WEECFP-SuRGE: Wide Embedded Extended Connectivity Fingerprint with Substructure Rotary Graph-distance Encoding
cs.LG
Quick verdict
Claims top non-pretrained TDC ADMET rank via WEECFP fingerprint + graph-distance RoPE; results credible, but SuRGE's marginal gain is small and untested.
Summary
The paper introduces WEECFP, a deterministic, parameter-free 1024-dimensional continuous molecular fingerprint that scatters each Morgan circular substructure across ~32 signed positions (position and sign both derived from MD5 hash bytes), plus SuRGE, a RoPE-style rotary attention encoding whose phase is a function of molecular shortest-path graph distance rather than sequence position, applied to WEECFP substructure tokens in a 3-layer transformer trained from scratch. A fixed 7-model blend (4 SuRGE transformers, an MLP, 2 CNNs, with Morgan and Avalon/ErG-augmented members) achieves the lowest average regression rank (3.9/9) on the TDC ADMET leaderboard, ranks second overall (5.1/22) behind only pretrained MapLight+GNN, and takes five #1 finishes (Pgp, Lipophilicity, CYP2D6 Substrate, microsomal clearance, LD50). On MoleculeNet it beats the best of 12 classical fingerprints x XGB/MLP on 3 of 4 regression tasks. The authors show tokenization is near-lossless (99.9% exact SMILES reconstruction in-distribution; 98.93% cross-dataset, where a Morgan-bit control ties) and that a three-reference distance encoding correlates r=0.901 with true pairwise distances at O(S) memory. Main limitations: no significance testing for small deltas (0.013 RMSE on Lipophilicity); the NoSuRGE blend ablation nearly matches (5.3 vs 5.1) and scores better on LD50/HIA; per-task variant selection appears test-driven; no code release is stated.
Review complete
Assessment at a glance
4
risk flags identified
● high 0
● medium 3
● low 1
4 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- A 7-model WEECFP-SuRGE blend achieves the lowest average regression rank (3.9/9) on the TDC ADMET leaderboard and is #1 overall among methods without external pretraining (5.1/22, behind only pretrained MapLight+GNN).
- SuRGE, a RoPE-like rotation parameterized by shortest-path graph distance, improves regression when applied inside every attention layer, with small gains over no-SuRGE controls (0.013 RMSE on Lipophilicity, 0.030 on QM9).
- WEECFP tokenization is near-lossless: greedy overlap-merge reconstruction recovers exact canonical SMILES for 99.9% of in-distribution molecules across 9 MoleculeNet datasets and 98.93% of a cross-dataset holdout (HIV to Lipophilicity).
- A three-reference farthest-first distance encoding preserves Pearson r=0.901 correlation with true pairwise graph distances while reducing positional-state memory from O(S^2) to O(S).
Strengths
Fully deterministic, exactly specified fingerprint
Eq. 1 gives the complete MD5 byte-to-(position, sign) mapping (~32 signed writes per substructure via two digests, second digest stride 31 on the low nibble), plus atom-property and neighbor blocks (§3.1); ~1 ms/molecule with no learned components — directly re-implementable.
Unusually disciplined and honest ablations
A full WEECFP-NoSuRGE Blend (Tables 2-3), Block-1-only WEECFP-Scatter controls (SI §C.3-C.4), a matched Morgan-bit reconstruction control (Table S13), and the leaky oracle explicitly quarantined as a diagnostic (SI §E.3); failures on FreeSolv, Bioavailability (rank 21) and CYP2C9-I are disclosed in §5.
Broad evaluation at official protocol scale
Sound O(S) positional-encoding analysis
Weaknesses
No statistical testing anywhere
Weak attribution of gains to SuRGE
Per-task variant selection on test performance
Blend composition muddies the method's contribution
No code or leaderboard snapshot released
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
The paper introduces two artifacts and one ensemble. WEECFP (§3.1) is a deterministic, parameter-free 1024-d continuous fingerprint: each Morgan circular substructure (RDKit, radius 2) is written to ~32 signed positions via two MD5 digests, with the same hash byte choosing position and sign (Eq. 1), plus atomic-property and neighbor-bond blocks; per-substructure vectors are log-compressed and summed (Eq. 2). SuRGE (§3.2) is a RoPE-style rotation whose phase is a function of shortest-path graph distance rather than sequence position, applied to per-substructure tokens in a 3-layer transformer (d_model 256, 8 heads, 300-token cap). The 7-model Blend (§3.3, Table 1) combines 4 SuRGE transformers (WEECFP or Morgan tokens, optionally with Avalon/ErG), an MLP and two CNNs. Headline claims: lowest average regression rank (3.9/9) on TDC ADMET, #2 overall (5.1/22) behind pretrained MapLight+GNN, #1 among no-pretraining methods, five leaderboard #1s, wins over all classical-fingerprint baselines on 3/4 MoleculeNet regression tasks, near-lossless tokenization, and an O(S) three-reference distance encoding at r=0.901.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 4 strengths & 5 weaknesses, analyzed in detail
- 4 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is WEECFP and how does it differ from ECFP?
WEECFP is a continuous, parameter-free molecular fingerprint. Instead of flipping one bit per Morgan circular substructure (as ECFP does), it writes each substructure into about 32 signed positions of a 1024-dimensional vector, with both the position and the sign derived from MD5 hash bytes. Collisions therefore tend to cancel rather than accumulate, and vector magnitudes encode substructure counts. A summed version (WEECFP-S) works as a drop-in fingerprint for MLP and XGBoost models.
What is SuRGE in this paper?
SuRGE (Substructure Rotary Graph-distance Encoding) adapts rotary position embeddings (RoPE) to molecules: instead of rotating token embeddings by their sequence position, the rotation angle is a function of shortest-path distance on the molecular graph — either each token's mean distance to all others, or its distances to three farthest-first reference atoms, which keeps memory linear in sequence length.
How well does WEECFP-SuRGE perform on the TDC ADMET leaderboard?
According to the paper, its 7-model blend has the best average rank across the 9 regression benchmarks (3.9) and ranks second overall across all 22 benchmarks (5.1), behind only the pretrained MapLight+GNN entry. Among methods that use no external pretraining it is first overall, with individual #1 results on Pgp, Lipophilicity, CYP2D6 Substrate, microsomal clearance, and LD50. These are leaderboard-position facts reported by the author, not statistically tested superiority.
Is the WEECFP fingerprint reversible — can it leak the original molecule?
The substructure tokens are nearly lossless in practice: a greedy overlap-merge reconstruction recovered exact canonical SMILES for 99.9% of in-distribution MoleculeNet molecules and 98.93% in a cross-dataset holdout. However, a matched 2048-d Morgan-bit control achieved exactly the same rate, so reversibility comes from substructure tokenization rather than the continuous signed embedding — relevant for audit scenarios where fingerprints must not reveal structures.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- The abstract in the metadata ends mid-sentence ('...enabling O(S) position'), but the embedded full paper text is present through SI §G.3 (p. 23); every number cited in this review comes from the complete text.
- Figures (Fig. 1, 2, S1-S4) appear only as captions in the extraction; visual/plot content could not be inspected.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.