When a tumor is sequenced in bulk, the DNA of millions of cells is pooled into one averaged readout, and the tumor's internal structure — its subclones, the distinct descendant lineages that evolve separately and respond differently to drugs — is hidden inside that average. The strongest existing remedy, a model called SubNetDL, recovers those subclones from patterns in mutation frequencies and maps each subclone's mutations onto a protein-interaction network to predict therapy response, reaching a median AUROC of 0.74 across ten cancer–drug pairs in The Cancer Genome Atlas and 0.77 in two lung-cancer immunotherapy cohorts A network-based deep learning model integrating subclonal architecture for therapy response prediction in cancer. But SubNetDL records every somatic mutation as a binary on/off switch: a stop codon that destroys a protein counts exactly as much as one that trims a dispensable tail.
That simplification collides with a well-established fact from the other half of this proposal's territory. The protein-truncation literature has shown, case by case, that a protein's biological action is concentrated in a compact active core, and that removing flanking or regulatory sequence is often harmless — sometimes beneficial. Truncated mini-dystrophins stabilize muscle membranes better than the full-length protein because they are structurally less fragile Duchenne muscular dystrophy treatment with lentiviral vector containing mini‐dystrophin gene in vivo; cutting the eight-residue hormone angiotensin II down to four residues keeps the antimalarial activity while deleting the segment that raises blood pressure New linear antiplasmodial peptides related to angiotensin II; and a minimal 23-amino-acid fragment of ACE2 retains enough binding to drive targeted degradation of the SARS-CoV-2 spike protein Targeted intracellular degradation of SARS-CoV-2 via computationally optimized peptide fusions. The same principle read in reverse is a prediction rule for cancer mutations: a truncating mutation that shears the active core destroys function, while one that removes only flanking or regulatory regions may leave the cell's biochemistry essentially intact. No subclone-resolved model of drug response uses this rule.
Core Claim
Replacing binary mutation indicators in subclonal drug-response models with graded, structure-resolved functional-core integrity scores — calibrated so their error rates are statistically controlled — will improve therapy-response prediction and will specifically resolve a documented class of false positives: truncating mutations currently counted as functional drivers that actually preserve the core. Formally, where SubNetDL assigns each subclone-specific mutation $m_i$ a node feature $x_i \in \{0,1\}$, we assign
$$x_i = 1 - \pi(m_i),$$
where $\pi(m)$ is the calibrated probability that mutation $m$ leaves the minimal functional core of its protein intact, so $x_i$ becomes an expected functional disruption rather than a presence flag. The hypothesis is falsifiable in the sharpest possible way: if core-preserving and core-shearing truncations predict drug sensitivity equally well, the truncation-biology prior does not transfer to somatic mutations and the idea dies.
Research Program
Three components, each reusing an existing technique rather than inventing one. First, a core-integrity annotator: for every protein in SubNetDL's 253-gene driver network, we extract minimal functional cores — the compact segments carrying binding or catalytic activity — using the Rosetta PeptiDerive protocol that Chatterjee et al. used to reduce full-length ACE2 to its 23-mer interface Targeted intracellular degradation of SARS-CoV-2 via computationally optimized peptide fusions, run over AlphaFold-predicted structures. To make this tractable at the scale of the whole driver network, we pre-filter structures with embedding-based search in the style of the conformal protein-mining work, which cut a 2.3-million-structure AlphaFold database search from roughly one day to 30 seconds while retaining more than 80% of high-quality hits Functional protein mining with conformal guarantees. Each mutation is then classified as core-shearing (a nonsense or frameshift change that cuts into or deletes the core), core-preserving (it removes only flanking or regulatory sequence — the mini-dystrophin regime), core-residue missense, or putatively neutral. Machine-learned subsequence scoring in the spirit of MLPD, which identified aptamers 70% shorter with equal binding at an 11-fold higher hit rate than random perturbation Machine learning guided aptamer refinement and discovery, supplies the activity-retention estimate for ambiguous cases.
Second, we wrap the annotation in conformal prediction — a wrapper that converts raw model scores into labels with a guaranteed bound on the fraction of errors — following the demonstration that base protein-embedding searchers can be given rigorous false-discovery-rate control without retraining, yielding confident functional assignments for 39.6% of previously unannotated essential genes at a 10% false-discovery threshold on the JCVI Syn3.0 minimal genome Functional protein mining with conformal guarantees. This matters because downstream clinical predictions inherit annotation noise, and we want an explicit, tunable error budget rather than silent misclassification.
Third, the truncation-aware subclonal model: we keep SubNetDL's entire pipeline fixed — SciClone clustering of bulk variant allele frequencies into one to six subclones, signal propagation over the STRING interaction network via multiscale PageRank with damping factors from 0.05 to 0.85, pruning to 253 driver genes, and a graph attention network (a network that learns which neighboring proteins matter for the prediction) — and change only the node features, from binary indicators to the graded vector above. Holding the architecture fixed is deliberate: any performance delta is attributable to the truncation-awareness alone.
Baselines
The naive reference point is tumor mutational burden (TMB), the total count of mutations in a bulk sample, used as a threshold-based responder classifier. SubNetDL's authors already benchmarked against it and won decisively in both lung-cancer immunotherapy cohorts ($P = 0.02$ in MSK_2020, $P = 6.0 \times 10^{-3}$ in MSK_2018), and it cannot distinguish which mutations matter; no reported result found on Lacuna for TMB's median AUROC, so we will run it ourselves on the identical cohorts and splits to get that number. The strongest competitor is SubNetDL itself in its exact published configuration, reporting median AUROC of 0.74 on the 507-patient, ten cancer–drug-pair TCGA cohort, 0.77 across the two MSK non-small-cell lung cancer immunotherapy cohorts (242 and 240 patients), and 0.81 on CCLE/GDSC cancer-cell-line drug-sensitivity pairs without cohort-specific re-tuning A network-based deep learning model integrating subclonal architecture for therapy response prediction in cancer. We will rerun it ourselves under matched conditions — same data, same cross-validation folds, same hyperparameter search ranges and tuning budget, same compute — so the comparison is controlled rather than transcribed. The key ablation is a domain-only variant: the identical architecture with features reduced to a binary "mutation falls inside an annotated domain" flag, which isolates whether structure-resolved core geometry adds anything beyond coarse domain membership; a second ablation randomly permutes core labels among mutations to confirm the signal is not an artifact of feature smoothness. Beating SubNetDL means a median AUROC at least 0.05 higher on the TCGA cohort and a statistically significant improvement in both MSK cohorts under paired tests, plus better reclassification of high-TMB non-responders — the false positives SubNetDL already showed TMB produces.
Why not just…?
Why not just use bulk biomarkers like TMB or bulk gene signatures? TMB is significantly outperformed by subclone-aware modeling in both MSK cohorts A network-based deep learning model integrating subclonal architecture for therapy response prediction in cancer, and bulk expression signatures can miss the mechanism of resistance entirely: tamoxifen resistance in breast cancer is driven by a rare stem-like subpopulation whose signal is averaged away in bulk measurements Integrative analyses of gene expression and DNA methylation profiles in breast cancer cell line models of tamoxifen-resistance indicate a potential role of cells with stem-like properties. Aggregate biomarkers cannot ask which mutations in which subclone break function, which is precisely our question.
Why not just use established deconvolution tools like CIBERSORTx, EPIC, or MuSiC? These answer a different question. They estimate what cell types are present in a bulk expression profile, and a 38,000-mixture benchmark built from a 100,064-cell breast cancer atlas showed they degrade badly at high tumor purity — misclassifying normal epithelial lineages as cancer — while high-specificity methods like hspe miss rare populations altogether Performance of tumour microenvironment deconvolution methods in breast cancer using single-cell simulated bulk mixtures. None of them model mutation-level functional consequence within a subclone. We do note honestly that our approach inherits a limitation from the same family: SciClone's clustering of bulk variant allele frequencies caps resolution at one to six subclones and cannot see ultra-low-frequency resistant lineages, so our claims are about subclone-level, not single-cell-level, prediction.
Why not just add domain-level annotations as features? This is the real threat, and we do not argue around it — we test it. The domain-only ablation is run with the same tuning budget as the full method, and if it matches the core-resolved variant, the conclusion is that domain membership suffices and the structural layer is unnecessary; we would report that. What domain annotation cannot capture, and what motivates the extra resolution, is the demonstrated fact that large annotated portions of proteins can be removed with function intact — mini-dystrophins outperform the full-length protein Duchenne muscular dystrophy treatment with lentiviral vector containing mini‐dystrophin gene in vivo and four-residue angiotensin fragments keep the activity while losing the toxicity New linear antiplasmodial peptides related to angiotensin II. Only core geometry predicts which truncations are tolerated.
Why not go straight to designing truncated therapeutics against the subsets that deconvolution identifies? The more ambitious intersection — feeding cell-type attribution scores from a tool like POPPY, which traces bulk drug responses back to specific effector populations such as memory B cells and NK cells Cross-modal transfer learning for mapping bulk transcriptomes at cellular level, into the generative objective of a discrete-diffusion protein designer like NOS/LaMBO-2, which achieved 99% expression and 40% binding in antibody optimization Protein Design with Guided Discrete Diffusion — currently has no ground truth to train or evaluate against: NOS optimizes against static purified targets, and POPPY outputs no per-subclone target expression. Our proposal is the prerequisite layer: until we can tell real subclone-specific functional disruptions from artifacts, conditioning a generator on them would optimize against noise.
Proposed Experiments
The first experiment validates the annotator before any downstream claim. We reproduce the conformal wrapper's reported behavior on its own benchmarks — 39.6% of unannotated essential genes assigned at a 10% false-discovery rate on the JCVI Syn3.0 minimal genome, and the enzyme-classification result on the Price-149 and New-392 sets, for which no numeric accuracy is given on Lacuna and will be measured directly Functional protein mining with conformal guarantees — and then apply it to the 253 driver genes, reporting the headline descriptive number the field lacks: what fraction of truncating driver mutations are core-preserving rather than core-shearing.
The second and third experiments are the head-to-head predictions. On the TCGA multi-cancer drug cohort (507 patients, ten cancer–drug pairs, six solid cancer types) we compare median AUROC against SubNetDL's published 0.74 under a matched protocol; on the MSK_2020 and MSK_2018 anti-PD-(L)1 lung cancer cohorts (242 and 240 patients) we compare against 0.77 and additionally count how many high-TMB non-responders each model correctly flags as false positives — SubNetDL reports reclassifying "dozens" of such patients, with no exact count on Lacuna, so we will quantify it on both models.
The fourth experiment is the decisive mechanistic check, run on CCLE/GDSC cell-line mutations paired with drug sensitivity measurements, where SubNetDL reports a median AUROC of 0.81 across six cancer–drug pairs without cohort-specific re-tuning. Among cell lines carrying truncating mutations in a given drug target or pathway gene, sensitivity should separate cleanly by annotation: cell lines with core-shearing truncations should behave like loss-of-function mutants, while core-preserving truncations should be indistinguishable from wild type. This per-mutation test, not the aggregate AUROC, is what would confirm or refute the core claim. As an exploratory fifth analysis, we stratify the 1,038-specimen TCGA breast cancer cohort by tumor purity, since the deconvolution benchmark validated purity-driven misclassification there Performance of tumour microenvironment deconvolution methods in breast cancer using single-cell simulated bulk mixtures, and ask whether the truncation-aware features remain stable across the purity range where lineage artifacts flourish.
If the graded features beat SubNetDL and the domain-only ablation, the case is made: the truncation principle is a missing, cheaply computable layer of biochemical realism for every model that maps subclonal mutations onto protein networks. If they beat SubNetDL but not the domain-only ablation, the honest finding is that coarse domain membership carries the signal. And if core-preserving truncations predict sensitivity just like core-shearing ones in the CCLE test, then the engineering intuition that function lives in compact cores does not transfer to somatic cancer mutations — a negative result, but one worth knowing before anyone builds a therapeutic-design pipeline on top of it.
References
- A network-based deep learning model integrating subclonal architecture for therapy response prediction in cancer
- Duchenne muscular dystrophy treatment with lentiviral vector containing mini‐dystrophin gene in vivo
- New linear antiplasmodial peptides related to angiotensin II
- Targeted intracellular degradation of SARS-CoV-2 via computationally optimized peptide fusions
- Functional protein mining with conformal guarantees
- Machine learning guided aptamer refinement and discovery
- Integrative analyses of gene expression and DNA methylation profiles in breast cancer cell line models of tamoxifen-resistance indicate a potential role of cells with stem-like properties
- Performance of tumour microenvironment deconvolution methods in breast cancer using single-cell simulated bulk mixtures
- Cross-modal transfer learning for mapping bulk transcriptomes at cellular level
- Protein Design with Guided Discrete Diffusion