Target Prediction as a Multi-Detector Hypothesis Test

When a collider experiment wants to know whether a tracked particle is a pion or a kaon, no single detector answers the question. Instead, each subsystem — ionization loss in the drift chamber, flight time, Cherenkov angle — is calibrated into a probability density for each species hypothesis, and the subsystems are combined into a joint likelihood or a delta-log-likelihood ratio that cleanly separates signal from background Measurement of the decay rate of $\Xi_c^0 \to p K^- K^- \pi^+$ relative to $\Xi_c^0 \to \Xi^- \pi^+$, Observation of the Decay $B^+ \to K^+ K^- \pi^+$. Computational target prediction in network pharmacology faces a structurally identical problem and solves it badly. A candidate drug is scored by several noisy "detectors" — chemical-similarity nearest neighbors, docking energies, sequence- or structure-based deep models, transcriptomic proximity — and these scores are then either intersected as set Venn diagrams or concatenated as uncalibrated features. The scores from different channels are never given explicit false-positive rates, never combined as a proper hypothesis test, and always binarized by an arbitrary cutoff before being fed into protein-network diffusion analyses. I propose to import the particle-identification (PID) fusion formalism wholesale: treat each target-prediction evidence channel as a calibrated sub-detector, combine channels into target-versus-decoy log-likelihood ratios with explicit correlation corrections, and propagate the resulting continuous likelihood ratios — not binary calls — as soft seeds through network diffusion.

The opportunity is concrete because the field's own diagnostics point at exactly what this fixes. TAPB: an interventional debiasing framework for alleviating target prior bias in drug-target interaction prediction proved that current drug-target interaction (DTI) models mostly memorize per-target label imbalance ("target prior bias") rather than binding mechanism, and collapse on novel targets. A likelihood-ratio formulation attacks this by construction: $\Delta\ln\mathcal{L} = \sum_i \ln \frac{P(x_i|\text{active})}{P(x_i|\text{decoy})}$ is a ratio of densities, so the per-target prior cancels from the decision statistic and only the evidence survives. Separately, Diffusion mapping of drug targets on disease signaling network elements reveals drug combination strategies showed that Random Walk with Restart (RWR) — a diffusion process that spreads influence from a starting set of proteins across the interaction network — recovers 76% of validated drug combinations within its top 30% of predictions, but only when fed clean, experimentally confirmed binary targets; nobody has shown it survives realistic, uncertain predicted targets. The HEP literature has already demonstrated, in its own domain, that hard cuts on noisy detectors destroy rare signals: only an unbinned likelihood fit over all detector layers let BABAR extract $429 \pm 43$ signal events at $9.6\sigma$ significance Observation of the Decay $B^+ \to K^+ K^- \pi^+$. The same intuition says binarizing predicted targets before diffusion throws away most of the useful information.

Core Claim

The core claim has two parts. First, computational target prediction can be reformulated as a multi-detector hypothesis test in which each evidence channel $i$ (2D fingerprint similarity, docking score, deep-model score such as SiamDTI's or SaBAN's output, transcriptomic co-expression) is calibrated into a pair of densities, $P(x_i|\text{active})$ and $P(x_i|\text{decoy})$, using target-matched decoy libraries, and the channels are fused into a single log-likelihood-ratio statistic — with a learned correlation model replacing the physics community's known-broken conditional-independence assumption. Second, this continuous statistic can replace binary seed sets in network diffusion, seeding RWR with per-target likelihood ratios so that downstream disease-mechanism and drug-combination analyses inherit well-calibrated uncertainty instead of arbitrary cutoffs.

This differs from the closest existing work in the way that matters. A Cross-Field Fusion Strategy for Drug-Target Interaction Prediction (SiamDTI) fuses information by subtracting one latent vector from another inside a Siamese neural network — a deterministic, uncalibrated operation on embeddings with no probabilistic decision layer and no downstream network use. Structure-Aware Contrastive Learning with Fine-Grained Binding Representations for Drug Discovery (SaBAN) is a single end-to-end affinity predictor whose output is a raw score, not a calibrated likelihood, and it stops at the drug-target pair without touching the interactome. Diffusion mapping of drug targets on disease signaling network elements reveals drug combination strategies is the downstream half alone, and explicitly assumes known binary targets. No published work, per the Lacuna search, supplies the middle layer: channel-level calibration, correlation-aware likelihood fusion, and continuous propagation into diffusion. The HEP papers that invented this machinery Direct measurements of the branching fractions for D0→K−e+νe and D0→π−e+νe and determinations of the form factors f+K(0) and f+π(0), First Observation of the Decay $\tau^- \to K^{*-} \eta \nu_\tau$ never left relativistic kinematics.

Research Program

Three components need to be built. The first is channel calibration. For each evidence channel I will estimate decoy-conditional densities using target-matched decoys of the kind LIT-PCBA already provides (verified inactive actives/decoys per target), plus the large curated binding data in BindingDB BindingDB in 2024: a FAIR knowledgebase of protein-small molecule binding data. This is the direct analogue of the sideband and continuum subtraction used in the BES-II and BABAR analyses: the null density is estimated from populations known to be background, so the ratio isolates the binding signal the way $\Delta\ln\mathcal{L}_{K-\pi}$ isolates the kaon signal. Calibration will be done per target family where labels exist and transferred to novel targets at the family level, since a genuinely cold target has no per-target labels — this transfer is itself a research question the program must answer.

The second component is correlation-aware fusion. The PID literature's own open problem is that the product-of-densities assumption fails because subsystems share error sources (common tracking errors, material effects). The biological analogue is worse: docking scores and sequence-model scores both see the same protein, so their errors are strongly correlated, and naive multiplication of densities will double-count evidence. I will fit a compact joint model over the channel responses — a Gaussian copula or a small conditional normalizing flow over the 4–6 channel scores, conditioned on the active/decoy hypothesis — so the fused statistic uses the modeled joint density rather than the product of marginals. This is the point of closest tension with SiamDTI, which handles cross-channel redundancy by subtractive latent cancellation; I will test whether explicit statistical modeling of the correlation beats latent subtraction at equal compute.

The third component is soft-seed diffusion. In the RWR framework of Xu et al., a walker restarts at a fixed set of seed proteins, and the stationary visitation distribution measures each network node's exposure to the drug's perturbation. I will replace the binary seed indicator with the per-target log-likelihood ratio (mapped through a softmax-weighted restart distribution), so a strongly evidenced target pulls more diffusion mass than a weakly evidenced one, and targets near the decision boundary contribute proportionally rather than being discarded or admitted wholesale. The downstream scoring, including the affinity-propagation community clustering for drug combinations, is otherwise unchanged from Xu et al., which isolates the soft-seed change as the only intervention.

Baselines

The naive reference point is the standard network-pharmacology intersection pipeline: query TCMSP-style target databases and GeneCards-style disease-gene databases, take the set intersection, and analyze the resulting subnetwork in STRING, as done in Bioinformatics-based study of the ceRNA network and the sub-network of the Ziyin Buyang Formula in premature ovarian failure due to ovarian tumour therapy and To explore immune synergistic function of Quercetin in inhibiting breast cancer cells. It is the field's default and must be beaten for the proposal to matter; success against it means higher drug-combination recovery at fixed network and dataset inputs. The strongest existing end-to-end method on the cold-split question is SiamDTI, which reports AUROC of 0.718 on novel-split BioSNAP, 0.627 on novel-split BindingDB, and 0.863 on novel-split Human A Cross-Field Fusion Strategy for Drug-Target Interaction Prediction; the proposal must match or exceed these on identical splits, with identical decoys, and under an identical tuning budget (the same number of random-search iterations over each method's hyperparameters on the same held-out validation folds, same GPU hardware). SaBAN is the strongest competitor on the virtual-screening regime, reporting AUROC of 68.16% and BEDROC of 13.35% on LIT-PCBA Structure-Aware Contrastive Learning with Fine-Grained Binding Representations for Drug Discovery; it will be run as a channel inside my fusion and also standalone as a baseline, which doubles as the key ablation: fusion-with-SaBAN-as-one-channel versus SaBAN-alone isolates the value of calibration and multi-channel combination. For the downstream task, the strongest baseline is the original binary-seed RWR of Xu et al. (76% recovery of validated combinations in its top 30%, versus 69% for SynGeNet) on the GSE15605 BRAF-mutant melanoma data Diffusion mapping of drug targets on disease signaling network elements reveals drug combination strategies; beating it means higher recovery at the same top-fraction cutoff when targets are predicted rather than known — the setting Xu et al. could not test. A further ablation removes the correlation model (pure independent-product likelihood) and another removes calibration (fuse raw scores), isolating each ingredient's contribution. For TAPB, its reported in-domain BioSNAP improvement is 2.3% AUROC over the next-best baseline PSICHIC TAPB: an interventional debiasing framework for alleviating target prior bias in drug-target interaction prediction; its code-based comparison on Davis and cold splits has no reported number found on Lacuna beyond that summary figure, so it will be re-run under my protocol rather than compared on quoted numbers.

Why not just…?

Why not just train XGBoost on the concatenated channel scores? This is the most serious alternative, and I will run it as a first-class baseline rather than argue around it. The evidence against it is that concatenation-trained models absorb target prior bias: TAPB showed such models memorize per-target label imbalance and collapse on cold splits TAPB: an interventional debiasing framework for alleviating target prior bias in drug-target interaction prediction, and SiamDTI's own baselines found plain DrugBAN-style concatenation reaching only about 0.582 AUROC on novel BindingDB splits versus 0.627 after subtractive isolation A Cross-Field Fusion Strategy for Drug-Target Interaction Prediction. The likelihood-ratio design cancels the prior analytically and gives per-channel, per-target interpretable evidence. But honestly, if XGBoost on the same channels matches my fused statistic on cold splits at the same tuning budget, that is a decisive finding that calibration structure is unnecessary — and the experiments are constructed so that this outcome would be visible and reportable, not hidden.

Why not just fine-tune one bigger end-to-end model (SaBAN or a protein language model)? SaBAN's own results show the ceiling of the pure end-to-end route: it is state of the art on Human and BioSNAP but only "competitive but not state-of-the-art" on BindingDB Structure-Aware Contrastive Learning with Fine-Grained Binding Representations for Drug Discovery, and sequence-only models without structural augmentation lost 10.99 AUROC points on LIT-PCBA relative to SaBAN itself — evidence that single-architecture capacity does not substitute for combining orthogonal evidence. More fundamentally, an end-to-end model outputs an uncalibrated score and stops at the drug-target pair; it cannot by construction feed calibrated uncertainty into downstream network diffusion, which is half of this proposal's claim. The SaBAN-standalone baseline and the SaBAN-as-a-channel ablation will quantify exactly how much the fusion layer adds on top of the strongest single model.

Why not just threshold the predictions before diffusion? This is the current default, and the HEP record is unambiguous about its cost: hard sub-detector cuts caused prohibitive signal loss in rare-decay searches, and only unbinned likelihood fusion recovered the signal Observation of the Decay $B^+ \to K^+ K^- \pi^+$, First Observation of the Decay $\tau^- \to K^{*-} \eta \nu_\tau$. The biological version of the argument is that polypharmacological efficacy comes from many weak bindings that a high cutoff truncates Diffusion mapping of drug targets on disease signaling network elements reveals drug combination strategies. Still, this alternative is cheap and could work, so the decisive experiment is a sweep: binarize predicted targets at several cutoffs (top-5, $P>0.9$, $P>0.5$) versus soft seeding with the same underlying predictions, all on the GSE15605 combination-recovery task. If some cutoff matches soft seeding, the soft-seed claim dies; if recovery degrades monotonically as information is discarded, the claim stands.

Why not just keep the standard database-intersection pipeline? It misses indirect mechanisms entirely: in Xu et al.'s melanoma analysis, 16 of 17 validated synergistic combinations paired drugs acting on different functional communities — no direct target-disease intersection would find them — and the intersection approach has no way to express that a predicted target is uncertain Diffusion mapping of drug targets on disease signaling network elements reveals drug combination strategies. It also cannot be evaluated on cold targets at all, since it requires the target to already be in a database. The intersection pipeline stays in the baseline set as the "do nothing" reference, not as a competitor expected to win.

Proposed Experiments

The evaluation runs in three stages, all on benchmarks present in the file. Stage one tests the fusion statistic head-to-head on cold splits. Following the SiamDTI protocol, I will evaluate on BioSNAP, BindingDB, and Human under known and novel (zero-shot drug and target) splits, reporting AUROC and AUPRC, and additionally on Davis for in-domain comparison with TAPB, and on LIT-PCBA with its verified decoys for the virtual-screening regime, reporting AUROC and BEDROC. The headline targets are the SiamDTI novel-split numbers (0.718 BioSNAP, 0.627 BindingDB, 0.863 Human) and SaBAN's LIT-PCBA AUROC of 68.16%; the proposal succeeds on this stage only if the fused statistic exceeds them on identical splits, or matches them while producing calibrated probabilities (measured by reliability curves and expected calibration error on cold splits, which no cited method reports).

Stage two tests the correlation claim. I will construct controlled degradations: deliberately corrupt one channel with noise that is shared with another channel (e.g., propagate the deep-model score's protein-identity dependence into the docking channel) and measure how the independent-product fusion, the copula/flow fusion, and SiamDTI-style latent subtraction each degrade. The prediction, from the PID open-problem literature on correlated sub-detector systematics, is that independent-product fusion degrades fastest; if the correlation model shows no measurable advantage, the simple product form wins on parsimony and the copula component should be dropped.

Stage three tests the downstream claim. On the GSE15605 melanoma dataset with its 131 prioritized signaling genes, I will regenerate Xu et al.'s drug-combination analysis, but with targets supplied by the stage-one predictor rather than by the database. The decisive comparison is soft-seed RWR versus binary-threshold RWR at matched prediction quality, with recovery of validated combinations in the top 30% of predictions as the metric; the number to beat is the 76% recovery Xu et al. report with ground-truth targets, which sets the ceiling, and the binary-threshold arms set the realistic floor. A secondary check on the same task: whether the likelihood-weighted seeds produce better-calibrated impact scores, evaluated by whether high-impact predicted targets are enriched in the literature-validated set.

The final decisive check ties the two fields together in the way only this design allows: propagate the per-channel calibration error (the spread of the bootstrapped density estimates) through the fusion and the diffusion, and test whether the reported confidence intervals on downstream combination scores actually cover the validation outcomes — the property that made likelihood-based PID analyses trustworthy enough to quote systematic uncertainties alongside statistical ones, and the property that no cited target-prediction or network-pharmacology method currently provides.

References

  1. Measurement of the decay rate of Ξc0→pK-K-π+ relative to Ξc0→Ξ-π+
  2. Observation of the Decay<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:msup><mml:mi>B</mml:mi><mml:mo>+</mml:mo></mml:msup><mml:mo>→</mml:mo><mml:msup><mml:mi>K</mml:mi><mml:mo>+</mml:mo></mml:msup><mml:msup><mml:mi>K</mml:mi><mml:mo>−</mml:mo></mml:msup><mml:msup><mml:mi>π</mml:mi><mml:mo>+</mml:mo></mml:msup></mml:math>
  3. TAPB: an interventional debiasing framework for alleviating target prior bias in drug-target interaction prediction
  4. Diffusion mapping of drug targets on disease signaling network elements reveals drug combination strategies
  5. A Cross-Field Fusion Strategy for Drug-Target Interaction Prediction
  6. Structure-Aware Contrastive Learning with Fine-Grained Binding Representations for Drug Discovery
  7. Direct measurements of the branching fractions for D0→K−e+νe and D0→π−e+νe and determinations of the form factors f+K(0) and f+π(0)
  8. First Observation of the Decay<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:mrow><mml:msup><mml:mrow><mml:mi mathvariant="italic">τ</mml:mi></mml:mrow><mml:mrow><mml:mo>−</mml:mo></mml:mrow></mml:msup></mml:mrow><mml:mi/><mml:mo>→</mml:mo><mml:mi/><mml:mrow><mml:msup><mml:mrow><mml:mi mathvariant="italic">K</mml:mi></mml:mrow><mml:mrow><mml:mi>*</mml:mi><mml:mo>−</mml:mo></mml:mrow></mml:msup></mml:mrow><mml:mi mathvariant="italic">η</mml:mi><mml:mrow><mml:msub><mml:mrow><mml:mi>ν</mml:mi></mml:mrow><mml:mrow><mml:mi>τ</mml:mi></mml:mrow></mml:msub></mml:mrow></mml:math>
  9. BindingDB in 2024: a FAIR knowledgebase of protein-small molecule binding data
  10. Bioinformatics-based study of the ceRNA network and the sub-network of the Ziyin Buyang Formula in premature ovarian failure due to ovarian tumour therapy
  11. To explore immune synergistic function of Quercetin in inhibiting breast cancer cells