Machine learning can now answer "is this gene pair synthetically lethal?" — two genes whose joint loss kills a cell that survives losing either one alone — with respectable accuracy. But the question that actually decides what a lab works on next is a ranking question: "given this tumor's missing suppressor genes, which of five hundred candidate partner genes should we spend the next two years on?" The benchmark study of Benchmarking machine learning methods for synthetic lethality prediction in cancer found precisely this asymmetry: models do well at binary classification and degrade sharply at prioritization, especially in the cold-start setting where neither gene of a candidate pair was seen in training. What fills that gap today is human judgment — medicinal chemists and oncologists informally weighing druggability, safety, evidence strength, and subtype specificity. And structured human judgment of exactly this kind is known to carry systematic bias: in The Research of Electronic Banking Risk Evaluation Based on Comprehensive Assessment AHP-Entropy, five experts assigned "security management" a weight of 0.910 that an objective entropy filter computed from the actual data corrected down to 0.614; in Operation Risk Assessment of Urban Dense Cable Passageway Based on Fuzzy-Analytic Hierarchy Process, panels passed mathematical consistency checks while remaining anchored to idiosyncratic priors; and in Scientific Challenges in Performing Life-Cycle Assessment in the Food Supply Chain, consensus dynamics were shown to codify a shared bias rather than remove it.
The proposal here is to build and rigorously evaluate a de-biased ranking layer that sits between synthetic-lethality predictors and the humans who act on them, for aggressive receptor-poor subtypes (triple-negative breast cancer, basal-like breast cancer, gastric signet ring cell carcinoma) where target choice is most consequential and most contested. This is, to my knowledge, the first attempt to port the hybrid objective-subjective weighting architecture — developed for banking risk, hazardous-material routing, and utility hardware — into molecular target prioritization, and the first to test whether its de-biasing effect can be measured against independently validated targets rather than against simulated panels.
Core Claim
The claim has two parts, one descriptive and one causal. Descriptively: target prioritization in aggressive-cancer biology is already a multi-criteria decision process, just an informal and biased one — the GSRCC study of A comprehensive proteomic analysis uncovers novel molecular subtypes of gastric signet ring cell carcinoma moved from 7,322 quantified proteins to four NMF subtypes and then to a single hand-picked repurposing candidate (neratinib against PFAS) through exactly this kind of uncalibrated expert triage. Causally: if we make the criteria explicit, elicit expert pairwise weights over them (Analytic Hierarchy Process), and temper each expert-assigned weight with a Shannon-entropy weight computed from how much each evidence channel actually discriminates among candidate genes within a specific molecular subtype, the resulting ranking will retrieve independently validated therapeutic targets at higher ranks than raw predictor scores, pure expert judgment, or pure data dispersion — and the entropy layer will demonstrably correct the expert bias, in the same direction and by a measurable magnitude as Wang's banking result.
The subtype-conditioning is the scientifically load-bearing part. eEF2K dependency exists only in the claudin-low TNBC created by combined Pten/p53 deletion (Combined deletion of Pten and p53 in mammary epithelium accelerates triple‐negative breast cancer with dependency on eEF2K — dual deletion collapsed tumor latency to 9.8–11.3 months versus 15.2–26.4 months for single Pten deletion, and TX-1918 sensitivity tracked AKT activation at $r = -0.70$). A global ranking that treats "the average tumor" would never surface a dependency this conditional. So the entropy of each evidence channel is computed per subtype, not once for the whole disease, and the expert weights are likewise allowed to differ by subtype. This also addresses the "static snapshot" limitation of prior hybrid models: as the candidate pool or cohort drifts, the objective weights recompute.
The Scoring Layer
For a candidate gene $g$ within a subtype $s$, we assemble five evidence channels: the synthetic-lethality predictor's score for pairs $(g, \text{lost suppressor})$; the subtype-conditional effect size of CRISPR knockout from DepMap; a Cox hazard ratio for expression of $g$ in subtype-$s$ tumors from TCGA-BRCA and METABRIC; an expert AHP prior over druggability, tractability, and safety elicited by pairwise comparison; and a cross-cohort reproducibility score. Each channel is normalized to $[0,1]$, and the channel weights combine multiplicatively in the style Wang validated:
$$w_c^{obj} = \frac{1 - H_c/\ln N}{\sum_k \big(1 - H_k/\ln N\big)}, \qquad\quad w_c = \frac{w_c^{AHP} \cdot w_c^{obj}}{\sum_k w_k^{AHP} \cdot w_k^{obj}},$$
where $H_c$ is the Shannon entropy of channel $c$'s values across the $N$ candidate genes in subtype $s$, $w_c^{AHP}$ is the expert weight from the pairwise matrix, and low entropy (high contrast between candidates) earns high objective weight. The intuition, carried over from the entropy-weight literature, is that a channel which assigns nearly the same value to every candidate is uninformative for ranking, however loud the experts are about it — this is what pulled Wang's security-management weight from 0.910 to 0.614. One design change is needed for molecular data: when a channel is near-flat across candidates — a known singularity in entropy weighting that would zero out clinically vital but low-dispersion signals, exactly the regime of master regulators like G9a in Targeting G9a Exerts Pleiotropic Suppression in Triple-Negative Breast Cancer Cells — we smooth the empirical distribution with a small Dirichlet prior before computing $H_c$, and we report rank sensitivity to the smoothing constant. Candidates are ranked by the weighted sum of channel scores.
Research Program
Three workstreams. First, criterion construction and subtype assignment: attach every TCGA-BRCA (966 patients) and METABRIC (1,904 patients) tumor to a subtype label, using either published calls or a reproducible classifier — MOGAT reports macro-F1 of 0.804 on TCGA-BRCA with three omics layers, and AET-net reaches accuracy 0.912 and AUC 0.9862 with an autoencoder-compressed multi-omics Transformer — and compute the objective channels per subtype. The basal-like subtype is a good internal anchor because Exportin-T Promotes Breast Cancer Progression gives an external reference for subtype-conditional biomarker behavior (XPOT classifies basal-like at AUC 71.8% in TCGA and 74.3% in METABRIC, with median survival 3,472 versus 4,267 days for high versus low expression).
Second, the expert-elicitation study — the part that makes this a bias experiment rather than a scoring heuristic. We recruit at least 15 oncologists and medicinal chemists (deliberately more than Wang's five, since small panels are a documented instability), elicit pairwise AHP matrices over the five criteria per subtype, and measure inter-rater variance in the resulting weights. We then run two deliberately biased conditions alongside the independent elicitation: a simulated Delphi panel (sequential disclosure of others' weights, the groupthink amplifier documented by Đjekić et al.) and a salience-primed condition where experts first read a vivid case report of one target. The quantity of interest is how much the entropy layer moves expert weights, whether it moves different experts toward each other (de-biasing) or apart (correcting for genuine disagreement), and whether the correction magnitude is comparable to the 33% reduction Wang observed.
Third, the retrieval evaluation: mask all validation evidence for a set of independently confirmed targets — eEF2K (whose 24-gene signature carries HR = 2.24, $P = 0.0124$ in claudin-low patients), XPOT, PFAS/neratinib in GSRCC, G9a, and PPARγ-after-HDAC-inhibition from Epigenetic derepression converts PPARγ into a druggable target — from the training data, and ask where each scoring scheme ranks them among hundreds of decoys. PPARγ is the hard negative control: experts rating "druggability" without the HDAC-context knowledge would rank it low, so it tests whether the objective channels can rescue a target that a static expert prior actively misleads on.
Baselines
The do-nothing reference is ranking candidates by the raw synthetic-lethality predictor score, a popularity proxy (count of recorded interactions in SynLethDB), and uniform channel weights. This is the de facto current practice, and it requires no elicitation, so any added machinery must justify itself against it on the same candidate pools. For the strongest existing competitor we run SLMGAE, the multi-view graph autoencoder that the benchmark study identified as the top overall model — it maintains an F1-score of about 0.73 in the cold-start CV3 split where neither gene of a candidate pair was seen in training, and where a competing deep model, GCATSL, collapsed to an uninformative score spike. That F1 is a binary-classification number; no reported ranking result (recall@k or similar) exists on Lacuna for SLMGAE or any other model in that benchmark, which is itself part of the motivation. We will re-run SLMGAE ourselves under the benchmark's CV1/CV2/CV3 splits and 1:1 to 1:50 imbalance ratios, using its published architecture and our identical candidate pools, so the comparison is same-data and same-tuning-budget (grid search over the same hyperparameter ranges for every method). We also include GRSMF, the matrix-factorization model the benchmark found "remarkably stable" as imbalance increased, as a shallow-learning competitor. Beating the baselines means: at CV3 cold start with 1:20 imbalance, the entropy-calibrated ranking places a strictly higher fraction of masked validated targets in the top 20 and top 50 than raw SLMGAE scores, GRSMF scores, and uniform weights, with the gap significant under bootstrap over candidate pools. The ablation that isolates the key ingredient is the full fusion against its two halves: pure AHP weights with no entropy layer (expert judgment alone), and pure entropy weights with no experts (data dispersion alone, the Zhang-style approach from Optimization Method of Calibration Cycle Based on State Evaluation Results of Electric Energy Meters, where pure CV weighting concentrated 73.26% of field faults into the abnormal 22% of meters). If the fused version does not beat both halves on retrieval, the central claim is wrong and we would report that.
Why not just…?
Why not just improve the graph neural network end-to-end? Because the failure the benchmark paper documents is not a capacity failure but an objective failure: models are trained on binary classification loss, and they rank poorly even when they classify well. A bigger or better-tuned SLMGAE still optimizes the wrong objective, and the CV3 collapse of GCATSL shows deep models degrade unpredictably exactly in the out-of-distribution regime where prioritization matters. That said, this is a real threat, and we settle it empirically rather than rhetorically: the SLMGAE baseline is given the same data, the same splits, and the same tuning budget, and if a retrained SLMGAE with a ranking-calibrated loss matches our retrieval performance, the honest conclusion is that the fix belongs in the model, not in a decision layer — though it would still not provide the auditable, per-criterion weight trail that clinicians require, which is a property no end-to-end embedding offers.
Why not just drop the experts and use pure entropy or coefficient-of-variation weighting? Because entropy measures informational contrast, not therapeutic importance. A noisy passenger gene with high expression variance across tumors would earn maximal objective weight while a tightly regulated master epigenetic modifier with low variance — G9a, or the KMT2D–YBX1 axis of KMT2D‐mediated H3K4me1 recruits YBX1 — would be smoothed toward zero, and no purely objective channel encodes "this compound class has a tolerable safety profile," which is exactly the knowledge that killed TX-1918's candidacy in the eEF2K work. The pure-entropy ablation is run precisely to quantify this failure mode; if it retrieves validated targets as well as the fusion, the expert layer adds nothing measurable and we would say so.
Why not just leave it to tumor boards, as clinical practice already does? Tumor boards handle patient-level nuance no tabular model captures, and nothing here replaces them. But the evidence that unstructured consensus carries systematic bias is direct, not hypothetical: Nie et al. showed panels pass consistency checks while remaining anchored to wrong priors; Wang documented a 33% expert over-weight corrected by entropy; Đjekić et al. showed Delphi dynamics standardize shared bias. Our simulated-Delphi condition is designed to reproduce this in the target-ranking setting — if it does, the status quo has a measured defect; if it does not, that is a genuine negative result worth reporting and would weaken the case for the whole program.
Proposed Experiments and Decisive Checks
The main experiment is the masked-target retrieval study on the SynLethDB 2.0 / BioGRID / DepMap setup, using the benchmark paper's exact split structure (CV1 through CV3) and imbalance ratios from 1:1 to 1:50, with metrics recall@20, recall@50, NDCG, and the rank of each masked target. The decisive checks are four. First, retrieval: does the calibrated fusion beat every baseline and both ablation halves at CV3, the setting that matters for a genuinely new target? Second, subtype-conditioning: does eEF2K surface only for claudin-low-like tumors (where its AKT correlation and signature hazard ratio were established) and not for other subtypes — a global-ranking failure would show the conditioning is doing real work? Third, the bias audit: does the entropy layer reduce rank churn caused by inter-expert disagreement (Kendall tau across bootstrap resamples of the expert panel), and by what factor relative to uncorrected AHP? Fourth, stability: under bounded perturbations of the fusion operator (multiplicative versus the $\alpha$-linear blend the literature uses ad hoc) and of the smoothing constant, how much do the top-20 ranks move — reported as a worst-case rank displacement, addressing the file's open problem that existing fusion operators come with no sensitivity bounds. A secondary, conditional experiment applies the same weighting to the GSRCC proteomic candidate pool to test whether PFAS lands in the top ranks without its neratinib evidence, contingent on cohort access. If the fusion fails to beat both of its own halves, or fails the subtype-conditioning check, the proposal's core claim is falsified and the honest output is a careful negative result about where de-biasing machinery does and does not transfer from engineering risk assessment to molecular target prioritization.
References
- Benchmarking machine learning methods for synthetic lethality prediction in cancer
- The Research of Electronic Banking Risk Evaluation Based on Comprehensive Assessment AHP-Entropy
- Operation Risk Assessment of Urban Dense Cable Passageway Based on Fuzzy-Analytic Hierarchy Process
- Scientific Challenges in Performing Life-Cycle Assessment in the Food Supply Chain
- A comprehensive proteomic analysis uncovers novel molecular subtypes of gastric signet ring cell carcinoma: Identification of potential prognostic biomarkers and therapeutic targets
- Combined deletion of <scp>P</scp> ten and p53 in mammary epithelium accelerates triple‐negative breast cancer with dependency on e <scp>EF</scp> 2 <scp>K</scp>
- Targeting G9a Exerts Pleiotropic Suppression in Triple-Negative Breast Cancer Cells: Cooperatively Inducing Pyroptosis and Apoptosis
- MOGAT: A Multi-Omics Integration Framework Using Graph Attention Networks for Cancer Subtype Prediction
- AET-net: A framework for subtype classification based on the multi-omics data of breast cancer
- Exportin-T Promotes Breast Cancer Progression via PI3K/AKT/mTOR Signaling Pathway
- Epigenetic derepression converts PPARγ into a druggable target in triple-negative and endocrine-resistant breast cancers
- Optimization Method of Calibration Cycle Based on State Evaluation Results of Electric Energy Meters
- KMT2D‐mediated H3K4me1 recruits YBX1 to facilitate triple‐negative breast cancer progression through epigenetic activation of c‐Myc