to-lead experiment, what is the contribution of each context input (docking pose vs residue list vs available volume vs attachment-atom type)? A factorial or leave-one-out ablation would let readers attribute the 8% gain (from 2.4% to 10.4%).

  4. Why is k=40 used in the main Table 1 when your own sweep (Table 14) shows it is the worst setting, and what explains the non-monotonic decline?

  5. The imatinib decomposition in Fig. 2c differs from the one used in Appendix F for the same molecule — which is correct, and what setting (f_min, vocabulary version) produced each? Also, Table 10 lists α-synuclein as Q5S007; the correct human SNCA ID is P37840 — how was Table 10 validated?

  6. Is that possible to evaluate the DrugCLIP fallback on genuinely novel / understudied targets against a standard virtual-screening baseline? This is the setting where the system would be most valuable.

  7. Is that possible to add a matched-compute control for adaptive orchestration (same total oracle budget, fixed vs adaptively allocated)?

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

MolLingo combines frequency-selected BRICS blocks, optional chemical names, scientific tools, and an agent workflow for molecular optimization. Experiments examine ADMET optimization, docking-guided fragment growing, and the effects of representation, memory, and adaptive search.

Critical strengths
  1. Within-backbone comparisons across three LLMs provide useful evidence for block-based inputs. Additional BRICS, naming, SELFIES, Graph-GA, and REINVENT comparisons broaden the empirical investigation.
  2. The paper separates docking-workflow comparisons from encoding comparisons and acknowledges the larger orchestration budget. Memory ablations examine both optimization outcomes and repeated-edit behavior, offering useful evidence about history-conditioned refinement.
Critical weaknesses
  1. The system-level novelty and independent methodological contributions remain insufficiently established. The paper cites existing tool-augmented and multi-agent chemistry systems, but does not clearly demonstrate what its architecture adds beyond integrating established components. The main benchmarks use only the Chemist Agent, while adaptive orchestration receives a larger sampling budget. Evidence for the proposed representation is also mixed: named and unnamed blocks obtain nearly identical matched-budget joint scores, and BFE improves the docking metric by only 0.569 percentage points over full-fragment SMILES under the same structural context. These findings leave the incremental value of naming, BFE, and multi-agent coordination uncertain.
  2. The evaluation provides limited evidence of meaningful molecular improvement beyond the scores used during optimization. The same ADMET predictors guide and evaluate candidates, without independent validation. Relative improvement in oracle logits may be unstable near zero, while multiplying aggregate improvement and similarity means does not measure per-molecule joint success. Candidate selection and aggregation are insufficiently specified. Consequently, it remains unclear whether individual candidates reliably improve the intended endpoint while satisfying the required molecular constraints.
Questions for authors
  1. What algorithmic contribution distinguishes MolLingo from existing systems such as PharmAgents, MADD, and ChemCrow? How should it be positioned relative to broader engines such as IsoDDE and AnewDDE, accounting for their different scopes?
  2. Does multi-agent coordination or adaptive orchestration outperform a single agent or fixed workflow with identical tools, information, and total candidate budget?
  3. Do BFE and chemical naming provide reproducible gains under matched backbones, budgets, and edit constraints, with paired uncertainty estimates and comparisons against Graph-GA?
  4. Do improvements persist under independent property predictors or docking rescoring, and what fraction of individual molecules simultaneously improve the objective and satisfy prespecified constraints?
  5. Could you provide the exact candidate-selection protocol and implemented BFE procedure, reconciling Algorithm 2’s two-boundary cuts with the five-block example, frequency filtering, and “finest-grained” fallback?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.