When a patient's cancer stops responding to a drug, the explanation often lives in two places at once: a rare subpopulation of cells that bulk sequencing averages away, and a mutation whose effect on the drug target is not a direct clash in the binding pocket but a subtle structural ripple that travels across the protein from a distant residue. The first problem belongs to the literature on cellular heterogeneity masking — how aggregate measurements from millions of cells hide low-frequency subclones and their treatment-resistant lineages. The second belongs to the literature on allosteric drug resistance — how out-of-pocket substitutions weaken drug binding through changed residue networks, conformational floppiness, or destabilized cofactors, without touching the pocket itself. These two fields have developed in complete isolation from each other, and the result is a blind spot with clinical consequences: a model can now reconstruct a tumor's subclonal architecture, and a simulation can now explain a single distal mutation's effect on binding, but nothing connects them.
The closest existing work makes the gap precise. SubNetDL A network-based deep learning model integrating subclonal architecture for therapy response prediction in cancer infers one to six subclones per tumor from bulk variant allele frequencies, projects mutations onto a protein-interaction network, and predicts therapy response better than aggregate biomarkers — but it encodes every mutation as a binary bit, so an in-pocket steric clash (ALK L1198F), a distal allosteric switch propagated along a ten-residue transmission chain (ALK I1171S), an entropy-driven loop destabilizer (JAK2 L884P), and a silent passenger all enter the model as the same value Deciphering the Shared and Specific Drug Resistance Mechanisms of Anaplastic Lymphoma Kinase via Binding Free Energy Computation; How Does the L884P Mutation Confer Resistance to Type-II Inhibitors of JAK2 Kinase: A Comprehensive Molecular Modeling Study. Conversely, the structural side has rich mechanistic detail but no notion of clonal context: MdrDB MdrDB: Mutation-induced drug resistance DataBase holds 100,537 mutation–drug–structure samples with free-energy labels, yet those labels come from bulk cell-line assays that average out subclonal diversity, and its features come from rigid docking that cannot see allosteric dynamics.
Core Claim
The claim of this proposal is that subclonal response prediction fails on exactly the patients it most needs to help — those whose resistance is driven by low-frequency subclones carrying distal, structurally subtle mutations — because the predictive signal is destroyed twice: once by bulk averaging (the heterogeneity mask) and once by binary mutation encoding (the biophysical mask). Replacing binary node labels with a drug-specific, structurally grounded allosteric resistance score, and pooling subclonal representations with clonal-frequency weighting, should recover that signal. Concretely, for each mutation $m$ in the target of drug $d$ we compute a score $a(m, d) \in [0, 1]$ that estimates how much the mutation weakens binding of that specific drug, combining intra-protein transmission centrality (how directly the mutated residue communicates with pocket residues through the residue contact network) with a learned binding free-energy shift. The model then predicts response through a clonal-frequency-weighted resistant fraction,
$$ R = \sum_{c} f_c \, \sigma\!\Big(\max_{m \in c} a(m, d)\Big), $$
where $f_c$ is the inferred frequency of subclone $c$, the maximum runs over mutations present in that subclone's target genes, and $\sigma$ is a logistic squashing function. The intuition is simple: a tumor should be classified as non-responding in proportion to how much of its cellular mass carries mutations that structurally compromise the drug's binding — not in proportion to how many mutated genes it happens to contain.
Research Program
Three components are built, in increasing order of integration. First, the allosteric score $a(m, d)$. For each target protein we construct an intra-protein residue contact network from experimental structures or AlphaFold3 models, as in the transthyretin landscape study Integrative structural profiling and ligand optimisation across the transthyretin mutational landscape, and compute for each mutable residue its transmission centrality with respect to the drug pocket — a PageRank-style propagation along contact paths, deliberately analogous to the multiscale propagation SubNetDL performs on the inter-protein network, but moved inside the protein where allosteric resistance actually operates. The centrality is combined with a predicted binding free-energy shift trained on MdrDB's 100,537 samples, using its 146 biochemical features as inputs. We acknowledge upfront that MdrDB's labels are bulk-derived and its docking is static; we therefore treat MdrDB as pretraining only, and calibrate the surrogate against the small set of systems where dynamics have been quantified by simulation: the JAK2 L884P system, where umbrella sampling showed the BBT594 binding well depth dropping from roughly 19.8 to 16.7 kcal/mol alongside an 11-fold IC50 increase, while CHZ868 suffered only a 4-fold shift How Does the L884P Mutation Confer Resistance to Type-II Inhibitors of JAK2 Kinase: A Comprehensive Molecular Modeling Study, and the ALK panel, where contact-network analysis traced the I1171S transmission chain residue by residue Deciphering the Shared and Specific Drug Resistance Mechanisms of Anaplastic Lymphoma Kinase via Binding Free Energy Computation. The surrogate must reproduce the relative drug sensitivity ordering in these systems — BBT594 more compromised than CHZ868 — to be considered validated, since this is precisely the drug-specificity that generic pathogenicity scores lack.
Second, the integrated model. We re-implement the SubNetDL architecture — SciClone subclone inference, STRING network propagation, graph attention over subclone-specific representations — and change exactly one thing per variant: the binary mutation indicator on each node becomes the allosteric score $a(m, d)$, and subclone pooling uses $f_c$ weights as above. Everything else (network, propagation scales, attention, tuning budget) is held identical so that any performance difference is attributable to the biophysical weighting alone. Third, the masking stress test, described under decisive experiments, which probes whether the frequency-weighted formulation degrades gracefully as resistant subclones are diluted below the roughly 5% variant-allele-frequency floor where bulk inference is known to struggle A network-based deep learning model integrating subclonal architecture for therapy response prediction in cancer.
Baselines
The do-nothing reference point is tumor mutational burden, the standard clinical aggregate biomarker: it is already computed from the same bulk data and requires no modeling. It is also already known to be beaten on these exact cohorts — SubNetDL significantly outperformed TMB on the MSK_2020 and MSK_2018 immunotherapy cohorts ($P = 0.02$ and $P = 6.0 \times 10^{-3}$ respectively) A network-based deep learning model integrating subclonal architecture for therapy response prediction in cancer — so the real question is whether anything beats the subclonal models, not the aggregate ones.
The strongest existing competitor is SubNetDL itself A network-based deep learning model integrating subclonal architecture for therapy response prediction in cancer, which reports a median AUROC of 0.74 across 10 TCGA cancer–drug combinations, 0.77 across the external MSK immunotherapy cohorts, and 0.81 across six CCLE/GDSC cancer–drug pairs. It is the strongest because it is the only published method that both reconstructs subclonal architecture and achieves clinically validated response prediction with it. We will run it on identical data, splits, and tuning budget, and count success as raising median AUROC above these reported values on the same benchmarks, with the decisive comparison restricted to the patient subset harboring out-of-pocket mutations in the drug target — where the binary encoding is most lossy and where our method should concentrate its gains. If gains appear only on in-pocket mutations, the allosteric component has failed its purpose.
Two ablations isolate the proposal's key ingredient within our own architecture. One swaps the allosteric score for static MdrDB-style features — docking scores and solvent-accessibility descriptors of the kind MdrDB computes MdrDB: Mutation-induced drug resistance DataBase — to test whether expensive transmission-centrality computation adds anything over cheap static annotation. The other swaps it for a protein-language-model pathogenicity score, since ESM2 embeddings achieved a ROC AUC of 0.9948 separating pathogenic from benign transthyretin variants, far above AlphaMissense at 0.6771 Integrative structural profiling and ligand optimisation across the transthyretin mutational landscape. If either ablation matches the full method, the structural machinery is unjustified and we will say so.
Why not just…?
Why not just use TMB or other bulk biomarkers? Because the subclonal literature has already settled this on the exact cohorts we will use: SubNetDL beat TMB with significance on both MSK cohorts and showed that perturbing the subclonal assignment of mutations — while keeping every mutation present — degrades prediction A network-based deep learning model integrating subclonal architecture for therapy response prediction in cancer. The signal we are after lives in the joint structure of which mutations sit in which subclones, which no aggregate can represent.
Why not just annotate every mutation with a language-model or pathogenicity score? This is the most serious alternative, because ESM2's 0.9948 ROC AUC on TTR variant classification Integrative structural profiling and ligand optimisation across the transthyretin mutational landscape is cheap to obtain and requires no structures. But pathogenicity is drug-agnostic: it says a variant is harmful to the protein, not that it weakens this drug's binding. The JAK2 case is the cleanest evidence — L884P differently compromises BBT594 (11-fold) and CHZ868 (4-fold), a drug-specific ordering no single per-variant score can produce How Does the L884P Mutation Confer Resistance to Type-II Inhibitors of JAK2 Kinase: A Comprehensive Molecular Modeling Study. We are nonetheless honest about the threat: the ESM2-weighted ablation is run with the same budget as the full method, and if it matches performance on the drug-specific benchmarks, the conclusion will be that pathogenicity is a sufficient proxy and the proposal's central claim is wrong.
Why not just add MdrDB's static features to the existing graph model? Also a real threat, and we test it directly as an ablation rather than argue around it. The biophysical evidence says static docking should fall short on exactly the allosteric cases: TTR's W61L and Y98F sit far outside the binding site yet impair stabilizer binding Integrative structural profiling and ligand optimisation across the transthyretin mutational landscape, and katG resistance operates by destabilizing a distal heme cofactor rather than affecting drug binding at all Computational analyses of drug resistance mutations in katG and emb complexes in Mycobacterium tuberculosis. If static features match transmission centrality on the out-of-pocket patient subset, the dynamic component earns nothing and the simpler design wins.
Why not just run molecular dynamics on every candidate mutation? It would be the gold standard, and we do not dispute its accuracy — the ALK and JAK2 studies show exactly what it buys Deciphering the Shared and Specific Drug Resistance Mechanisms of Anaplastic Lymphoma Kinase via Binding Free Energy Computation; How Does the L884P Mutation Confer Resistance to Type-II Inhibitors of JAK2 Kinase: A Comprehensive Molecular Modeling Study. But those studies examine one or a handful of systems at 30-nanosecond-to-umbrella-sampling cost, while a single TCGA cohort carries thousands of mutations across hundreds of patients. The compromise here is explicit: MD serves as the calibration and validation set for a fast surrogate, and every claim about the surrogate is checked against simulation numbers, not against other fast models.
Decisive Experiments and Checks
The first experiment validates the allosteric score on systems with known answers. On the JAK2 L884P system the score must rank BBT594 as more compromised than CHZ868, matching the simulated 11-fold versus 4-fold IC50 shifts and the PMF well-depth drop from about 19.8 to 16.7 kcal/mol How Does the L884P Mutation Confer Resistance to Type-II Inhibitors of JAK2 Kinase: A Comprehensive Molecular Modeling Study. On ALK, residues ranked high in transmission centrality should recover the documented ten-residue I1171S chain terminating at the lorlatinib pocket Deciphering the Shared and Specific Drug Resistance Mechanisms of Anaplastic Lymphoma Kinase via Binding Free Energy Computation. On the 133-variant TTR landscape, distal variants such as W61L and Y98F should score as compromising for tafamidis and acoramidis despite their distance from the pocket Integrative structural profiling and ligand optimisation across the transthyretin mutational landscape.
The second experiment is the head-to-head response prediction on SubNetDL's own benchmarks: 507 patients across 10 TCGA cancer–drug combinations, the MSK_2020 and MSK_2018 NSCLC immunotherapy cohorts, and six CCLE/GDSC pairs, where the numbers to beat are the reported median AUROCs of 0.74, 0.77, and 0.81 A network-based deep learning model integrating subclonal architecture for therapy response prediction in cancer. Success is defined in two tiers: overall median AUROC improvement, and — the primary endpoint — improvement specifically among patients whose tumors carry out-of-pocket mutations in the drug target, the tier where the mechanism we propose should concentrate.
The third experiment is the masking stress test, the one that distinguishes this proposal from a feature-engineering exercise. Starting from cell lines and patients with a known resistance driver, we synthesize bulk mixtures in which the resistant lineage is diluted to frequencies from 20% down to below 5%, the regime where SciClone-style VAF clustering is known to lose rare subclones. We then measure how the frequency-weighted resistant fraction $R$ degrades relative to ground truth, comparing against deconvolution-based alternatives in the spirit of Count Bridges, which outperformed CIBERSORTx and MuSiC at recovering cell-type proportions though no numerical value is reported on Lacuna for those comparisons Count Bridges enable Modeling and Deconvolving Transcriptomic Data. The expected signature of the core claim is asymmetry: dilution should hurt prediction for orthosteric and allosteric drivers roughly equally in a binary model, but the frequency-weighted formulation should retain ranking of allosteric drivers longer, because their per-cell effect is encoded as a graded score rather than a bit. If no such asymmetry appears, the hypothesis that the biophysical and clonal masks compound each other is wrong, and the proposal should not proceed past this point.
A note on scope: we deliberately do not attempt single-cell phylogenetic reconstruction at SCIΦ scale SCIΦ: Single-Cell Mutation Identification via Phylogenetic Inference, whose MCMC scheme is documented to scale poorly beyond hundreds of cells; subclone frequencies come from bulk VAF inference exactly as in SubNetDL, so all comparisons are on shared ground. Similarly, transcriptomic deconvolution frameworks such as POPPY Cross-modal transfer learning for mapping bulk transcriptomes at cellular level and the single-cell interaction maps of scPGI-finder Charting cell-type-specific positive genetic interaction at single-cell resolution for lung adenocarcinoma are complementary rather than competing: they resolve which cells are present, we resolve what those cells' mutations do to the drug target. A natural extension, if the core claim survives the three experiments above, is to use POPPY-style attribution to locate the resistant subclone in cell-type space once $R$ flags it — but that is future work, not this proposal.
References
- A network-based deep learning model integrating subclonal architecture for therapy response prediction in cancer
- Deciphering the Shared and Specific Drug Resistance Mechanisms of Anaplastic Lymphoma Kinase via Binding Free Energy Computation
- How Does the L884P Mutation Confer Resistance to Type-II Inhibitors of JAK2 Kinase: A Comprehensive Molecular Modeling Study
- MdrDB: Mutation-induced drug resistance DataBase
- Integrative structural profiling and ligand optimisation across the transthyretin mutational landscape
- Computational analyses of drug resistance mutations in katG and emb complexes in Mycobacterium tuberculosis
- Count Bridges enable Modeling and Deconvolving Transcriptomic Data
- SCIΦ: Single-Cell Mutation Identification via Phylogenetic Inference
- Cross-modal transfer learning for mapping bulk transcriptomes at cellular level
- Charting cell-type-specific positive genetic interaction at single-cell resolution for lung adenocarcinoma