Generative Thermal Sensitizers: Designing Molecules Against a Bioheat Damage Objective

When surgeons or interventional oncologists destroy tissue with heat — radiofrequency probes, lasers, focused ultrasound — the energy that kills the tumor also diffuses into neighboring structures. Above roughly 50–60°C, this "heat-affected zone" reliably damages healthy tissue: nerves during prostate surgery, bladder-tumor margins destined for pathology, normal liver around a hepatocellular carcinoma. One promising way out is biological rather than physical. Cells defend themselves against sub-lethal heat with chaperone proteins — chiefly Heat Shock Protein 70 and 90 (HSP70/HSP90) — that refold damaged proteins and block cell death. If a drug disables these chaperones first, tumor cells die at a mild 43–45°C, a range at which adjacent healthy tissue is spared, as demonstrated in mice with gambogic acid (an Hsp90 inhibitor) at ~43°C Rapid synthesis of a Bi@ZIF-8 composite nanomaterial as a near-infrared-II (NIR-II) photothermal agent for the low-temperature photothermal therapy of hepatocellular carcinoma and with a copper nanozyme that degrades HSP70 Visualization of HSP70-regulated mild-photothermal therapy for synergistic tumor treatment: a precise space-time mild-temperature photothermal ablation strategy.

The molecules doing this today are poor drugs: a toxic, poorly soluble natural product that must be stuffed into a metal-organic framework to work, or an all-inorganic nanoparticle with clearance and reproducibility problems. Meanwhile, generative molecular design — neural networks that sample new molecular structures directly, rather than searching fixed libraries — has never been aimed at this problem with modern tools. The one prior attempt used a 2017-era text-based adversarial generator to propose Hsp70 binders scored only by docking and solubility De Novo Design of Hsp70 Inhibitors with Objective Reinforced Generative Adversarial Networks (ORGAN) and Docking, with no constraint that the molecules can actually be synthesized and no connection to the physics of heat damage.

The core claim of this proposal is that the right objective for such molecules is not a static docking score but a thermal margin: the temperature gap between "tumor cells die" and "adjacent healthy tissue is injured." We propose to build a molecule generator whose reward function is computed by a bioheat simulation — a Pennes heat-diffusion model with an Arrhenius damage integral whose parameters shift with predicted chaperone inhibition — and to close the loop with wet-lab validation in an established mouse hyperthermia model.

Core Claim

Existing generative design for HSP inhibitors optimizes the wrong endpoint. Docking affinity to the chaperone binding pocket tells you a molecule binds; it does not tell you by how many degrees the ablation threshold drops, or whether the dose that sensitizes the tumor is far from the dose that harms healthy tissue. The physics of that question is already written down: thermal cell kill follows the Arrhenius integral

$$\Omega(t) = \int_0^t A\, e^{-E_a / (R\,T(\tau))}\, d\tau$$

where $A$ and $E_a$ are the pre-exponential factor and activation energy of the damage process, $R$ is the gas constant, and $T(\tau)$ is the local tissue temperature over time, itself computed from the Pennes bioheat equation Development of a Computational Paradigm for Laser Treatment of Cancer. That same literature established that HSP70 kinetics — not temperature alone — determine whether cells at the edge of the heat-affected zone survive Development of a Computational Paradigm for Laser Treatment of Cancer. Our claim: if the HSP-expression and cell-death terms in this model are made functions of a candidate molecule's predicted chaperone inhibition, then the generator can be trained to maximize the sensitization margin

$$\Delta = T_{\text{injury}}^{\text{healthy}} - T_{\text{kill}}^{\text{tumor+sensitizer}},$$

the degrees of operating-temperature headroom a molecule buys, which is precisely the quantity that shrinks collateral damage. This couples the two fields at the level of the loss function, not merely the application: collateral thermal damage supplies the objective, generative molecular design supplies the search procedure, and synthesizability supplies the constraint.

Research Program

Three components, in order of dependency.

First, the thermal-margin surrogate. We will extend the validated Pennes/Arrhenius/HSP70 closed-loop modeling stack of Development of a Computational Paradigm for Laser Treatment of Cancer — which showed good agreement between finite-element predictions and real-time magnetic-resonance temperature imaging in canine prostate — with a pharmacodynamics term: chaperone inhibition level as a function of molecule properties (predicted binding affinity to the pocket, plus a simple pharmacokinetic surrogate for tumor-versus-healthy-tissue exposure). Calibrating the HSP70-inhibition-to-$E_a$-shift relationship is the key scientific unknown; we will fit it against the published in vivo dose-response landmarks: near-complete 4T1 tumor ablation at a steady 43–45°C under 1064 nm irradiation at 0.75 W/cm² when HSP70 is suppressed Visualization of HSP70-regulated mild-photothermal therapy, and complete Huh7 tumor regression over 14 days at ~43°C with a photothermal conversion efficiency of 24.4% under gambogic-acid Hsp90 inhibition Rapid synthesis of a Bi@ZIF-8 composite nanomaterial. The surrogate must run in milliseconds per molecule, so it will be a fast reduced model (radially symmetric Pennes solution, lookup-table Arrhenius), not a full finite-element solve — the literature itself flags that live 3D inversion needs high-performance clusters Development of a Computational Paradigm for Laser Treatment of Cancer].

Second, the generator. A pocket-conditioned 3D molecular generator — a diffusion model, which learns to turn random noise into molecular structures by iteratively denoising, respecting the rotational/translational symmetries of 3D space — conditioned on the Hsp70 binding site (PDB ID 4H5T, the crystal structure used by the prior ORGAN work De Novo Design of Hsp70 Inhibitors with ORGAN and Docking) and, in a second instantiation, an Hsp90 pocket. We adopt the hybrid continuous-categorical design of Navigating the Design Space of Equivariant Diffusion-Based Generative Models for De Novo 3D Molecule Generation because its categorical treatment of atom and bond identities avoids element hallucination and reaches over 95% molecular stability on GEOM-Drugs (versus 40.3% for EDM and 89.7% for MiDi) after pre-training on 95.7 million molecules from PubChem3D — a property that matters because DiffBP: generative diffusion of 3D molecules for target protein binding leaves bond orders and rings to post-hoc perception, generating exactly the synthetic-monster problem we must avoid. Two non-diffusion ingredients are grafted on: reaction-constrained sampling in the spirit of ClickGen: Directed exploration of synthesizable chemical space via modular reactions and reinforcement learning, which guarantees synthetic accessibility by assembling molecules from modular real reactions, and pharmacophore-level conditioning in the spirit of Accelerating discovery of bioactive ligands with pharmacophore-informed generative models, which conditions on interaction blueprints rather than scaffolds and is the demonstrated route to escape "me-too" chemotype bias (that model scaffold-hopped to a 5.1 nM PLK1 inhibitor, IIP0943). The reward combines the thermal margin $\Delta$ from the surrogate with synthesizability and a mild novelty pressure against known HSP inhibitors.

Third, experimental validation in the hyperthermia workflow. Top-ranked molecules undergo synthesis (guaranteed by construction under the reaction constraints), binding assays against Hsp70/Hsp90, and testing in the 4T1 murine breast-cancer model under 1064 nm irradiation at 0.75 W/cm² — the exact protocol of Visualization of HSP70-regulated mild-photothermal therapy, including its measured HSP70 kinetic timeline (photoacoustic signal peak at 10 hours post-injection, therapeutic window 10–12 hours, HSP70 reduction detectable from 2 hours post-treatment). The decisive measurement is the dose-temperature response surface: tumor kill and adjacent-tissue injury as functions of applied temperature, per molecule. This addresses the open problem the nanozyme work itself names — there is no real-time readout of when chaperone suppression is complete, so heating applied out of phase fails Visualization of HSP70-regulated mild-photothermal therapy — by testing whether small molecules with predictable pharmacokinetics give a wider, more forgiving timing window than oxidative HSP70 degradation does.

Proposed Experiments

Generation quality will be reported on the standard benchmarks so the community can compare: validity, uniqueness, novelty, and Fréchet ChemNet Distance via the GuacaMol suite, whose distribution-learning baselines include a SMILES LSTM achieving 77% acceptable compounds, matching the "Best of Dataset" baseline GuacaMol: Benchmarking Models for de Novo Molecular Design, and ChEMBL 24 filtered as that paper specifies. Pocket-conditioning quality will be reported as AutoDock Vina scores against 4H5T, directly comparable to the ORGAN paper's reported leads (Compound 464 at −8.6 kcal/mol against a reference-inhibitor S1g-10 at −5.5 kcal/mol) and as ligand efficiency in the format DiffBP reports on CrossDocked2020 (overall 41.07% versus GraphBP's 29.54%) DiffBP. The headline number, however, is new: predicted and then measured $\Delta$ — the degrees of thermal headroom — per molecule, plus the fraction of generated molecules that are both synthesizable (SA score below a fixed cutoff) and predicted to increase $\Delta$ by at least 5°C. For the in vivo arm, the reference outcomes are near-complete tumor inhibition at 43–45°C with healthy-tissue sparing Visualization of HSP70-regulated mild-photothermal therapy and complete regression over 14 days at ~43°C Rapid synthesis of a Bi@ZIF-8 composite nanomaterial; we additionally measure adjacent-tissue injury histologically, which neither predecessor quantified as a function of temperature.

Baselines

The "do nothing" reference point is conventional high-temperature ablation, which by the physics of heat diffusion guarantees a heat-affected zone in adjacent structures — the charring/insulation failure mode in radiofrequency probes Research on High Precision Temperature Acquisition System for Radiofrequency Ablation System and the thermal-artifact problem at resection margins Monopolar versus bipolar transurethral resection of bladder Tumour: post-hoc analysis of a prospective trial. We will quantify its collateral-damage margin in our surrogate and in histology to anchor the scale of $\Delta$.

The strongest existing computational competitor is the ORGAN pipeline of De Novo Design of Hsp70 Inhibitors with ORGAN and Docking: a SMILES adversarial generator pre-trained on 800,000 ZINC molecules, reinforcement-tuned on docking and logP, producing leads up to −8.6 kcal/mol against 4H5T. It is the strongest competitor because it is the only published generative system aimed at this exact target and this exact clinical rationale. We will rerun it ourselves under matched conditions — same target structure, same docking engine and scoring, same compute and tuning budget, same number of final candidates — and evaluate both pipelines on all metrics (Vina score, SA score, GuacaMol novelty, and our thermal-margin $\Delta$). Beating it requires strictly better on the shared metrics (mean Vina score of final candidates below −8.6 kcal/mol at matched SA score) and nonzero where it scores zero: a positive measured $\Delta$ in vivo, since its own authors note it had no synthesis constraints, no ADMET filtering, and no thermal-damage integration.

The strongest wet-lab comparator is gambogic acid delivered in Bi@ZIF-8, the current best-in-class small-molecule sensitization result: complete tumor disappearance in BALB/c nude mice over 14 days at ~43°C Rapid synthesis of a Bi@ZIF-8 composite nanomaterial. Our molecules beat it only if they achieve comparable tumor control at 43–45°C with a simple formulation (no metal-organic carrier), better solubility, and a wider therapeutic window — this is a stretch target and we say so.

The key ablation removes the thermal-margin reward: the identical generator conditioned on the same pocket, tuned only on docking + logP + SA (i.e., the ORGAN objective transplanted into a modern 3D generator). If this ablation matches the full model's measured $\Delta$, the bioheat objective is decorative and the proposal's claim fails; if the full model's candidates show larger measured thermal headroom at equal docking scores, the objective — not the architecture — is doing the work. This ablation also doubles as the honest test of the "just fine-tune an off-the-shelf 3D diffusion model" objection.

Why not just…?

Why not just control the energy delivery physically? Wavelength matching genuinely solves the problem where it applies: the Er:YAG laser at 2.94 µm confines the heat-affected zone to 5–15 µm by explosive surface vaporization Position paper, and closed-loop magnetic-resonance-temperature-guided control dynamically adapts to perfusion Development of a Computational Paradigm for Laser Treatment of Cancer. But both are bounded by physics: deep-tumor wavelengths like 1064 nm must penetrate and therefore diffuse, charring insulates probes and spreads damage outward Research on High Precision Temperature Acquisition System for Radiofrequency Ablation System, and when a tumor sits against a neurovascular bundle, heat reaches the nerve before the last tumor cell dies unless the tumor's death threshold is lowered. Our proposal is complementary, not competing — the surrogate's predictions are computed under the same closed-loop delivery parameters — and the decisive check is whether sensitized tumors reach full kill at temperatures where the physical-control-only arm fails.

Why not just screen existing libraries for HSP inhibitors? Virtual screening of purchasable compounds guarantees synthesizability and avoids generative-model risk. We will in fact run this as an additional baseline (docking screen of a matched ZINC subset against 4H5T) because it is a real threat: if screening finds compounds at −8 kcal/mol with good SA and predicted $\Delta$, generative search is unjustified here. The evidence against it being sufficient: synthesizable space is ~10^60 molecules Generative Models as an Emerging Paradigm in the Chemical Sciences, static libraries over-sample known chemotypes Accelerating discovery of bioactive ligands with pharmacophore-informed generative models, and the one head-to-head data point in the file shows de novo generation finding binders (−8.6 kcal/mol) well beyond the known reference inhibitor (−5.5 kcal/mol) De Novo Design of Hsp70 Inhibitors with ORGAN and Docking. The screening baseline settles it empirically rather than by argument.

Why not just repurpose or derivatize gambogic acid? It already works in mice Rapid synthesis of a Bi@ZIF-8 composite nanomaterial, so medicinal-chemistry optimization of a known scaffold is the cheap path. The counter-evidence is that it needed a metal-organic-framework carrier for bioavailability, and its own paper flags systemic off-target toxicity and poor modifyability — the classic profile that scaffold-hopping generative models exist to escape, with demonstrated reach (IIP0943, 5.1 nM, on a chemically novel backbone) Accelerating discovery of bioactive ligands with pharmacophore-informed generative models. Still, a gambogic-acid analog series belongs in our experimental matrix as a comparison arm, because if simple derivatization matches generative candidates on $\Delta$ at equal solubility, the generative machinery has not earned its cost.

Why not just use the nanozyme platform that already works? The Cu-ABTS@CCM construct achieves the full clinical goal — near-complete ablation at 43–45°C with real-time HSP70 readout Visualization of HSP70-regulated mild-photothermal therapy — so a reviewer could ask why a small-molecule program is needed at all. Our answer is the translational profile: cell-membrane-coated inorganic nanoparticles carry batch reproducibility, clearance, and immunogenicity burdens that the file itself notes, while the open problem of timing (heating must coincide with chaperone suppression) may be easier to control with a molecule whose pharmacokinetics are predictable than with oxidative degradation whose own authors flag as lacking a timing indicator Visualization of HSP70-regulated mild-photothermal therapy. This is a genuine strategic bet, not a settled point, and the in vivo arm measures the timing-window width for both approaches directly.

Decisive Checks

Three findings would falsify the core claim. If the docking-only ablation matches the thermal-margin model on measured $\Delta$ in vivo, the bioheat objective adds nothing over binding affinity. If the HSP70-inhibition-to-$E_a$-shift relationship cannot be fit consistently across the two published in vivo regimes (4T1 and Huh7), the surrogate has no predictive calibration and the whole reward is unanchored. And if the ZINC screening baseline produces candidates equal on all metrics, the generative component is gratuitous. Conversely, one result would validate the program decisively: a synthesized, generatively designed molecule that shifts the measured tumor-kill threshold below the adjacent-tissue injury threshold by several degrees in the 4T1 protocol — the first time a molecule designed by a generative model would have been evaluated inside an energy-based ablation workflow at all.