Natural products that kill cancer cells by triggering two death programs at once — say, blocking the PI3K/Akt survival signal while simultaneously forcing the mitochondrial apoptotic cascade — often owe their potency to a single, precisely oriented stereocenter. The clearest documented case in the file: among seven epipolythiodioxopiperazine fungal metabolites sharing one scaffold, only prenylemestrin A, whose activity depends on its 3D bridge and hemiterpene arrangement, is cytotoxic against L1210 leukemia cells ($IC_{50} = 14.7\ \mu$M) through a dual PI3K/Akt-plus-mitochondrial mechanism, while its stereochemically perturbed congeners are nearly inert (Emestrin-type epipolythiodioxopiperazines from Aspergillus nidulans with cytotoxic activities by regulating PI3K/AKT and mitochondrial apoptotic pathways). Yet the machine learning tools we use to find such molecules under label scarcity are, as a matter of arithmetic, blind to this: standard graph networks sum or average over neighbor atoms, so an active molecule and its inactive mirror image receive identical predicted activity (Stereochemical Awareness in Molecular Graph Neural Networks).
The proposal here is to close that gap with one system: a few-shot meta-learner — a model that learns how to learn from a handful of labeled examples per new assay — whose molecular encoder is built to distinguish stereoisomers, evaluated on a new benchmark of stereoisomer-paired cell-death assays. The encoder reuses two existing techniques: ChIRo's torsion-angle encoder, which represents a molecule with periodic waves in its dihedral angles so that it is stable under bond rotations yet still separates enantiomers (Learning 3D Representations of Molecular Chirality with Invariance to Bond Rotations), and ChiDeK's signed-volume determinant features, which extend explicit chirality encoding to axial stereocenters common in natural products (Learning Molecular Chirality via Chiral Determinant Kernels). Neither has ever been plugged into a few-shot biological prediction pipeline; both have only been tested on chiroptical or R/S-labeling exercises. Conversely, the strongest few-shot drug models — UniMatch's hierarchical matching over 2D graphs (UniMatch: Universal Matching from Atom to Task for Few-Shot Drug Discovery) and MHNfs's Hopfield-network associative memory (Context-enriched molecule representations improve few-shot drug discovery) — cannot in principle separate an active stereoisomer from an inactive one. Dual-pathway apoptosis induction is the application regime where this blindness matters most, because the phenotype is exquisitely stereochemistry-dependent and the labeled data (a few dozen compounds per assay, per cell line) is exactly the regime where few-shot methods operate.
Core Claim
The central claim is empirical and falsifiable: a prototypical few-shot learner whose embedding fuses a chirality-discriminative 3D encoder with a standard 2D graph network will (a) match or beat the best 2D few-shot models on conventional activity benchmarks, and (b) discriminate active from inactive stereoisomers in paired cell-death assays — a task on which every 2D model is stuck at chance, since it outputs the same prediction for both members of a pair. The stereochemical axis is the core mechanism; dual-pathway cell death is the evaluation regime; label scarcity is the constraint that dictates the meta-learning design rather than a brute-force supervised approach.
Two design choices follow directly from documented failure modes. First, ChiDeK's authors report that predicting chiroptical spectra from quickly generated conformers fails when those conformers drift from density-functional-theory-optimized geometries (Learning Molecular Chirality via Chiral Determinant Kernels); ChIRo's rotation-invariant encoder was built precisely to survive such conformer noise — trained on a single conformer per molecule it kept roughly 79% accuracy on optical-rotation sign while the SE(3)-equivariant SphereNet baseline fell from 65% to 55% (Learning 3D Representations of Molecular Chirality with Invariance to Bond Rotations). So the encoder uses rotation-invariant torsion features as the backbone, with determinant-based chiral features for axial centers, computed on cheap cheminformatics conformers. Second, the meta-learning head is prototypical rather than memory- or MAML-based, following the evidence that prototypical networks predict stereoselectivity in asymmetric catalysis from as few as 32–64 support examples, reaching an AUPRC of 0.9341 ± 0.0053 on out-of-sample selectivity tasks and beating random forests trained on the full dataset (A meta-learning approach for selectivity prediction in asymmetric catalysis) — direct precedent that meta-learning transfers across stereochemical selectivity tasks, just not yet into pharmacology. The prototype head also avoids MHNfs's roughly 15,000 A100 GPU hours (Context-enriched molecule representations improve few-shot drug discovery), which matters for a method intended to be retrained by ordinary labs.
Because generic activity prediction already leans heavily on topology, the architecture fuses the chiral encoder with a standard 2D graph network, and adds one auxiliary loss that directly targets the failure mode: given a molecule $M$ and its mirror image $\bar{M}$, push their embeddings $h(M)$ and $h(\bar{M})$ apart by at least a margin $m$ whenever their measured activities differ,
$$\mathcal{L} = \mathcal{L}_{\text{proto}} + \lambda \, \max\!\bigl(0,\; m - \|h(M) - h(\bar{M})\|\bigr),$$
where $\mathcal{L}_{\text{proto}}$ is the usual prototype-classification loss over support sets, and $\lambda$ balances the two terms. On molecules without stereocenters the second term is simply zero, so the model degrades gracefully to a standard few-shot learner.
Research Program
Three deliverables. The first is a benchmark, provisionally called StereoFS-Mol, assembled by filtering few-shot assay tasks for compounds with annotated stereocenters and pairing stereoisomers that share a scaffold but differ in measured outcome. The source pool is FS-Mol's 5,120 ChEMBL27 assays (4938 training, 40 validation, 157 test tasks, mean 94 compounds per task) (FS-Mol: A Few-Shot Learning Dataset of Molecules), supplemented by a literature-mined held-out set of natural-product stereoisomer pairs with quantified cell-death readouts — the epipolythiodioxopiperazine family with its 14.7 µM activity switch (Emestrin-type epipolythiodioxopiperazines from Aspergillus nidulans with cytotoxic activities by regulating PI3K/AKT and mitochondrial apoptotic pathways), erianin-class bibenzyls whose apoptosis-plus-pyroptosis phenotype is stereochemically defined ($IC_{50} = 76.27$ nM in CAL62 cells, combination index consistently below 0.9 with anlotinib) (Erianin induces apoptosis and pyroptosis via MAPK/ERK and PI3K/Akt pathways and synergizes with anlotinib in anaplastic thyroid carcinoma), and marine-fungal alkaloids and heterodimers whose absolute configurations were established by ECD calculation (Versiquinazolines L–Q, new polycyclic alkaloids from the marine-derived fungus Aspergillus versicolor; Chalasoergodimers A–E, heterodimers with multiple polymerization modes from a marine-derived Chaetomium sp. fungus). Each pair carries, where reported, the pathway annotation (intrinsic/extrinsic apoptosis, pyroptosis, ferroptosis, autophagy blockade), so the benchmark can score pathway-selective prediction, not just "toxic or not."
The second deliverable is the model itself: the fused chirality-aware prototypical network described above, trained on the FS-Mol task distribution with the enantiomer-margin loss applied to any stereoisomer pairs found within the training tasks. The third is an analysis: which stereocenter types (tetrahedral, axial, bridgehead) drive the model's discrimination, and whether the model's attention to chiral features correlates with the assay's target class — testing the hypothesis that assays against kinases in survival cascades (PI3K/Akt, MAPK/ERK) are more stereochemistry-sensitive than, say, broad cytotoxicity counterscreens.
Proposed Experiments
Experiment 1 is the conventional few-shot comparison on FS-Mol and MoleculeNet at support size 16, following the protocols of UniMatch and MHNfs, with the same task splits, the same tuning budget (random search over the same hyperparameter ranges for every method), and identical compute for training and inference. Success means staying within or beating the published gaps: UniMatch reports a 6.52% improvement in $\Delta$AUPRC over its baselines and an average 2.87% AUROC improvement on MoleculeNet (UniMatch: Universal Matching from Atom to Task for Few-Shot Drug Discovery); MHNfs reports $\Delta$AUC-PR of 0.241 at support size 16 versus 0.234 for ADKF-IFT and 0.223 for Siamese networks (Context-enriched molecule representations improve few-shot drug discovery). For Tox21 domain shift, MHNfs reports no number on Lacuna, so that comparison will be run from scratch. This experiment guards against the known trap where added architectural complexity costs accuracy on ordinary tasks.
Experiment 2 is the decisive one: stereoisomer-pair discrimination on the held-out natural-product set. The metric is pairwise ranking accuracy — the fraction of stereoisomer pairs for which the model assigns higher predicted dual-pathway activity to the experimentally active member. Any 2D model necessarily scores 50% here (it produces tied predictions), which is the built-in null. A secondary metric is calibration of the margin between pair members against the measured potency gap, using the reported $IC_{50}$ values (14.7 µM for prenylemestrin A against effectively inactive congeners) as anchors. The image-based CellDeathPred system, which classifies ferroptotic, apoptotic, and healthy cells from microscopy at 95% accuracy on non-confocal data (CellDeathPred: a deep learning framework for ferroptosis and apoptosis prediction based on cell painting), is deliberately not a baseline — it predicts from cell images, not molecular structure, so it cannot rank an untested compound — but it sets the ceiling for what a phenotypic readout can resolve, and its modality labels define the pathway classes we predict.
Experiment 3 is the frequency-controlled ablation. Because MHNfs's authors showed that a naive "frequent hitters" baseline — predicting activity by how often a molecule is active across training assays — outperforms several published meta-learners (Context-enriched molecule representations improve few-shot drug discovery), every few-shot comparison is repeated with the frequency signal regressed out of the evaluation tasks, and the frequency baseline is reported alongside. Beating the frequency baseline on stereo-poor tasks is not claimed; the claim is specifically that on stereochemistry-bearing tasks the chiral encoder extracts signal no frequency count or fingerprint can see.
Baselines
The do-nothing reference point is the frequent-hitters predictor: for each query molecule, output the fraction of training-assay tasks in which that molecule (by canonical 2D structure) was labeled active. It requires no learning, and the MHNfs paper's finding that it beats several meta-learning algorithms (Context-enriched molecule representations improve few-shot drug discovery) makes it the credibility test any new few-shot architecture must pass. On stereoisomer pairs it is a tie by construction, so its pair-ranking accuracy is exactly 50%. We run it with identical task splits and report it everywhere.
The strongest existing competitor on the few-shot axis is UniMatch, which combines hierarchical attention matching across graph-network layers with MAML-style task adaptation and reports the best numbers on the benchmarks we use: a 6.52% $\Delta$AUPRC improvement on FS-Mol and 2.87% average AUROC improvement on MoleculeNet (UniMatch: Universal Matching from Atom to Task for Few-Shot Drug Discovery). We will re-implement or obtain its code, tune it on the validation tasks with the same search budget and compute as our model, and run it on both the conventional benchmarks and StereoFS-Mol. On the stereoisomer-pair metric it is expected to sit at 50% — that prediction is itself a falsifiable check of the blindness claim, and if UniMatch somehow exceeds 50% the premise of this proposal weakens and we will report that honestly. MHNfs serves as a second reference point at $\Delta$AUC-PR 0.241 at support size 16 (Context-enriched molecule representations improve few-shot drug discovery); given its 15,000-GPU-hour training cost we will use its published numbers on shared tasks rather than retrain it, and flag any protocol mismatches.
On the representation axis, the strongest competitor for chirality discrimination is ChiDeK, whose determinant-kernel architecture reached 99.8% on R/S classification, 69.2% on axial optical-rotation prediction where baseline models hovered near 50%, and beat the next best model by 10.8% on ECD peak-sign prediction (Learning Molecular Chirality via Chiral Determinant Kernels). ChiDeK has never been applied to biological assays or few-shot settings, so we run it as a frozen feature extractor feeding the same prototypical head — a like-for-like test of whether its chiral features, rather than ChIRo's torsion features, are the better substrate for pharmacological few-shot learning. ChIRo's own encoder is run the same way; its published optical-rotation sign accuracy of 79.3% versus 74.4% for 2D models and 65.5% for SphereNet (Learning 3D Representations of Molecular Chirality with Invariance to Bond Rotations) sets expectations for what geometric chirality signal adds over topology.
The key ablation isolates the proposal's single ingredient: the identical prototypical meta-learner, identical training schedule, identical tuning budget, with the chiral encoder replaced by a plain graph-isomorphism network on 2D graphs. Any gap between this ablation and the full model on Experiment 2, with the two tied on Experiment 1, is the evidence that stereochemical encoding — not meta-learning, not extra parameters — is doing the work. A second ablation removes only the enantiomer-margin loss from the full model, to test whether explicit pair supervision is necessary or whether chirality-aware features alone suffice.
Why not just…?
Why not just use 2D fingerprints with random forests or nearest neighbors? This is the most serious objection, and the file gives it real ammunition: the original FS-Mol study found few-shot performance highly task-dependent, with well-tuned single-task random forests not always beaten (FS-Mol: A Few-Shot Learning Dataset of Molecules), and the frequent-hitters result shows many meta-learners add nothing over promiscuity counting (Context-enriched molecule representations improve few-shot drug discovery). We concede the stereochemistry-free regime entirely: on tasks whose molecules lack stereocenters, classical baselines may well win, and we will report that. The proposal's claim lives precisely where classical methods cannot follow: a Morgan fingerprint is identical for both members of a stereoisomer pair, so its pair-ranking accuracy is mathematically fixed at 50%, while the epipolythiodioxopiperazine data shows the true signal is a 14.7 µM-versus-inactive separation (Emestrin-type epipolythiodioxopiperazines from Aspergillus nidulans with cytotoxic activities by regulating PI3K/AKT and mitochondrial apoptotic pathways). The catalysis meta-learning result — prototypical networks with 64 support examples beating random forests trained on the entire dataset on stereoselectivity (A meta-learning approach for selectivity prediction in asymmetric catalysis) — is the strongest existing evidence that this concession costs little where it matters. Experiment 3 settles the residual risk by running every comparison with the frequency signal controlled.
Why not just run TDDFT simulations or add chiral tags to the SMILES string? TDDFT with ECD spectral matching is the accepted gold standard for assigning absolute configuration (Absolute Configuration and Biological Evaluation of Novel Triterpenes as Possible Anti-Inflammatory or Anti-Tumor Agents; Norsampsone E, an unprecedented decarbonyl polycyclic polyprenylated acylphloroglucinol with a homoadamantyl core from Hypericum sampsonii), and we do not propose to replace it for assignment — only for prediction of biological activity, which TDDFT does not address at all: it computes spectra, not caspase cleavage, and the $2^n$ stereoisomer explosion makes per-isomer quantum runs infeasible for screening libraries. Textual chiral tags are cheaper but demonstrably weaker: on optical-rotation sign, 2D models with such tags reached only 74.4% versus ChIRo's 79.3%, and on axial-chirality tasks tag-based models sat at chance (~50%) where geometric kernels reached 69.2% (Learning 3D Representations of Molecular Chirality with Invariance to Bond Rotations; Learning Molecular Chirality via Chiral Determinant Kernels). A tag-augmented 2D baseline will still be included in the ablation suite, so if tags turn out to suffice for pharmacology the experiments will show it.
Why not just co-administer two known drugs instead of finding one stereochemically precise dual-pathway molecule? Combinations genuinely work in the file: erianin plus anlotinib yields combination indices below 0.9 in anaplastic thyroid carcinoma (Erianin induces apoptosis and pyroptosis via MAPK/ERK and PI3K/Akt pathways and synergizes with anlotinib in anaplastic thyroid carcinoma). But this objection targets a different output than ours. The proposal is a predictive model for which stereochemical arrangements of a single compound engage two death pathways, not a design doctrine that single entities are clinically superior. Even for combination therapy, the model answers a question combinations cannot: whether a given stereoisomer of a new natural product is the active partner — prenylemestrin A's dual PI3K/Akt-plus-mitochondrial mechanism lives in one scaffold whose stereochemical perturbation destroys it (Emestrin-type epipolythiodioxopiperazines from Aspergillus nidulans with cytotoxic activities by regulating PI3K/AKT and mitochondrial apoptotic pathways), and CombiMOTS's demonstration that single dual-target entities avoid mismatched pharmacokinetics while sustaining polypharmacology (CombiMOTS: Combinatorial Multi-Objective Tree Search for Dual-Target Molecule Generation) explains why anyone wants that question answered. A fair related point: CombiMOTS itself already does dual-target generation, but purely in 2D over fragment space — the stereochemical axis it ignores is exactly the one our benchmark measures, and the two systems would ultimately be complementary (its search, our stereochemistry-aware scoring) rather than competing.
Decisive Checks
The proposal stands or falls on three numbers. First, pairwise ranking accuracy on held-out natural-product stereoisomer pairs: the full model must beat 50% by a margin outside the bootstrap interval, while the 2D ablation sits at 50%. Second, FS-Mol $\Delta$AUPRC within the range set by UniMatch's reported 6.52% improvement (UniMatch: Universal Matching from Atom to Task for Few-Shot Drug Discovery) and MHNfs's 0.241 $\Delta$AUC-PR (Context-enriched molecule representations improve few-shot drug discovery) — chirality awareness must not cost ordinary competence. Third, the frequency-controlled rerun must show the stereoisomer-pair advantage survives removal of the promiscuity signal. If the first number comes back at chance, the honest conclusion would be that current chirality encoders, tuned on chiroptical data, do not transfer to pharmacological assays — itself a finding the field currently has no benchmark to discover.
References
- Emestrin-type epipolythiodioxopiperazines from Aspergillus nidulans with cytotoxic activities by regulating PI3K/AKT and mitochondrial apoptotic pathways
- Learning 3D Representations of Molecular Chirality with Invariance to Bond Rotations
- Learning Molecular Chirality via Chiral Determinant Kernels
- UniMatch: Universal Matching from Atom to Task for Few-Shot Drug Discovery
- Context-enriched molecule representations improve few-shot drug discovery
- A meta-learning approach for selectivity prediction in asymmetric catalysis
- FS-Mol: A Few-Shot Learning Dataset of Molecules
- Erianin induces apoptosis and pyroptosis via MAPK/ERK and PI3K/Akt pathways and synergizes with anlotinib in anaplastic thyroid carcinoma
- Versiquinazolines L–Q, new polycyclic alkaloids from the marine-derived fungus Aspergillus versicolor
- Chalasoergodimers A–E, heterodimers with multiple polymerization modes from a marine-derived Chaetomium sp. fungus
- CellDeathPred: a deep learning framework for ferroptosis and apoptosis prediction based on cell painting
- Absolute Configuration and Biological Evaluation of Novel Triterpenes as Possible Anti-Inflammatory or Anti-Tumor Agents
- Norsampsone E, an unprecedented decarbonyl polycyclic polyprenylated acylphloroglucinol with a homoadamantyl core from Hypericum sampsonii
- CombiMOTS: Combinatorial Multi-Objective Tree Search for Dual-Target Molecule Generation