acceptodds
Under review as a conference paper at ICLR 2027

MolFiber: From Global Anchors to Local Refinement in Molecular Learning

Abstract

Molecular property prediction increasingly relies on combining complementary representations, yet standard fusion can be ineffective when one representation is already strong. In this regime, the key question is not how to combine all views globally, but where additional representations provide information that the strong representation misses. We introduce **MolFiber**, a Morgan-anchored framework that treats Morgan fingerprints as a global reference geometry for chemical space and uses GraphGrid and MolFormer as local refiners within neighborhoods of Morgan-similar molecules. Rather than acting as competing global predictors, the auxiliary views resolve distinctions that Morgan cannot see locally. **MolFiber-G** uses the full fiber-conditioned auxiliary evidence, whereas **MolFiber-L** uses only locally centered contrast; their ensemble, **MolFiber-ENS**, combines these complementary inductive biases. Across 24 molecular property prediction benchmarks, MolFiber-ENS achieves the strongest aggregate performance among the compared methods and outperforms controlled global-fusion baselines using the same representations. These results support a broader principle: complementary representations can be most valuable not when fused globally, but when used to refine the local ambiguities of a strong anchor.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Molecular property prediction, Multiview learning, Molecular learning, Molecular fingerprints, Local representation refinement

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_Fn2nOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces MolFiber, a molecular property prediction framework that leverages Morgan fingerprints as a global reference anchor while utilizing GraphGrid and MolFormer as local refiners within Morgan-similar neighborhoods. By decomposing auxiliary evidence into local contrast and fiber-level signals, the framework offers two main variants, MolFiber-G and MolFiber-L, along with their high-performing ensemble, MolFiber-ENS. Extensive evaluations across 24 benchmark datasets demonstrate that this asymmetric, anchor-relative organization significantly outperforms traditional global fusion strategies.

Critical strengths

This paper starts from the traditional multi-view fusion methods and clearly distinguishes the roles of different representations: Morgan fingerprints provide a stable reference for global prediction, while GraphGrid and MolFormer offer additional information by comparing local similar molecules. This asymmetric design is intuitive and has potential value, as it provides a principled method that can introduce auxiliary representations without relying on unrestricted feature fusion. The control experiments further support the importance of the proposed neighborhood mechanism.

The evaluation covers 24 molecular property prediction benchmarks and is compared with representative fingerprint-based, graph-structured, pre-trained, and hybrid methods. In addition to overall performance, the authors also conducted ablation experiments on neighborhood construction, neighborhood size, and alternative context correction strategies. These analyses help verify the robustness of the empirical results and provide valuable insights, indicating under what circumstances retaining the neighborhood hierarchy information or emphasizing relative molecular differences is more beneficial for prediction. The complementary behavior of the two variants also prompted the proposal of an integrated design.

Critical weaknesses

The innovation of this method and the underlying concepts it contributes have not been fully verified. Although MolFiber introduces a neighborhood-based strategy to combine molecular representations, its main components still rely on existing concepts, including nearest neighbor retrieval based on similarity, pairwise feature interaction, residual prediction correction, and ensemble learning. Therefore, the main contribution of this paper lies in how these components are organized around an anchor point based on Morgan, rather than proposing a significantly novel learning principle. Although ablation experiments show that local conditioning is beneficial, they do not fully prove why the proposed model is essentially superior to traditional similarity-weighted local models or other neighborhood-based correction mechanisms. This limits the strength of the methodological contribution of the paper.

The experimental results are not yet convincing enough to prove the practical significance and universality of this method. Although this evaluation covers 24 molecular property prediction benchmarks, the overall improvement of MolFiber-ENS compared to existing strong methods is relatively limited. The summary results alone cannot determine whether these improvements are statistically reliable or consistent across different datasets. More importantly, the core motivation of this method is to assist representations in resolving the ambiguity within the neighborhood defined by Morgan. However, the current evaluation does not fully prove that these improvements are mainly concentrated in such challenging cases, nor does it directly quantify the degree to which local ambiguity is resolved. Since this method relies on Morgan fingerprints to define these neighborhood structures, it is necessary to provide stronger evidence to verify the core claim and clearly identify under what circumstances the proposed method can bring meaningful advantages.

Questions for authors
  1. In Appendix A.3, the authors note that for a fiber containing K+1K+1 members, the pair encoder evaluates all ordered pairs, yielding a training complexity of O(NK2)O(NK^2). While K=8K=8 offers a good trade-off and maintains linear scaling with respect to dataset size for fixed small KK, Appendix Table 7 shows that training time increases significantly as dataset size grows (e.g., from ~11s for BACE with n=1,513n=1,513 to ~607s for HIV with n=41,120n=41,120). Furthermore, scaling KK to 16 or 24 causes a quadratic jump in pair computations (K2=256K^2 = 256 and 576576, respectively). The authors should explicitly discuss the scalability limitations of this pair-based approach on ultra-large molecular datasets (e.g., million-scale ZINC or PubChem subsets). Does this pairwise local computation hit severe GPU memory or time bottlenecks in industrial or ultra-large pretraining scenarios?

  2. The entire MolFiber architecture relies heavily on Morgan fingerprints (ECFP4) to define both the global coordinate space and the kk-NN "approximate fibers." However, different fingerprints or representations can define molecular similarity in vastly different ways. If the Morgan anchor fails in certain chemical spaces (e.g., activity cliffs where structurally similar molecules have vastly different properties, or hash collisions distorting ECFP similarity), does the local refinement mechanism misfire? The authors should discuss the sensitivity of their framework to the quality of the global anchor. Would MolFiber remain robust if substituted with alternative global anchors (such as 3D conformation skeletons, different graph kernels, or different fingerprint radii)? Are there specific datasets where a poor Morgan performance severely restricts MolFiber's correction space?

  3. The authors candidly acknowledge (in Section 4.1 and Appendix E.2) that CLAMP performs exceptionally well on TDC datasets because it was pretrained via supervised contrastive learning on ChEMBL bioactivity assay data, leading to domain overlap. However, MolFiber also relies on a frozen pretrained MolFormer (ibm/MoLFormer-XL-both-10pct) and a Morgan predictor. The authors need to further clarify whether the auxiliary views (MolFormer) introduce similar domain overlap or data leakage risks. Specifically, when fine-tuning the projection heads PGP_G and PFP_F, does the model implicitly harvest downstream-task-related priors from the pretraining representations? Providing controlled discussions or experiments on pretraining-to-downstream domain gaps would strengthen this point.

7 comments
AuthorsOpenReview ↗

Responses to Reviewer Fn2n - part 1

We thank the reviewer for their careful and constructive feedback. We have revised the paper accordingly (updated PDF) and summarize the main changes in our general response above; we address each point individually below.

W1: Novelty

The innovation of this method and the underlying concepts it contributes have not been fully verified.

We thank the reviewer for raising this concern. We agree the components — similarity-based nearest-neighbor retrieval, pairwise feature interaction, residual correction, ensembling — are individually established, and we claim none as a new primitive. The contribution is the organizing principle, and the reviewer's sharper question — why this is superior to a traditional similarity-weighted local model or another neighborhood-correction mechanism — is the one we can answer concretely.

The key distinction: the neighborhood and its resolution come from different representations. A traditional similarity-weighted local model (e.g., weighted k-NN) predicts from neighbor labels on the same representation that defines the neighborhood. On this benchmark that representation is Morgan — and that is exactly the model that fails on activity cliffs: when a query's Morgan-nearest neighbors carry conflicting labels, label-weighting predicts the neighborhood consensus and gets the cliff wrong. MolFiber breaks this circularity. The neighborhood is defined by Morgan, but the resolution comes from different views — GraphGrid and MolFormer — that see distinctions invisible to the Morgan geometry, and retrieval is label-free: the readout never uses neighbor labels, only fiber-relative comparisons of the auxiliary embeddings (App. A.2). A similarity-weighted local model on Morgan cannot do this, because it has only Morgan.

This is not a conceptual claim alone — it is what the cliff analysis shows. The frozen Morgan anchor is itself a similarity-based global model; on activity cliffs it errs on 1,557 molecules, and MolFiber corrects 287 of them (18.4%) while breaking only 4.9% of those it had right (net +109, p<10⁻⁶; App. E.1), using the auxiliary views to separate molecules Morgan places together. No Morgan-only local model could produce these corrections.

We also control against the alternatives directly. Replacing the Morgan neighborhood with a random one of the same size collapses the gain (App. D.5), so it is the Morgan-confused neighborhood specifically that matters, not generic locality. Removing the fiber entirely while keeping the anchor, views, and bounded residual (Anchor-MLP) drops OGB AUC from 80.15 to 77.68 (App. D.1). And alternative neighborhood-correction mechanisms — learned per-task or per-molecule retention, and per-view routing — do not beat the simple design (App. D.2).

These choices are domain-specific, not generic: Morgan is the anchor because fair benchmarking shows it is cheap and hard to beat (Praski et al.); the fibers are Morgan–Tanimoto neighborhoods because that is the medicinal-chemistry definition of an activity cliff; and the bounded, zero-initialized residual encodes the prior that a weaker view should refine, not overturn, a strong anchor. The result is a modular recipe — swapping the anchor for Atom Pair or MACCS keeps MolFiber-ENS at 79.3–80.2 AUC (App. D.4).

We are glad to state explicitly that our contribution is a domain-motivated, empirically validated framework — a fusion method engineered for the activity-cliff structure of MPP, in which the neighborhood and its resolution come from different representations — rather than a new ML primitive. That is the right level at which to judge the novelty.

Revision: We added the controls and evidence that distinguish MolFiber from single-representation local models and other neighborhood-correction mechanisms — the cliff-correction analysis (App. E.1), the random-neighborhood control (App. D.5), the Anchor-MLP/fusion controls (App. D.1), the learned-variant comparison (App. D.2), and the anchor-sensitivity ablation (App. D.4).

AuthorsOpenReview ↗

Responses to Reviewer Fn2n - part 2

W2: Experimental results

The experimental results are not yet convincing enough to prove the practical significance and universality of this method.

We thank the reviewer. This raises three concerns — statistical reliability and consistency, whether the gains concentrate on the hard cases, and when the method helps — and we address each with evidence beyond the summary table.

Reliability and consistency. We do not rely on summary numbers alone. Per-dataset results with standard deviations over n=5 seeds are reported in full (App. F.1–F.2), and significance is tested directly (App. D.1): MolFiber-ENS is significantly better than the fusion controls under Welch's test with Holm–Bonferroni correction and a Fisher omnibus. Consistency is in fact one of MolFiber's clearest advantages: every MolFiber variant stays within ±1.4 AUC across seeds, whereas strong baselines swing by ±3–6 and occasionally ±10–17 (e.g., GROVER on ClinTox, CDDD on HIA). Low, stable variance is what makes the ranking trustworthy rather than an artifact of a lucky seed.

The "limited" overall margin is limited only against one outlier. MolFiber-ENS (82.10) edges CLAMP (82.01) — but CLAMP is a specially-pretrained outlier standing nearly a full point above every other baseline. Against the rest of the field the margin is decisive: +1.13 AUC over the next-best method and average rank 6.46 → 4.77. The gain is also intentionally bounded — MolFiber refines a strong anchor, it does not replace it — so a large aggregate jump is not the right expectation; a reliable, broad improvement with rare regressions is.

Does it concentrate on the hard cases, and by how much? We measure it directly (App. E.1). This is the core of the reviewer's concern, and rather than infer it from aggregate AUC we go molecule-level and count. We isolate the Morgan-ambiguous molecules — activity cliffs, where a near-identical training neighbor carries the opposite label — across all 24 datasets (5,222 of them). On exactly these cases the Morgan anchor errs 1,557 times; MolFiber corrects 287 (18.4% of its cliff errors) and breaks only 178 of the 3,665 it had right (4.9%), net +109, raising cliff accuracy 70.2% → 72.3% (sign test p<10⁻⁶). This directly quantifies the degree of local-ambiguity resolution the reviewer asks for, and App. E.2 shows it molecule by molecule.

When does it help? The same analysis tells us. The benefit concentrates where Morgan-ambiguous cliffs are common (e.g., hERG_Karim +30 net, CYP2C19 +17) and is flat where cliffs or anchor errors are few (BACE, ClinTox) — net-positive on 15 datasets, tied on 7, negative on 2. We report this in full rather than omit it. The other condition is the anchor itself: the anchor-sensitivity ablation (App. D.4) shows the method helps across strong fingerprint anchors and weakens only with one lacking a meaningful similarity geometry. The circumstances under which MolFiber brings an advantage are therefore explicit and tested: a strong fingerprint anchor, applied to property spaces where Morgan-local ambiguity is present.

Revision: added the cliff-level correction analysis across all 24 datasets quantifying local-ambiguity resolution (App. E.1) with molecule-level examples (App. E.2); per-dataset results and significance tests are in App. F.1–F.2 and D.1, and the anchor-sensitivity boundary in App. D.4.

AuthorsOpenReview ↗

Responses to Reviewer Fn2n - part 3

Q1: Scalability

The authors should explicitly discuss the scalability limitations of this pair-based approach on ultra-large molecular datasets (e.g., million-scale ZINC or PubChem subsets).

We thank the reviewer. The worry is natural — "pairwise" suggests quadratic cost — but the pairwise computation here is local: each molecule compares only against its own K = 8 neighbors, never against the dataset. Concretely, MolFiber's cost splits into a one-time stage computed once per dataset and cached, and a lightweight training stage that reuses it (Table 7, App. C.4).

One-time cached stage. Fingerprints and the Morgan anchor, GraphGrid tensors, frozen MolFormer embeddings, and the Morgan–Tanimoto fibers are each computed once and reused across all variants, seeds, and baselines. On HIV (n ≈ 41k) this is 4,699 s, of which the MolFormer embedding is 84% (3,930 s) — a cost the MolFormer baseline and the fusion controls pay identically. The only MolFiber-specific part, fiber construction, is just 134 s (2.8%).

Training stage: linear in N. With cached inputs, training requires O(NK²) pair-encoder evaluations, and since K is a small fixed constant, this is linear in dataset size. The runtimes confirm it: MolFiber-G/-L trains in 11 s (BACE), 56 s (Tox21), and 298 s (HIV) — about 7.2 ms per molecule across a 27× range in dataset size, i.e., a per-molecule cost independent of N. This is only 5–8% of the one-time cost. MolFiber-ENS is exactly twice a single model (G and L train independently and in parallel), and each extra seed or variant adds only this marginal amount.

GPU memory is independent of N. We never materialize a dataset-wide pairwise matrix. Each fiber is (K+1)² = 81 pairs, processed one minibatch at a time (batch 64), so peak memory is a fixed footprint set by batch × K² — it does not grow with dataset size. The only quadratic factor is K², and K is a knob we never scale up: App. D.3 (Table 4) shows K = 8 captures essentially all the benefit (K = 16, 24 cost ~4× and ~9× the pairwise computation for at most +0.05 AUC).

At million-scale (ZINC/PubChem). The two dominant costs are one-time feature extraction — the same embedding any method using these representations pays — and nearest-neighbor retrieval. For the latter, exact scanning can be replaced by approximate NN search on the binary fingerprints (HNSW, or MinHash-based LSH for Tanimoto): an index built once over the training pool lets each query examine a small candidate set. Retrieval is label-free and cached, so this swap leaves the linear, fixed-memory O(NK²) training stage unchanged. We have evaluated up to n ≈ 41k rather than 10⁶, so we present this as a scalability analysis, but the method has no inherent super-linear-time or N²-memory bottleneck.

Revision: we added a scalability discussion (App. A.3/C.4) noting the one-time cached stage vs. the linear, fixed-memory training stage (observed ~7.2 ms/molecule across a 27× dataset range), that peak GPU memory is independent of N, and that million-scale retrieval can use approximate-NN indices while leaving the O(NK²) training unchanged.

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

A frozen Morgan random forest is used as an anchor to provide global predictions, while GraphGrid and MolFormer perform bounded residual correction (|Δ| ≤ λ, initialized to zero) only within the approximate fibers formed by Morgan-Tanimoto neighbors, i.e., S(x) = b_M(x) + λ·tanh(u). The context score within a fiber can be decomposed into local contrast + fiber-level signal; the two endpoints correspond to MolFiber-G (retaining all) and MolFiber-L (retaining only the mean-subtracted contrast), respectively, which are integrated into ENS. ENS achieves the best aggregate AUC/rank/gap (82.10) across 24 OGB/TDC datasets, and controlled ablation experiments (Anchor-MLP, early/late fusion) indicate that the gains stem from the fiber-based architecture rather than mere multi-view fusion.

Critical strengths

S1. The auxiliary representation performs bounded residual correction only within Morgan’s neighborhood (fibers); the zero-initialization + tanh bounded residual formulation is clean, and the unified ρ-family decomposition of G/L is concise and interpretable.

S2. Anchor-MLP removes the fiber structure while retaining anchors and bounded residuals (80.15 vs. 77.68), directly demonstrating that the gains stem from the fiber organization rather than the multi-view approach itself; with 24 datasets and open-source code, the model exhibits significantly lower variance than the baseline and demonstrates good stability.

Critical weaknesses

W1. The improvement over the Relative Best Baseline (CLAMP) is minimal: the overall AUC is only 0.09 higher, and it still underperforms CLAMP on TDC; results vary across datasets, and the conclusion that it is the strongest aggregate is primarily based on a slight difference in average rankings.

W2. All baseline figures are taken from an external study (Praski et al.) and were not rerun under the protocol described in this paper; mean/variance and stability cannot be strictly compared across papers, and the credibility of the conclusions depends on the reliability of others’ data.

W3. Core premises remain unverified: No attempts were made to use anchors other than Morgan or alternative definitions of fingerprints/similarity; the applicable boundaries for “how strong an anchor must be to work” remain unknown, and hyperparameters such as k=8 and λ=4 lack systematic ablation analysis.

W4. Lack of evidence for mechanistic claims: The study is motivated by the activity cliff and Morgan’s local ambiguity, but no analysis indicates that the correction terms are concentrated on molecules with ambiguous anchors or high intra-fiber divergence; furthermore, G significantly outperforms L, and variants that learn ρ all fail, creating tension with the narrative that local contrast is key, yet this tension is not discussed.

Questions for authors

Q1. What happens if we replace the anchor representation (e.g., using MolFormer as the anchor) or change the definition of fingerprints/similarity? How sensitive is the framework to the strength of the anchors?

Q2. Are the correction terms concentrated on molecules with large Morgan anchor errors or significant intra-fiber label discrepancies? Please directly validate the “activity-cliff” motivation.

Q3. How were λ=4, ECFP4(r=2), and k=8 selected? Since the absolute embedding product terms (g⊙g, m⊙m) in r carry global information, is the claim that the auxiliary view has no global pathways accurate? Would performance deteriorate if the product terms were removed, leaving only the difference terms?

Q4. Given that G significantly outperforms L and learning ρ does not exceed G, what is the practical value of this local de-averaging comparison bias? Does the gain from ENS primarily stem from a simple average of the two models?

6 comments
AuthorsOpenReview ↗

Responses to Reviewer zzYV - part 1

We thank the reviewer for their careful and constructive feedback. We have revised the paper accordingly (updated PDF) and summarize the main changes in our general response above; we address each point individually below.

We want to reframe this comparison, because we think it strengthens the paper rather than weakening it — and we address the TDC and "slight ranking" points directly.

CLAMP is not a typical baseline — it is a clear outlier at the top. On the full 24-dataset suite, CLAMP reaches 82.01 AUC while the best of the other twelve baselines (R-MAT) reaches only 80.97, and the rest of the field sits below that. CLAMP stands roughly a full AUC point above every other published baseline — a wide margin on a benchmark where methods routinely separate by tenths. It is, by some distance, the strongest competitor, and we treat it as the bar to clear, not one baseline among thirteen.

Against that bar MolFiber is competitive; against everyone else it is decisively ahead. MolFiber-ENS (82.10) edges CLAMP on all three aggregate metrics overall (AUC 82.10 vs. 82.01, rank 4.77 vs. 5.63, gap 3.52 vs. 3.61) and leads on OGB by nearly a point (80.16 vs. 79.18); even the single-model MolFiber-G matches CLAMP's overall AUC (82.01). The ranking advantage is not a sliver — it holds across AUC, rank, and gap simultaneously, and MolFiber is +1.13 AUC over the next-best non-CLAMP baseline, improving average rank from 6.46 to 4.77. On the TDC point specifically: CLAMP's TDC lead is in raw AUC (82.96 vs. 82.75) and is expected — CLAMP is contrastively pretrained on ChEMBL bioactivity assays, whose domain overlaps the ADME/toxicity endpoints that dominate TDC — yet MolFiber-ENS still attains the best TDC rank (4.47 vs. 5.36). On OGB, where that domain overlap does not apply, MolFiber leads outright. So "underperforms on TDC" holds only for one metric on the subset most favorable to CLAMP's pretraining.

MolFiber is not one model but a small, interpretable family. MolFiber-G retains the full fiber-conditioned evidence; MolFiber-L keeps only the locally-centered contrast (invariant to a common shift within a fiber); and MolFiber-ENS averages the two, combining complementary inductive biases — absolute fiber evidence versus purely relative local correction. This is why the method is robust: it does not rest on a single modeling assumption holding everywhere.

Crucially, MolFiber and CLAMP reach the top by opposite routes. CLAMP buys its performance with data: supervised contrastive pretraining on a large, task-adjacent bioactivity corpus. MolFiber buys it with structure: an anchor-relative refinement built on frozen, off-the-shelf components (a random-forest Morgan anchor and a frozen MolFormer embedding), with no bioactivity pretraining and no task-matched supervision. Reaching CLAMP's level from this direction, without its data advantage, is the result.

Why this matters for the field — and why MolFiber is the more hands-on contribution. Combining molecular representations has, in practice, disappointed: fair benchmarking shows naive fusion often fails to beat a single fingerprint, and our own controls confirm it (early/late fusion 79.2/79.0 vs. 80.15 for MolFiber-G). MolFiber is a practical recipe for getting genuine synergy instead: it needs no pretraining corpus, runs with frozen off-the-shelf encoders and one default configuration (no per-dataset tuning), and is modular — one can improve the model by swapping in a stronger anchor or auxiliary view (App. D.4), including CLAMP's own embedding. CLAMP, by contrast, reaches its result only where a large task-matched pretraining corpus exists; for the many properties where it does not, that route is unavailable, while MolFiber's is not. The two are therefore complementary, not competing — CLAMP can serve as an anchor or a view inside MolFiber — and CLAMP's own design, which keeps a fingerprint encoder as a load-bearing component, corroborates our central premise that fingerprints remain hard to beat except through careful conditioning of an auxiliary signal.

Revision: We revised the main results (Sec. 4.1) positioning CLAMP as a clear top-end outlier and noting that MolFiber-ENS matches it while using none of its task-matched bioactivity pretraining, which makes the two complementary rather than competing.

AuthorsOpenReview ↗

Responses to Reviewer zzYV - part 2

W2: External baseline report

All baseline figures are taken from an external study (Praski et al.)

We thank the reviewer for this point. We think this choice makes the comparison stronger, not weaker — and we handle the one narrow point where the reviewer is right transparently.

The source is a benchmark built for exactly this. Praski et al. is a dedicated fair comparison of 57 embedder variants on the same 24 datasets we use, under the same scaffold splits and the same classifier heads. These are not figures lifted from thirteen papers with thirteen protocols; they are one standardized, internally consistent evaluation. Just as importantly, using an independent benchmark removes the most serious comparability risk — author bias. Had we re-run the baselines ourselves, we would be open to the standard objection that we tuned our own method more carefully than the competitors; sourcing their numbers from a study whose explicit goal was fair evaluation makes that impossible. We regard this as a feature of the design.

We read the benchmark conservatively. Rather than taking a per-dataset maximum over all 57 variants — which manufactures an oracle "best baseline" on every column — we fixed one representative variant per family by aggregate AUROC and used it consistently (App. C.2). If anything this makes the baselines harder to beat.

MolFiber is evaluated under that same protocol, so the comparison is like-for-like. As a benchmark paper, Praski et al. release their full code and every protocol detail, so reproducing their evaluation exactly is both possible and verifiable — not a matter of approximating a described setup. We follow it directly: the same 24 datasets, the same scaffold splits, and the same n=5 mean±std reporting, and within our pipeline the splits, fingerprints, anchor logits, and auxiliary features are byte-identical across every method and seed (App. C.3). Our numbers are comparable to the benchmark's because we reproduced its published protocol, not a different one.

The reviewer is right on one specific point, and we address it explicitly. Since we did not re-run the baselines, their per-seed values are unavailable, so a paired test cannot be reconstructed. This is exactly why our significance analysis (App. D.1) uses Welch's unpaired t-test from the reported means, standard deviations, and n=5, with Holm–Bonferroni correction and a Fisher omnibus — stated as a limitation — and our stability claims are made within the benchmark's own reported-variance framework, applied identically to every method.

If the reviewer would find it reassuring, we are happy to re-run the strongest baselines (e.g., CLAMP, R-MAT) under our exact pipeline as a spot-check, though we believe the independent benchmark is the more credible basis for the main comparison.


W4: Hyperparameter sensitivity

Core hyperparameter sensitivity (k=8, λ=4) and applicability boundaries across alternative anchors/fingerprints remain unverified.

We thank the reviewer for this important point. The two hyperparameters the reviewer names are both already studied.

Fiber size k is systematically ablated (App. D.3, Table 12). Across K ∈ {1, 8, 16, 24} on the six OGB datasets, performance rises sharply from K = 1 to K = 8 (MolFiber-ENS 79.75 → 80.16; MolFiber-L gains 1.18), then plateaus — K = 8, 16, 24 span only 0.13 AUC (80.16 / 80.21 / 80.08). The method is insensitive to k once the fiber holds enough neighbors; we use K = 8 as it matches K = 16 within 0.05 AUC at a quarter of the pairwise cost (K² = 64 vs. 256).

λ and the other hyperparameters were covered by a randomized search showing no sensitivity. We searched over λ ∈ {1, 2, 4}, k ∈ {4, 8, 16}, the temperatures τ and τ_c, learning rate, dropout, widths, and batch size, scored on held-out training-pool data (never the test split). Performance varied by less than 1% AUC on every dataset, so we report one default (k = 8, λ = 4, both within the grid) across all 24 datasets with no per-dataset tuning — the absence of tuning is itself evidence of robustness. We add a sentence to App. D.3 documenting this search.

Applicability across anchors/fingerprints is tested in App. D.4. With four anchors (Atom Pair, MACCS, MolFormer, Morgan), MolFiber is robust across the fingerprint anchors (MolFiber-ENS 79.3–80.2 AUC, Morgan best) and weakens only with a non-fingerprint anchor lacking a natural similarity geometry. The boundary is thus clear: the method wants an anchor that is both a strong predictor and a meaningful similarity metric.

Revision: added a sentence to App. D.3 documenting the randomized hyperparameter search (λ, temperatures, optimization settings); the fiber-size ablation was already in D.3 and the anchor/fingerprint study is in D.4.

AuthorsOpenReview ↗

Responses to Reviewer zzYV - part 3

W4: Activity Cliffs and Evidence

Lack of evidence for mechanistic claims regarding local contrast and activity cliffs, along with unaddressed tensions where global models outperform local ones.

We thank the reviewer for this point. We address both halves of this concern — the mechanistic evidence and the apparent tension between the variants.

Mechanistic evidence for the activity-cliff claim. We test the mechanism directly on exactly the high-similarity, opposite-label pairs that define the regime, across all 24 datasets (App. E.1). There are 5,222 such cliffs; the frozen Morgan anchor errs on 1,557, and MolFiber corrects 287 (18.4% of its cliff errors) while breaking only 178 of the 3,665 it had right (4.9%) — net +109, cliff accuracy 70.2% → 72.3%, sign test p < 10⁻⁶. Appendix E.2 shows individual molecules: the anchor's probability sitting on an opposite-label neighbor's side of the boundary, and MolFiber moving it back. This is direct evidence, not an inference from aggregate accuracy.

On the G-vs-L tension. The reviewer reads "G significantly outperforms L, and learned-ρ variants do not win" as evidence against local contrast mattering. But our claim is not that the purely-centered contrast (L) is the best predictor; it is that auxiliary evidence conditioned on Morgan-local neighborhoods refines the anchor. G and L are the two endpoints of one decomposition — the fiber-conditioned score splits into local contrast plus a fiber-level signal, and L keeps only the first term while G keeps both. That G > L simply means the fiber-level signal is usually informative, so discarding it unconditionally (as L does) loses more than it protects (App. D.2). That is a finding about the decomposition, not a contradiction of it.

The decisive point is the ensemble. MolFiber-ENS — the plain average of G and L — beats G alone on every aggregate metric (overall AUC 82.10 vs. 82.01, rank 4.77 vs. 5.19, gap 3.52 vs. 3.62) and on all nine Overall/OGB/TDC summary cells in App. D.2. If L's local-contrast component were redundant or noise, averaging it in could not improve on G — the fact that it does is direct evidence that local contrast carries complementary signal the full-evidence endpoint does not fully exploit. The two endpoints make partly uncorrelated errors, which is exactly why combining them is strongest.

Finally, the learned-ρ variants (M, A) not exceeding the fixed endpoints is a parsimony result, not a failure of the mechanism: the two interpretable endpoints and their ensemble already capture the useful structure, so the added per-task or per-molecule estimation machinery does not pay for itself (App. D.2).

Revision: added the cliff-level correction analysis across all 24 datasets (App. E.1) and concrete molecule-level examples (App. E.2) as direct evidence for the mechanism; the G/L/ENS and learned-ρ comparison that resolves the variant tension is in App. D.2.

Reviewer_z2wxOpenReview ↗
Rating
3
Confidence
1
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces MolFiber, a Morgan-anchored framework that treats Morgan fingerprints as a global reference geometry for chemical space and uses GraphGrid and MolFormer as local refiners within neighborhoods of Morgan-similar molecules. Rather than allowing the auxiliary views to act as competing global predictors, the method uses them to resolve distinctions that Morgan may miss locally. The paper evaluates the method on 24 molecular classification datasets.

Critical strengths
  • The paper considers a realistic problem: when Morgan fingerprints already provide a strong predictor, additional representations may be more useful for correcting local ambiguities than for unrestricted global fusion. This is a reasonable and worthwhile perspective.

  • The paper focuses on how auxiliary representations can complement information that the anchor representation fails to distinguish, rather than simply increasing the number of input representations. This is a meaningful question in multirepresentation learning.

  • The frozen Morgan predictor, zero-initialized residual branch, and bounded correction allow the model to refine the anchor prediction without completely replacing it.

  • The Anchor-MLP baseline preserves the anchor, auxiliary representations, and bounded residual while removing the neighborhood comparison. This is a useful control for assessing the contribution of the proposed local conditioning.

Critical weaknesses
  • The main components of the method—Morgan k-nearest-neighbor neighborhoods, pairwise features, neighborhood aggregation, and bounded residual corrections—are individually familiar. The proposed “fiber” mainly reorganizes these components around a local neighborhood, while the score decomposition is largely an algebraic reformulation. The contribution would therefore be better characterized as a motivated framework for locally conditioned fusion rather than a new theory.

  • The paper argues that MolFiber mainly corrects cases that Morgan cannot distinguish, but this mechanism is not directly validated. The experiments do not show whether the largest improvements occur for highly Morgan-similar molecules with substantially different properties. There is also no analysis relating the correction magnitude to neighborhood similarity, label disagreement, or activity-cliff measures. The current results show that the model can improve prediction, but they do not establish that local disambiguation is the reason for the improvement.

  • In Table 2, Morgan-RF achieves an average AUC of 72.94, below MolFormer at 77.20 and below the ECFP baseline in Table 1 at 76.67. This raises the possibility that part of MolFiber’s gain comes from compensating for a relatively weak anchor, rather than demonstrating the effectiveness of local refinement on top of a strong Morgan predictor.

  • The current ablations do not sufficiently separate the effects of the Morgan neighborhood, the pairwise encoder, GraphGrid, MolFormer, and generic ensembling. Additional controls, such as random neighborhoods, alternative neighborhood definitions, single-view variants, and an equal-budget ensemble of two MolFiber-G models, would be useful.

Questions for authors
  1. Does MolFiber retain its advantage when the anchor is replaced by a stronger and more carefully tuned Morgan/ECFP predictor?

  2. Can the authors provide concrete molecular-neighborhood examples showing that the method improves predictions precisely where Morgan assigns high similarity but the molecular properties differ?

  3. Can the authors disentangle the contributions of the neighborhood construction, pairwise encoder, GraphGrid, MolFormer, and ensembling through additional ablations?

8 comments
AuthorsOpenReview ↗

Responses to Reviewer z2wx - part 1

We thank the reviewer for their careful and constructive feedback. We have revised the paper accordingly (updated PDF) and summarize the main changes in our general response above; we address each point individually below.

Novelty

The main components of the method are individually familiar.

We thank the reviewer, and we want to use this point to state the contribution more sharply, because the "heuristic local fusion" reading misses what the method is for.

MolFiber is not another fusion method — it is a fusion method built around the one failure mode specific to this domain. Standard multi-view fusion is domain-agnostic: it concatenates or attends over all representations and lets the model learn from data when each view is trustworthy, treating molecular property prediction like any other multi-view problem. But this domain has a specific, well-known Achilles' heel — activity cliffs: pairs of molecules a fingerprint places almost on top of each other yet whose properties differ sharply. Cliffs are exactly where fingerprint-based prediction demonstrably breaks, and they are the central difficulty medicinal chemistry cares about. MolFiber goes at this head-on. Morgan provides a strong, cheap global map of chemical space; the auxiliary graph and transformer views are deployed only inside the neighborhoods where that map is locally self-contradictory — i.e., on cliffs — to resolve precisely the distinctions Morgan cannot see. Fusion is directed where it is needed rather than applied blindly everywhere. We think of this as smart fusion: a fusion recipe shaped by the structure of the problem, not agnostic to it.

The three design choices follow from domain facts, not generic intuition:

  1. The anchor is not arbitrary. Fair benchmarking (Praski et al.) shows Morgan/ECFP is cheap, learning-free, and still hard to beat — which is what licenses treating one view as the global reference rather than one co-equal input among many.
  2. The neighborhood is the phenomenon. Morgan–Tanimoto neighborhoods are the very definition used to identify activity cliffs. Our "fibers" are not a generic similarity trick; they are exactly the regions where the strong anchor is known to fail, and the auxiliary views are given access only there.
  3. The bounded, zero-initialized residual encodes the domain prior that a weaker view should never overturn a strong cheap predictor globally — only refine it locally.

This delivers the synergy that flat fusion had not. In this domain, naively combining strong and weak views tends to dilute rather than help, and our own controls confirm it: early/late fusion of the same three views reach 79.2 / 79.0 AUC on OGB, and an Anchor-MLP that keeps the anchor, the views, and the bounded residual but removes the fibers drops to 77.68 — versus 80.15 for MolFiber-G (App. D.1). With the fiber organization, each view does what it is best at — Morgan carries the global geometry and prediction, the auxiliary views carry local disambiguation — and the combination exceeds the parts, which is what "combining representations" in MPP had struggled to achieve.

Because the organization is modular, it is a recipe, not a fixed architecture. The anchor is a replaceable slot: swapping Morgan for Atom Pair or MACCS keeps MolFiber-ENS at 79.3–80.2 AUC (App. D.4), so the model can be improved simply by substituting a stronger component — and the same holds for the auxiliary views. This is what makes MolFiber a framework rather than a one-off model.

The claim is empirical, not heuristic, and we validate the mechanism directly: across all 24 datasets, on exactly the activity-cliff molecules where Morgan's neighborhood is self-contradictory, MolFiber corrects 18.4% of the anchor's errors while breaking only 4.9% of its correct predictions (net +109; sign test p<10−6p<10^{-6}; App. E.1), with concrete molecule-level examples in App. E.2 — the anchor sitting on an opposite-label neighbor's side of the boundary, and MolFiber moving it back.

We are glad to state explicitly in the paper that our contribution is a domain-motivated, empirically validated framework — a fusion method engineered for the activity-cliff structure of MPP — rather than a new ML primitive. That is the right level at which to judge the novelty, and at that level the contribution is clear.

AuthorsOpenReview ↗

Responses to Reviewer z2wx - part 2

Activity Cliffs and Validation

The paper argues that MolFiber mainly corrects cases that Morgan cannot distinguish, but this mechanism is not directly validated.

This is the sharpest possible test of our claim, and we have run it directly — on exactly the high-similarity, opposite-label pairs the reviewer names, across all 24 datasets rather than a favorable subset. The full analysis is Appendix E.1; molecule-level examples are Appendix E.2.

Setup. We flag a test molecule as an activity cliff when its nearest opposite-label training neighbor has Morgan–Tanimoto similarity sM≥τs_M \geq \tau: Morgan's own neighborhood then contains a near-identical molecule with the opposite label, so the frozen anchor is being pulled toward a conflicting example and cannot separate the two. Cliff membership depends only on structure and labels — never on whether the anchor or the model happens to be right — so the test cannot be biased toward MolFiber.

Result. Across the 24 datasets there are 5,222 cliffs. The frozen Morgan anchor errs on 1,557 of them; MolFiber-G corrects 287 (18.4% of the anchor's cliff errors) and breaks only 178 of the 3,665 it had right (4.9%), for a net of +109 and a cliff-accuracy gain of 70.2% → 72.3%. A sign test on corrections versus breaks is decisive (p<10−6p<10^{-6}). The table restricts to the ten datasets with at least 30 anchor errors on cliffs — an outcome-independent threshold, since it filters on error count, not on the correction outcome — where a per-dataset rate is meaningful; its pooled rate (18.4%) equals the all-24 figure, so the effect is not an artifact of low-count datasets. The complete 24-dataset breakdown is in Appendix E.1, Table E.1.

DatasetCliffs (N)Anchor wrongCorrectedBrokenNetCorr./Wrong
hERG_Karim8132275727+3025.1%
CYP2C19_Veith411913013+1733.0%
SIDER7193083724+1312.0%
CYP3A4_Veith7311903727+1019.5%
HIV18757156+926.3%
CYP2C9_Veith5451322518+718.9%
CYP2D6_Veith500862215+725.6%
CYP1A2_Veith330591917+232.2%
Tox21347156141409.0%
AMES32412168−25.0%
Pooled49071427262 (18.4%)169 (4.9%)+93

What this shows. The two numbers that matter move together. Where Morgan fails — its neighborhood is self-contradictory — MolFiber corrects a substantial fraction of its errors (287), and it does so while almost never damaging the predictions Morgan already gets right (178 breaks, a 4.9% rate that is stable across datasets). That asymmetry is the mechanism the paper claims: the auxiliary views supply exactly the local disambiguation Morgan cannot, directed only at the cliffs. It is also consistent with the architecture by construction — the readout receives only the aggregated fiber-relative context and never the absolute embeddings or the anchor score (App. A.2), so every correction is produced by within-fiber comparison.

We also report the result honestly in full: the effect is partial (≈18%, not a majority) and is net-zero or slightly negative on a few datasets (BACE, Tox21, AMES), shown rather than omitted. This is the intended behavior of a bounded refinement — the zero-initialized, λ\lambda-bounded residual nudges the anchor where the auxiliary views carry local signal but cannot override a strong anchor globally, so it neither resolves every cliff nor diverges from the anchor when the local evidence is uninformative. Appendix E.2 makes this concrete at the molecule level: the anchor's probability sitting on the opposite-label neighbor's side of the boundary, and MolFiber moving it back.

Revision: added Appendix E — E.1 provides the cliff-level correction analysis across all 24 datasets (Table E.1), and E.2 gives concrete molecule-level activity-cliff examples.

AuthorsOpenReview ↗

Responses to Reviewer z2wx - part 3

W3: Morgan performance as anchor

In Table 2, Morgan-RF achieves an average AUC of 72.94, below MolFormer at 77.20 and below the ECFP baseline in Table 1 at 76.67.

We thank the reviewer, and we think this concern dissolves once three distinct objects are separated — and that, properly understood, it reframes our contribution rather than undercutting it.

The comparison is not like-for-like. Our anchor is a deliberately simple, untuned 200-tree random forest on binary ECFP4 — the plainest possible strong predictor, chosen so that any gain is attributable to the refinement, not to anchor engineering. The ECFP (76.67) and MolFormer (77.20) rows the reviewer cites are from Praski et al.: the count-based ECFP variant and the MolFormer embedding, each with a classifier head tuned and selected per dataset. Their higher standalone scores reflect that tuning and selection, not a defect in our anchor. We could have adopted a tuned anchor and scored higher; we intentionally did not.

The anchor's role is dual — predictor and geometry — which is where Morgan is uniquely suited. The anchor idea requires more than a strong global predictor: it requires a meaningful parametrization of molecule space in which to define the local neighborhoods where the auxiliary views operate. Morgan supplies both. It is a strong predictor, and it comes with a canonical, chemically grounded metric — Tanimoto similarity — which is the very measure medicinal chemistry uses to define structural neighborhoods and activity cliffs. Most alternatives do not: a dense MolFormer embedding admits no unambiguous similarity (cosine? Euclidean?) and none chemically interpretable the way Tanimoto is. This is visible in our anchor-sensitivity ablation (App. D.4): the fingerprint anchors, which carry natural similarity metrics, all support strong MolFiber performance (ENS 79.3–80.2 AUC), whereas MolFormer — a competitive representation but one without a well-defined geometry for the fibers — is the weakest anchor (76.41). That Morgan remains hard to beat despite its simplicity (Praski et al.) is precisely why it is the right organizing anchor: not because it is the strongest predictor in isolation, but because it couples a strong, cheap predictor with a well-defined geometry. The open problem in this domain is not whether Morgan is strong enough on its own, but how to combine these complementary models so their strengths compound — which is exactly what MolFiber addresses.

Even from this simple anchor, the full method surpasses the tuned baselines. Across all 24 datasets, MolFiber-ENS reaches 82.10 AUC — above the tuned ECFP (80.14) and MolFormer (79.80) baselines (Table 1). The gain cannot be read as merely closing a gap to a weak anchor; starting from the simple anchor, the method ends up ahead of the strong, tuned single-representation models.

The control the reviewer's hypothesis predicts already exists — and it fails. If the gain were just extra capacity compensating for a weak anchor, Anchor-MLP — which keeps the identical frozen Morgan-RF anchor, the same bounded residual, and the same auxiliary views, and removes only the fibers — should recover it. It does not: 77.68 versus 80.15 for MolFiber-G (Table 2), with significant Fisher-combined tests against Early Fusion, Late Fusion, and Anchor-MLP. The same anchor and views, organized globally instead of through fibers, leave 2.5 AUC on the table.

Finally, cliffs are a limit of the fingerprint representation, not of the anchor's training. Two molecules near-identical in ECFP space receive the same anchor prediction however well the predictor is tuned — one cannot tune a random forest out of an activity cliff. MolFiber's corrections are concentrated exactly there (App. E.1: 18.4% of the anchor's cliff errors corrected, 4.9% broken, net +109, p<10−6p<10^{-6}), resolving distinctions no stronger or better-calibrated fingerprint anchor could reach. That is categorically different from patching an undertrained predictor.

Revision: added Appendix D.4 (anchor-swap ablation across four anchors) and Appendix E.1 (cliff-level correction analysis), which together show the gain comes from the fiber organization and from resolving representation-level cliffs — and that the anchor's value lies in coupling a strong predictor with a well-defined Tanimoto geometry, not in raw predictor strength.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Rebuttal

We are grateful to all three reviewers for their careful, constructive, and genuinely engaged reviews — the questions pushed us to sharpen both the framing and the evidence, and the paper is stronger for them.

What MolFiber is. MolFiber is not another generic fusion method; it is a fusion method built around the one failure mode specific to molecular property prediction — activity cliffs, where molecules that a fingerprint places almost on top of each other have sharply different properties. A strong, cheap Morgan anchor charts the global space and supplies the Tanimoto geometry, and the auxiliary views are deployed only inside the local neighborhoods where that anchor is ambiguous, to resolve what it cannot see. The novelty is this asymmetric, domain-motivated organization — and the evidence that it works as claimed — rather than any single component.

What we added in response to your reviews. The concerns converged, so our main additions help several of you at once:

  • Direct mechanism validation (App. E). Across all 24 datasets we isolate the activity cliffs and show the corrections concentrate there — correcting 18.4% of the anchor's cliff errors while breaking only 4.9% (net +109, p<10⁻⁶) — with concrete molecule-level examples.
  • Anchor-sensitivity ablation (App. D.4). MolFiber is robust across fingerprint anchors (not Morgan-specific) and degrades gracefully with a geometry-poor one.
  • Random-neighborhood control (App. D.5). Replacing the Morgan fiber with a random one of the same size collapses the gain — it is the Morgan-confused neighborhood, not generic locality, that matters.
  • Plus a documented hyperparameter search, a scalability analysis, and a clarification of MolFormer's (label-free) pretraining.

Point-by-point responses follow below. Thank you again for the time and care.

Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.