CROSS-REFERENCES
CROSS-REFERENCE TO RELATED APPLICATIONS
This application claims the benefit of and priority to U.S. Provisional Patent Applications No. 63/967,576, No. 63/974,083, No. 63/974,099, No. 63/988,460, No. 63/988,475, No. 63/988,480, No. 63/991,254, and No. 63/991,263. All prior provisional applications are incorporated herein by reference in their entireties.
I.FIELD OF THE INVENTION
[0001]The present disclosure relates generally to artificial intelligence, computational biology, and automated laboratory systems. More specifically, the invention relates to machine learning systems and methods for predicting drug combination synergy using structured latent representations, symmetric bilinear neural network architectures, and automated diagnostic control loops to execute closed-loop active learning of downstream physical assay systems.
II.BACKGROUND OF THE INVENTION
1.Limitations of Standard Synergy Prediction
[0002]The prediction of synergistic drug combinations is a critical challenge in personalized oncology and pharmaceutical development. Standard deep learning approaches for synergy prediction typically concatenate drug structural descriptors (such as chemical fingerprints or graph representations) with basal cell-line gene expression data, passing the concatenated vector through a multi-layer perceptron to regress a synergy score. However, these conventional architectures suffer from severe technical limitations.
[0003]First, they predominantly operate by “identity memorization” rather than learning underlying biological mechanisms. Because neural networks easily memorize the chemical fingerprints of frequently synergistic drugs, they fail to generalize to novel drugs or cold-start scenarios. Second, standard concatenation-based neural networks are not inherently order-invariant — Drug A + Drug B yields a different prediction than Drug B + Drug A unless artificially augmented, leading to computational inefficiencies and representational inconsistencies. Third, prior art methods fail to isolate the mechanistic pathway dependencies of the drugs.
III.SUMMARY OF THE INVENTION
[0004]The invention provides a system that: (1) computes monotherapy-derived pathway sensitivity profiles by correlating drug response (IC50) with structured latent pathway activations across cell lines, producing a per-drug 50-dimensional pathway sensitivity vector grounded in biological mechanism rather than chemical identity; (2) predicts combination efficacy using a symmetric bilinear interaction architecture that is mathematically guaranteed to be order-invariant (predicting the same synergy score regardless of drug input order) and uses canonical indexing to halve computational memory writes; (3) integrates an automated laboratory control loop with a fail-closed state machine and deterministic plate-layout compilation for physically testing mechanism-driven combination hypotheses; and (4) enforces an architecture-level prohibition on chemical descriptor inputs, structurally preventing identity memorization.
[0005]The system achieves zero-shot combination prediction — accurately predicting synergistic drug pairs from monotherapy data alone, without requiring any combination training data. Validated on 1,209 GDSC drug pairs, the method achieves Spearman ρ=0.800 between predicted and actual combination response patterns. The pathway orthogonality principle correctly identifies that drugs targeting different pathways show higher predicted synergy than drugs targeting the same pathway, validated by leave-target-family-out evaluation (ρ=0.689).
IV.CLAIMS
20 claims covering monotherapy-derived pathway sensitivity profiling, symmetric bilinear interaction architecture, order-invariant synergy prediction, zero-shot combination discovery, automated laboratory control loop integration, fail-closed assay state machine, canonical indexing for computational efficiency, and architecture-level prohibition of chemical descriptor inputs.
Full specification available upon request. Contact us for the complete patent application document.