COLONYVAQ™, a Quantum-Classical Guided Personalized Neoantigen Vaccine for MSS Stage III Colorectal Cancer

Colorectal cancer is a leading cause of cancer-related mortality. In stage III disease, recurrence remains frequent despite curative-intent surgery and adjuvant oxaliplatin-based chemotherapy. Immune checkpoint inhibitors have transformed outcomes in mismatch repair-deficient / microsatellite instability-high colorectal cancer, but microsatellite-stable / pMMR tumors typically exhibit a lower tumor mutational burden and a poorly inflamed, immunosuppressive microenvironment. As a result, conventional PD-1 blockade alone provides minimal benefit in MSS/pMMR disease. Earlier vaccine approaches in colorectal cancer focused on tumor-associated antigens such as CEA, MUC1, survivin, MAGE and multi-TAA peptide cocktails. These studies showed that peptide and dendritic-cell-based vaccines can induce antigen-specific T-cell and B-cell responses, yet objective responses were rare, clinical benefit was modest, and off-tumor toxicities were a concern because TAAs are frequently expressed in normal tissues. Tumor-specific neoantigens, generated by non-synonymous somatic mutations, are in contrast restricted to malignant cells, escape central tolerance, can elicit higher-avidity T-cell responses, and minimize off-tumor toxicity. Early colorectal and pan-cancer neoantigen trials, as well as shared-neoantigen programs such as SLATE-KRAS and fully personalized viral-RNA platforms such as GRANITE, have demonstrated that multi-neoantigen vaccination is feasible, safe, and immunogenic, particularly in low-burden or maintenance settings and when combined with checkpoint blockade. The adjuvant neoantigen dendritic cell vaccine plus nivolumab trial in resected hepatocellular carcinoma and colorectal liver metastases further supports the idea that personalized neoantigen vaccination in the minimal residual disease (MRD) setting can augment neoantigen-specific T-cell responses and potentially improve relapse-free survival. Oxaliplatin-based regimens (mFOLFOX6 or CAPOX) can induce immunogenic cell death, exposing calreticulin and other danger signals that enhance dendritic cell uptake and cross-presentation of tumor antigens. Nivolumab, by blocking PD-1, relieves inhibitory signaling on activated T cells. Combining a personalized multi-neoantigen peptide vaccine with immunogenic chemotherapy and PD-1 blockade is therefore expected to increase antigen release, improve antigen presentation, and augment effector function, potentially converting immunologically "cold" MSS tumors into more inflamed, "hot" lesions amenable to durable immune surveillance in the adjuvant setting.

COLONYVAQ-CRC: Quantum-Classical, Physics-Aware Neoantigen Prioritization Most existing neoantigen pipelines treat epitope ranking as mainly statistical. COLONYVAQ-CRC introduces a physics-aware, quantum-classical AI layer, adapted from Tamavaq, to generate an auditable, mechanistic chain from sequencing to clinical peptide selection. For each candidate peptide-HLA pair p, the system constructs a unified feature representation Φ(p), which concatenates sequence-based, biological, quantum, structural, and energetic evidence: Φ(p)=[e_"CNN" (p),"" aux(p),"" z_Q (p),"" ϕ_"struct" (p),""ϕ_"dock" (p)]. The term e_"CNN" (p) denotes a deep sequence/HLA embedding derived from convolutional or transformer models trained on large immunopeptidome datasets. The auxiliary block aux(p) compiles antigen processing and expression priors such as proteasomal cleavage likelihood, TAP transport propensity, transcript abundance, clonality and, when available, ctDNA/MRD information to approximate the effective antigen source strength. The quantum descriptor z_Q (p) is a low-dimensional classical vector that parameterizes a quantum circuit embedding. The structural term ϕ_"struct" (p) summarizes pocket occupancy and residue-residue contacts in modeled peptide-HLA complexes. Finally, ϕ_"dock" (p) aggregates docking ensemble statistics including pose energies, dispersion and conformational diversity.

Similarity between two candidates p and q is captured by a composite positive semi-definite kernel K_"total" (p,q)=αK_"CNN" (p,q)+βK_"aux" (p,q)+γK_Q (p,q)+δK_"struct" (p,q)+εK_"dock" (p,q), where the non-negative weights α,β,γ,δ,ε adjust the relative contribution of each modality. Because each component kernel is constructed to be positive semi-definite, their non-negative linear combination remains positive semi-definite, ensuring that K_"total" can be used consistently in kernel logistic regression or related methods. A decision function can be written as f(p)=∑_(i=1)^M▒α_i K_"total" (p,p_i)+b, where {p_i } are training peptides and α_i,b are learned coefficients. The immunogenicity probability is then modeled as I ̂(p)=σ(f(p)), where σ(z)=1/(1+e^(-z)) is the logistic function. On the quantum side, each peptide x is encoded as a normalized state ∣ψ(x)⟩ in a Hilbert space H of dimension 2^n, constructed via a feature map U(z_Q (x),θ) acting on a reference state ∣0⟩^(⊗n): ∣ψ(x)⟩=U(z_Q (x),θ)" "∣0⟩^(⊗n). The overlap between two peptide states is ⟨ψ(x)∣ψ(y)⟩. Quantum-geometric similarity is quantified by the Fubini-Study distance d_"FS" (x,y)=arccos⁡(∣⟨ψ(x)∣ψ(y)⟩∣), which lies in [0ⓜ,π/2], where d_"FS" =0 corresponds to identical rays and d_"FS" =π/2 to orthogonal states. From this distance, a quantum similarity kernel is defined as K_q (x,y)=〖∣⟨ψ(x)∣ψ(y)⟩∣〗^2=〖cos⁡〗^2 (d_"FS" (x,y)). This kernel can be interpreted as the probability that the state ∣ψ(x)⟩ is projected onto ∣ψ(y)⟩. When low-sequence-identity peptides share higher-order physicochemical structure, they may map to nearby points on this complex projective manifold, generating large K_q values even when classical sequence similarity is low. The internal structure and entanglement of ∣ψ(x)⟩ are monitored by forming reduced density matrices on subsystems. For a bipartition into subsystems A and B, the reduced state is ρ_A (x)=Tr_B (∣ψ(x)⟩⟨ψ(x)∣). The von Neumann entropy S_A (x)=-Tr[ρ_A (x)log⁡ρ_A (x)] quantifies entanglement between A and B. A regularization term encourages entropy within a target range, avoiding trivial product states (too little entanglement) and excessively entangled states that can be numerically unstable and difficult to approximate on noisy intermediate-scale quantum (NISQ) hardware.

The sensitivity of the quantum embedding to parameter changes is characterized by the quantum Fisher information matrix F(θ) with entries F_ij (θ)=R[⟨∂_i ψ∣∂_j ψ⟩-⟨∂_i ψ∣ψ⟩⟨ψ∣∂_j ψ⟩], where ∣∂_i ψ⟩=∂∣ψ(θ)⟩/∂θ_i. Ill-conditioned Fisher matrices, with very small eigenvalues, can lead to large variances in parameter estimates and unstable kernel values. COLONYVAQ therefore introduces a penalty proportional to tr(F(θ)^(-1)), which diverges when eigenvalues approach zero; minimizing this term nudges optimization toward parameter regions where all directions in parameter space are well informed by the data. Energetics are treated in a thermodynamically calibrated way. For each peptide-HLA docking pose i with standard free energy ΔG_i^∘, the microstate association and dissociation constants are K_(a,i)=exp⁡(ⓜ (ΔG_i^∘)/RT),K_(d,i)=exp⁡((ΔG_i^∘)/RT), with R=1.987×10^(-3) " " kcal⋅mol^(-1)⋅K^(-1) and T=310"" K, such that RT≈0.616" " kcal⋅mol^(-1). The docking ensemble is summarized as a Boltzmann-weighted effective association constant K_a^"eff" =∑_i▒w_i exp⁡(ⓜ-(ΔG_i^∘)/RT),∑_i▒w_i =1, yielding an effective free energy ΔG_"eff" ^∘=-RTln⁡K_a^"eff" and corresponding dissociation constant K_d^"eff" =1/K_a^"eff" . These values are reported in units familiar to experimentalists (kcal·mol-¹ for ΔG_"eff" ^∘, nM for K_d^"eff" ). If the spread of free energies in the ensemble is σ_ΔG, then the associated uncertainty in K_dcan be expressed multiplicatively as exp⁡(±σ_ΔG/(RT)). For example, at T=310"" K, a change of 1.2"" kcal⋅mol^(-1) in ΔG^∘ changes K_d by a factor of approximately exp⁡(1.2/0.616)≈6.3. A docking loss term L_"dock" =λ_1 E ˉ+λ_2 σ_E, where E ˉ and σ_E are the mean and standard deviation of docking energies, biases the model toward low-energy, low-variance ensembles that are empirically associated with robust peptide-MHC display. On top of Φ(p) and the kernel K_"total" , COLONYVAQ trains a calibrated logistic head I ̂(p)=σ(w^⊤ Φ(p)+b), which is interpreted as the probability that peptide p is recognized by T cells, optimized for both discrimination (for example AUC) and calibration (for example Brier score, expected calibration error). In parallel, a linear thermodynamic head predicts (ΔG) ̂^∘ (p)=η^⊤ Φ(p)+η_0, from which a predicted dissociation constant K ̂_d (p)=exp⁡((ΔG) ̂^∘ (p)/(RT)) is derived. The total loss couples prediction, structure, docking and quantum Fisher regularization into a single objective L=L_"pred" +L_"struct" +L_"dock" +L_"QFIM", with L_"QFIM" proportional to tr(F(θ)^(-1)).

Candidate peptides are passed through a three-gate "physics + geometry + immunology" oracle. First, a quantum-geometric gate requires that the Fubini-Study distance between ∣ψ(x)⟩ and a centroid ∣ψ(P)⟩ of empirically validated immunogenic peptides satisfy d_"FS" (ψ(x),ψ(P))≤d^"\*" . Second, a thermodynamic gate requires that the effective free energy and dissociation constant meet minimum binding strength criteria, ΔG_"eff" ^∘ (x)≤ΔG^"\*" or equivalently K_d^"eff" (x)≤K_d^"\*" . Third, an immunogenicity gate requires that the calibrated probability exceed a threshold, I ̂(x)≥I^"\*" . Let the total number of candidates be N, with M peptides passing all three filters. In abstract quantum terms, a uniform superposition over all candidates is ∣Ψ_0⟩=1/√N ∑_(j=1)^N▒〖∣j⟩,which can be decomposed into "marked" and "unmarked" subspaces as ∣Ψ_0⟩=sin⁡θ" "∣Ψ_"good" ⟩+cos⁡θ" "∣Ψ_"bad" ⟩, where 〖sin⁡〗^2 θ=M/N. A Grover-like amplitude amplification operator G is defined as the product of an oracle O that flips the phase of marked states and a diffusion operator D that reflects about ∣Ψ_0⟩. After r iterations, the state becomes ∣Ψ_r⟩=G^r∣Ψ_0⟩=sin⁡((2r+1)θ)∣Ψ_"good" ⟩+cos⁡((2r+1)θ)∣Ψ_"bad" ⟩, and the probability of measuring a marked index is P_r=〖sin⁡〗^2 ((2r+1)θ).

When θ is small (few good candidates), the optimal number of iterations that maximizes P_r is approximately r_"opt" ≈π/4 √(N/M), but in the NISQ regime and in the presence of uncertainty in M, COLONYVAQ uses a small number of iterations (typically one to three) to reliably amplify the weight of marked peptides without over-rotation. In practice, this Grover-style step is simulated or approximated in a manner compatible with available hardware and serves to focus GMP peptide synthesis on a compact, high-confidence subset.

Within the set of marked peptides, residual ties are broken using a composite score S(x)=αK_q (x,P)+β" " σ" ⁣" ((I ̂(x)-I^"\*" )/τ_I )+γ" " σ" ⁣" ((K_d^"\*" -K_d^"eff" (x))/τ_K ), where K_q (x,P)=〖∣⟨ψ(x)∣ψ(P)⟩∣〗^2, τ_I and τ_K set the steepness of transitions, and σ is the logistic function. This expression makes explicit how similarity to known positive controls, modeled immune potency, and predicted binding strength jointly determine the final ranking used to define the COLONYVAQ-CRC peptide cargo for each patient.

Rationale for Combination with mFOLFOX6 or CAPOX and Nivolumab

Oxaliplatin and fluoropyrimidines are standard components of adjuvant therapy in stage III colorectal cancer and can induce immunogenic cell death, increasing the release of tumor antigens and danger signals, which in turn enhances dendritic cell activation and antigen cross-presentation. Nivolumab 3 mg/kg every 2 weeks blocks PD-1, preventing exhaustion and functional suppression of vaccine-induced and chemotherapy-released tumor-specific T cells. The quantum-classical COLONYVAQ-CRC engine is intended to maximize the quality of neoantigen targets; immunogenic chemotherapy increases antigen availability; and PD-1 blockade sustains T-cell effector function. The early phase I trial will test the safety and feasibility of this three-component strategy in the adjuvant MRD setting and generate preliminary immune and molecular response data.