In Parkinson's and Alzheimer's disease, proteins such as α-synuclein and amyloid-β clump together, and the real killers are not the large deposits but the small, soluble clusters called oligomers. Two facts make these clusters hard to fight. First, mature fibrils act as catalytic surfaces: monomers landing on a fibril's outer wall nucleate into fresh oligomers far faster than they would in free solution. This process, secondary nucleation, turns aggregation into an autocatalytic loop in which every fibril manufactures more seeds — it is the dominant source of toxic oligomers for α-synuclein under physiological conditions (α-Synuclein oligomers form by secondary nucleation, which used single-molecule microfluidics to watch fibrils act as oligomer factories). Second, the oligomers kill cells by inserting exposed hydrophobic patches into lipid membranes and perforating them; one can defuse this without touching the protein at all, by giving the membrane a more attractive binding partner that outcompetes and displaces the oligomers — the mechanism by which the aminosterol trodusquemine displaces 65–71% of membrane-bound oligomers across three unrelated proteins and restores cell viability (Trodusquemine displaces protein misfolded oligomers from cell membranes and abrogates their cytotoxicity through a generic mechanism, which showed by infrared spectroscopy and light scattering that displacement, not oligomer remodeling, is at work).
Machine-learning discovery campaigns exist for each half of this problem separately, and they share a quiet methodological defect that text-generation researchers diagnosed a decade ago: they score candidates with surface-level similarity and a single scalar number. This proposal argues that molecular scoring for amyloid therapeutics is failing in exactly the way BLEU and ROUGE fail for language evaluation, and that the fix discovered in natural language processing — scoring by asking calibrated questions, one per quality axis, instead of matching strings — transfers to this problem and unlocks a search that no existing single-objective pipeline can perform: compounds that act on both the fibril surface and the membrane at once.
Core Claim
The claim has a conceptual half and an operational half. Conceptually: structural similarity between molecules is the chemical analogue of lexical overlap between texts, with the same two failure modes. The paraphrase penalty — texts that say the same thing with different words score near zero — appears chemically as scaffold hopping, where fingerprint similarity search misses active compounds on novel scaffolds; Discovery of potent inhibitors of α-synuclein aggregation using structure-based iterative learning, an active-learning study that optimized secondary-nucleation inhibitors from under 200 starting samples, measured exactly this: similarity search achieved a 4% hit rate while its learned model reached 21.4% by iteration three. The keyword illusion — texts that share every keyword but invert the meaning score perfectly — appears chemically as the aminosterols: squalamine and trodusquemine are close structural relatives, yet Trodusquemine enhances Aβ42 aggregation but suppresses its toxicity by displacing oligomers from cell membranes showed that trodusquemine accelerates Aβ42 secondary nucleation while protecting cells, and the accompanying review (Squalamine and trodusquemine: two natural products for neurodegenerative diseases, from physical chemistry to the clinic) documents that the same compound class suppresses lipid-induced nucleation for α-synuclein. High overlap, opposite function — a 100% match with inverted semantics.
Operationally: we will replace single-scalar QSAR filters (quantitative structure–activity models that map a molecule to one potency number) with a multi-axis question-answering scorer, built on the design of Towards a Unified Multi-Dimensional Evaluator for Text Generation, which reframed text evaluation as Boolean questions answered by a language model whose probability of "yes" serves as a calibrated score. For each candidate molecule we will construct a structured "molecular document" (descriptors, predicted physicochemical profile, and a natural-language property description) and score it on separate axes — direction and magnitude of its effect on secondary nucleation, membrane-displacement efficacy, blood-brain-barrier penetrance, and cytotoxicity — never collapsing them prematurely into one number:
$$S(m) = \big(s_{\mathrm{kin}}(m),\; s_{\mathrm{disp}}(m),\; s_{\mathrm{BBB}}(m),\; s_{\mathrm{tox}}(m)\big)$$
where each $s_a(m)$ is the calibrated yes-probability for one question about molecule $m$. The one collapsed variant we will study deliberately, as the key ablation, is the weighted sum $S_{\mathrm{scalar}}(m) = \sum_a w_a s_a(m)$.
The payoff target is a compound class that no current pipeline can even rank correctly: dual-action molecules that inhibit α-synuclein secondary nucleation at substoichiometric ratios (the catalytic sites are vanishingly rare relative to monomer, so blockers work at inhibitor-to-protein ratios below 1:100, as established mathematically by Paolo Arosio et al. and exploited by Optimization of a small molecule inhibitor of secondary nucleation in α-synuclein aggregation) and displace pre-formed oligomers from membranes. Existing generative campaigns optimize only the kinetic half: Exploration and Exploitation Approaches Based on Generative Machine Learning to Identify Potent Small Molecule Inhibitors of α-Synuclein Secondary Nucleation discovered compound CLM.1 ($KIC_{50} = 0.42$ µM against secondary nucleation) but never measured membrane interactions, and its authors explicitly flagged the omission. Nobody has joined the two objectives — the closest existing work either couples the mechanisms by accident in a single natural product (Trodusquemine enhances Aβ42 aggregation...) or optimizes one axis with machine learning (Discovery of potent inhibitors...). We differ from both: we treat the tension between the mechanisms as an explicit scoring axis, because whether accelerating secondary nucleation is protective or harmful depends on whether displacement is simultaneously present — the Aβ42 trodusquemine result showed a 6-fold increase in aggregate mass coexisting with a 50% improvement in worm motility. A scalar filter cannot represent "accelerate only if displacing"; a sign-aware, protein-conditional question can.
Research Program
Four workstreams, in dependency order. First, a dual-annotation dataset. We start from the 453-molecule experimental set of secondary-nucleation inhibitors compiled by Exploration and Exploitation Approaches... and prioritize roughly 150–200 compounds spanning diverse scaffolds plus a positive-control aminosterol series. For each, we run the displacement assay battery established in Trodusquemine displaces protein misfolded oligomers...: confocal quantification of membrane-bound α-synuclein oligomers on SH-SY5Y neuroblastoma cells, MTT viability, and hydrophobic-dye controls confirming the oligomers are displaced intact rather than remodeled. We also re-derive each compound's microscopic rate constants ($k_n$, $k_+$, $k_2$ — primary nucleation, elongation, secondary nucleation) by Thioflavin-T kinetics combined with master-equation fitting, the deconvolution method used by Trodusquemine enhances Aβ42 aggregation... to isolate which microscopic step a compound touches. This yields the first dataset in which every molecule carries both a kinetic phenotype (with sign) and a displacement phenotype.
Second, the multi-axis scorer. We compare four representation-plus-objective combinations on identical data: SMILES string embeddings (a molecule written as a text string — the chemical equivalent of token matching, already shown by Exploration and Exploitation Approaches... to suffer severe false-negative rates here), hand-engineered descriptor QSAR (their rescue, reaching accuracy 0.98 and AUC 0.85 on the kinetic task), graph neural networks, and our Boolean-QA scorer, which follows Towards a Unified Multi-Dimensional Evaluator for Text Generation in training on auxiliary question-formulated tasks with synthetic negatives (scaffold-pair perturbations playing the role of its entity-swap perturbations) and reading the axis score from the yes-token probability. Where an axis lacks labels — displacement, before workstream one completes — the QA formulation lets the model answer from pretrained knowledge, the molecular analogue of UniEval's reference-free coherence scoring; we will also borrow probability-weighted scoring over graded answers from G-Eval: NLG Evaluation using Gpt-4 with Better Human Alignment, which improved human correlation by computing an expected score from token log-probabilities.
Third, a meta-evaluation benchmark — a SummEval for molecules. SummEval: Re-evaluating Summarization Evaluation benchmarked 14 automatic metrics against 12,800 human ratings and found that crowd annotations are unreliable (Krippendorff's alpha 0.49) relative to experts (0.71); we mirror this by having two independent kineticists assign per-axis ground truth on a held-out set, reporting agreement, and measuring each scorer's Spearman correlation with expert labels per axis. The NLP yardstick makes the target concrete: on SummEval, ROUGE-L correlates at only 0.128 while UniEval reaches 0.474 and G-Eval-4 reaches 0.514 (G-Eval); our hypothesis is that descriptor-QSAR and fingerprint similarity sit near the ROUGE-L end on the axes they were never designed to capture, and that QA scoring closes a comparable gap on the dual phenotype.
Fourth, the closed discovery loop. We plug the multi-axis scorer into the generative architecture of Exploration and Exploitation Approaches... — a chemical language model pretrained on ChEMBL24 and fine-tuned in the low-data regime, with reinforcement learning over the scorer as reward, drawing candidate structures from the ChemDiv CNS library and the MEGx natural product library so proposals remain purchasable or synthesizable, a limitation those authors explicitly worked to manage. The loop's Inform–Predict–Test cadence and per-iteration molecule budget are fixed in advance and shared by every competing method.
Baselines
The naive reference point is Tanimoto fingerprint similarity search: screen the commercial libraries for molecules most similar to known hits and test the top of that ranking. This is the baseline Discovery of potent inhibitors of α-synuclein aggregation using structure-based iterative learning actually ran, and it is the direct chemical embodiment of lexical overlap; that paper reports a 4% optimization hit rate for it, which is the number any learning-based method must clear to justify its existence. We will run it under the identical library and iteration budget.
The strongest competitor is the descriptor-QSAR-plus-CLM pipeline of Exploration and Exploitation Approaches..., which is the current best system on this problem family: its descriptor-based QSAR reached accuracy 0.98 and average AUC 0.85 with zero false positives on the 453-compound kinetic dataset, and its generative loop produced CLM.1 at $KIC_{50} = 0.42$ µM. That system is single-objective, so the fair multi-objective version of it — and the comparison we will actually run — is the same architecture with stacked scalar filters (a kinetic QSAR threshold, a displacement QSAR threshold, and the blood-brain-barrier classifier trained on B3DB, for which no reported result found on Lacuna for its accuracy), given exactly our dual-annotation data, the same CLM, the same compute, and the same tuning budget. Beating it means two things: a higher Spearman correlation with held-out expert labels on the displacement and sign-of-effect axes, and a higher dual-action hit rate — the fraction of experimentally tested compounds that both inhibit α-synuclein secondary nucleation ($KIC_{50} \le 1$ µM) and displace at least 50% of membrane-bound oligomers at no more than 10-fold molar excess without their own cytotoxicity.
The key ablation isolates the multi-axis ingredient: our own scorer with the axes collapsed into the single weighted sum $S_{\mathrm{scalar}}$, trained and tuned identically. If the collapsed version matches the multi-axis version on every metric, then the contribution of question-answering structure, as opposed to the representations and data behind it, is nil — and we will say so.
Why not just…?
Why not just stack more scalar QSAR filters? This is the most serious alternative, and we do not argue around it; we test it as the primary baseline above. The evidence against it is indirect but consistent: in text evaluation, Towards a Unified Multi-Dimensional Evaluator for Text Generation showed that collapsing multiple quality dimensions into unified scalar metrics loses 23% of achievable human correlation relative to per-axis question answering, and SummEval showed that single-number metrics correlate poorly with exactly the judgments that require weighing dimensions against each other. Chemically, scalar filters cannot express the conditional that motivates this proposal — "accelerating secondary nucleation is acceptable only when displacement is present," the combination that Trodusquemine enhances Aβ42 aggregation... proved protective in worms — because a weighted sum is blind to which axis drives the score. But if the scalar-collapse ablation matches our multi-axis scorer on the dual-action hit rate, the honest conclusion is that calibration and data, not question structure, did the work, and the proposal's central claim would be falsified in that form.
Why not just fix the representation — bigger graph networks or fine-tuned molecular foundation models? Representation genuinely matters here: Exploration and Exploitation Approaches... found that SMILES string embeddings broke down on this exact task while descriptors reached 0.98 accuracy, so a skeptic is right that backbone choice is a live variable. We therefore include graph networks and fine-tuned string models as representation baselines under identical data. But representation and evaluation objective are orthogonal axes, and the NLP evidence says the objective axis carries weight independently of backbone scale: UniEval won with a T5-large backbone, not a frontier model, by changing how evaluation was formulated. If a graph network with scalar heads beats our QA scorer on dual-phenotype prediction, we will report a representation result and a negative result on our claim — the experiment is powered to settle it either way.
Why not just run a medicinal-chemistry campaign on trodusquemine analogs? The aminosterol scaffold already displaces 65–71% of oligomers (Trodusquemine displaces protein misfolded oligomers...), so optimizing that series is the cheapest route to a displacement compound, and squalamine's clinical derivative ENT-01 has already reversed severe constipation in over 80% of Parkinson's patients in the RASMET trial (Squalamine and trodusquemine...). The catch is that the aminosterols' kinetic behavior flips across protein species and the molecular determinants of that flip are unknown — flagged as open in both the Aβ42 paper and the review — and squalamine itself barely crosses the blood-brain barrier. A single-scaffold series cannot map that landscape; our scorer can, and to keep the comparison concrete we include an aminosterol analog subset as positive controls and require the framework to reproduce their known qualitative behavior before trusting its rankings on new scaffolds.
Decisive Checks
Three outcomes decide the proposal. First, the meta-evaluation: our scorer must beat the descriptor-QSAR stack on Spearman correlation with expert labels for the displacement axis and the sign of the kinetic effect, with the gap reported per axis exactly as SummEval reported per dimension. Second, sign retrodiction: with all aminosterols held out of training, the scorer must correctly recover that trodusquemine accelerates Aβ42 secondary nucleation while the related aminosterols suppress α-synuclein lipid-induced nucleation — a leave-one-family-out test that a similarity-based method fails by construction, since these molecules are near-neighbors with opposite labels. Third, the closed loop: across a fixed testing budget, the multi-axis reward must produce a dual-action hit rate exceeding both the 4% similarity-search floor and the stacked-scalar CLM baseline, with at least one validated compound achieving $KIC_{50} \le 0.42$ µM (matching CLM.1) and ≥50% oligomer displacement without intrinsic cytotoxicity, confirmed by ThT kinetics with master-equation deconvolution, confocal displacement imaging, and MTT assays. Failing the first two checks would still yield a usable dual-annotation dataset and a clear negative result on the transfer of question-answering evaluation across modalities; failing only the third would locate the limit at the generator rather than the scorer, which is itself a finding the field currently has no way to see.
References
- α-Synuclein oligomers form by secondary nucleation
- Trodusquemine displaces protein misfolded oligomers from cell membranes and abrogates their cytotoxicity through a generic mechanism
- Discovery of potent inhibitors of α-synuclein aggregation using structure-based iterative learning
- Trodusquemine enhances Aβ42 aggregation but suppresses its toxicity by displacing oligomers from cell membranes
- Squalamine and trodusquemine: two natural products for neurodegenerative diseases, from physical chemistry to the clinic
- Towards a Unified Multi-Dimensional Evaluator for Text Generation
- Kinetic analysis reveals the diversity of microscopic mechanisms through which molecular chaperones suppress amyloid formation
- Optimization of a small molecule inhibitor of secondary nucleation in α-synuclein aggregation
- Exploration and Exploitation Approaches Based on Generative Machine Learning to Identify Potent Small Molecule Inhibitors of α-Synuclein Secondary Nucleation
- G-Eval: NLG Evaluation using Gpt-4 with Better Human Alignment
- SummEval: Re-evaluating Summarization Evaluation