Design-Time Selectivity: Fragment Assembly Against Metabolic Immunosuppression, Cell Type by Cell Type

Drugs that attack a tumor's metabolism risk killing the immune cells that are supposed to fight the tumor, because cytotoxic T cells and pro-inflammatory macrophages run on the same fuel — glucose and aerobic glycolysis — that the tumor uses to build its immunosuppressive shield. The literature maps this dilemma precisely: tumor-derived lactate rewrites immune-checkpoint gene expression through a histone modification called lactylation (Histone lactylation-driven B7-H3 expression promotes tumor immune evasion), and the metabolite methylthioadenosine pushes macrophages into a suppressive state via the adenosine A2B receptor (MTAP loss correlates with an immunosuppressive profile in GBM and its substrate MTA stimulates alternative macrophage polarization) — yet every existing chemical tool for these axes, from sodium oxamate to 2-deoxyglucose, is non-selective and would starve the very T cells it is meant to rescue (Metabolic gatekeepers: harnessing tumor-derived metabolites to optimize T cell-based immunotherapy efficacy in the tumor microenvironment).

This proposal moves selectivity from a late-stage toxicity filter into the molecular generation objective itself. The plan is to build a fragment-based generative model — a model that assembles molecules from chemically valid building blocks rather than atom by atom, so that valency errors are impossible by construction (Junction Tree Variational Autoencoder for Molecular Graph Generation) — that is conditioned on single-cell perturbation responses and trained with a two-channel reward: strong predicted perturbation of tumor and suppressive-myeloid cells, weak predicted perturbation of cytotoxic T cells. The enabling observation is that the data and the machinery both exist separately but have never been joined. The harmonized single-cell perturbation atlas scPerturb provides 44 datasets spanning CRISPR knockouts and small molecules across more than 2.5 million cells, together with a rigorous statistic (the E-distance, which measures how far a perturbed cell population sits from its unperturbed control) for quantifying per-cell-type responses (scPerturb); and fragment-level discrete flow matching — a generative method that gradually turns random noise on a coarse fragment graph into structured molecules — already generates valid, diverse molecules efficiently (FragFM: Hierarchical Framework for Efficient Molecule Generation via Fragment-Level Discrete Flow Matching). Nobody has connected them.

Core Claim

A fragment-level generative model conditioned on cell-type-resolved perturbation signatures can design molecules that suppress the metabolic programs driving immunosuppression (macrophage M2 polarization, lactate-driven B7-H3 upregulation, CCL2-mediated myeloid recruitment) while explicitly sparing the glycolytic fitness of CD8+ T cells — and this selectivity can be measured at design time, before any compound is synthesized. The claim is not merely that single-cell data is better conditioning than bulk data; it is that the structure of single-cell data — separate response vectors per cell type within one experiment — makes the tumor-versus-T-cell objective expressible at all, which bulk signatures like those in LINCS L1000 cannot do, since each bulk profile is an average over one homogeneous cell line.

Concretely, the model learns a generator $p(m \mid s_{\text{tumor}}, s_{\text{T-cell}})$ over fragment assemblies, and molecules are ranked by a selectivity score

$$S(m) = d_{\text{tumor}}(m) - \lambda \, d_{\text{T-cell}}(m),$$

where $d_c(m)$ is the predicted E-distance between the perturbed and unperturbed single-cell profile of cell type $c$ under compound $m$, and $\lambda$ is a tunable penalty weight on harming T cells. Because the E-distance framework from scPerturb already tells us the statistical resolution floor — a minimum of 300 cells per perturbation and about 1,000 UMI counts per cell to distinguish signal from noise (scPerturb) — the reward model can be trained with honest uncertainty rather than noise-fitting.

Research Program

The program has three parts. First, build the conditioning corpus: align scPerturb's perturbation-response datasets with cell-type annotations, splitting each experiment into tumor-cell, myeloid, and lymphocyte response channels; where scPerturb lacks tumor-microenvironment composition, supplement with single-cell tumor repositories of the kind used by Pan et al. to resolve macrophage and T-cell states in ovarian cancer (The MYBL2–CCL2 axis promotes tumor progression and resistance to anti-PD-1 therapy in ovarian cancer by inducing immunosuppressive macrophages). Second, build the generator: a FragFM-style discrete flow matcher over BRICS fragment graphs, using its coarse-to-fine autoencoder and stochastic fragment bag so the vocabulary scales without computational bottlenecks (FragFM), but replacing FragFM's unconditional noise source with cell-type-tokenized conditioning embeddings, following the precedent that GPT-style fragment decoders can be steered by biological embeddings — GGIFragGPT showed that Geneformer-derived embeddings of gene-interaction profiles can drive fragment assembly toward known inhibitors of CDK7, EGFR, and mTOR while keeping 100% validity (Transcriptome-conditioned molecule generation via gene interaction-aware fragment modeling with a GPT-based architecture). Third, build the two-channel reward: a model that maps a molecular structure (via Morgan fingerprints and a graph encoder) to predicted per-cell-type E-distances, trained on held-out scPerturb perturbations, and used either as a ranking filter over generated samples or as guidance during sampling. A scaffold-optimization variant following CLaSMO — conditional VAE latent space with Gaussian-process Bayesian optimization under an explicit evaluation budget — will be run in parallel for the low-budget regime, since real biological assays of immune phenotypes are exactly the expensive, multi-parametric evaluations CLaSMO was built to economize (Conditional Latent Space Molecular Scaffold Optimization for Accelerated Molecular Design).

The proposal deliberately targets the two best-mapped axes as test cases: selective A2B-receptor antagonism for MTAP-deleted tumors, where Hansen et al. showed low MTAP correlates with elevated M2 macrophages (p = 0.0064) and reduced CD4+ T cells (p < 0.03) in 385 TCGA glioblastoma samples, and where A2B blockade reverses polarization even after it is established but existing tool compounds were never optimized for brain penetration or selectivity over A1, A2A, and A3 receptors (MTAP loss correlates with an immunosuppressive profile in GBM); and CDK2-pathway inhibitors that shut down the MYBL2–CCL2 axis without the pan-CDK cytotoxicity that Pan et al. flag, since MYBL2 itself has no ligandable pocket and CDK2 inhibition with CVT-313 is the proven upstream entry point (The MYBL2–CCL2 axis). This is where the fragment approach earns its keep: the required selectivity is a fine-grained structure–activity problem across a receptor family, exactly what scaffold-hopping over pre-validated fragments does well, and exactly what generic metabolic poisons cannot do.

Baselines

The naive reference point is blanket metabolic inhibition: sodium oxamate for LDHA and 2-deoxyglucose for hexokinase, the compounds actually used in the lactate–B7-H3 work and the macrophage-polarization literature (Histone lactylation-driven B7-H3 expression; Metabolic Reprogramming Induces Macrophage Polarization in the Tumor Microenvironment). These serve as the "do the obvious thing" comparator in every in silico efficacy check; no reported result found on Lacuna for a quantitative selectivity metric for either compound, which is itself part of the gap. Beating this baseline means generating candidates whose predicted T-cell E-distance is materially below that of these non-selective inhibitors at matched predicted tumor-cell E-distance.

The strongest existing method is GGIFragGPT's bulk-transcriptome-conditioned fragment generation (Transcriptome-conditioned molecule generation), which is the only published system that conditions fragment assembly on a biological perturbation signature and recovers known inhibitors; it reports 100% validity and a uniqueness score of 0.860 on its transcriptome-conditioned task (versus 0.011 for the atom-level baseline TransGEM). It will be run as a direct competitor: same fragment vocabulary (BRICS), same training compounds, same compute and tuning budget, conditioned on the bulk-aggregated version of our single-cell signatures, with its outputs scored by our selectivity metric S(m). Beating it means achieving higher selectivity at comparable validity, uniqueness, and QED — in particular, producing molecules whose predicted T-cell channel is flat while the tumor channel moves, which bulk conditioning cannot even represent. A second strong competitor is FragFM itself, retrained with the dual-objective reward as a post-hoc filter over unconditional samples; it reports a Fréchet ChemNet Distance of 0.58 on MOSES (against 1.19 for the atom-level diffusion model DiGress) (FragFM), so any conditioning we add must not degrade distributional quality below that standard. Finally, the key ablation is our own model with the T-cell penalty removed ($\lambda = 0$): same architecture, same data, same budget, tumor-channel reward only. If the $\lambda > 0$ model does not achieve better T-cell sparing at equal tumor-channel efficacy than this ablation, the central claim of design-time selectivity is false and the simpler single-objective model should be preferred.

Why not just…?

Why not just generate molecules atom-by-atom with a standard graph or SMILES model? The evidence against this is quantitative and repeated. Atom-level construction produces illegal valencies and broken aromatic topologies because a single misplaced atom early in generation compounds downstream (Deep Learning Methods for Small Molecule Drug Discovery: A Survey; Classifier-free graph diffusion for molecular property targeting), whereas fragment vocabularies guarantee validity by construction (Junction Tree Variational Autoencoder for Molecular Graph Generation). On MOSES the atom-level DiGress scores an FCD of 1.19 against FragFM's 0.58, and on natural products FragFM samples 5x faster (FragFM). For a search over a fine selectivity frontier, where thousands of candidate structures must be scored, the 5x sampling cost and the validity burden are decisive, not cosmetic.

Why not just run a high-throughput screen or a CRISPR knockout screen? CRISPR screens are not actually a competitor here — they are our training data; scPerturb's CRISPR perturbations, including the Replogle collection with roughly 10,000 perturbations across more than 2.5 million cells, are what teach the reward model which structures move which cell types (scPerturb). The screening alternative fails on budget and on targetability: biological evaluation is expensive enough that CLaSMO was explicitly designed for a 100-evaluation budget and beat Graph-GA and MARS on 13 of 20 tasks under it (CLaSMO), and the central target MYBL2 has no ligand-binding pocket to screen against, requiring upstream CDK2 design instead (The MYBL2–CCL2 axis).

Why not just use the existing generic metabolic inhibitors (2-DG, sodium oxamate)? This is the real scientific alternative, and the literature is blunt about its failure mode: M1 macrophages and CD8+ T cells depend on the same glycolytic machinery the inhibitors shut down, so blanket inhibition disarms the immune response along with the tumor (Metabolic gatekeepers; Metabolic Reprogramming Induces Macrophage Polarization in the Tumor Microenvironment), and sodium oxamate in particular is a millimolar, non-drug-like probe rather than a therapeutic starting point (Histone lactylation-driven B7-H3 expression). Our decisive experiment addresses it head-on: in the predicted-response evaluation, 2-DG and oxamate will be scored with the same S(m) metric; if they match our generated candidates on T-cell sparing, the entire premise of selective design collapses, and we would report that.

Why not just give anti-PD-1 or anti-PD-L1 antibodies? Because the metabolic shield is upstream of, and resistant to, checkpoint blockade. Anti-PD-1 monotherapy response rates in ovarian cancer run between only 6% and 22% (The MYBL2–CCL2 axis), and phase III trials in glioblastoma showed no survival benefit for anti-PD-1 monotherapy, with myeloid-driven immunosuppression and the MTAP/MTA/A2B axis implicated (Next-generation immunotherapy biologics for glioblastoma). The same papers show that combining metabolic or pathway blockade with anti-PD-1 suppresses tumor volume where monotherapy does not — which is exactly the combination our designed small molecules are meant to enable.

Why not just fine-tune GGIFragGPT on single-cell data? This is the closest real threat, and it may partially work — if it does, we will say so. The honest experiment is included as the baseline above: GGIFragGPT conditioned on bulk-aggregated versions of our signatures, run under identical budget. Our prediction, grounded in its design, is that it cannot express the two-channel objective because its conditioning is a single signature vector per prompt, so its selectivity curve will track its training distribution rather than actively minimizing T-cell perturbation; the decisive check is whether the fine-tuned bulk model's T-cell channel degrades as its tumor channel improves, while ours does not. If both behave identically, the contribution reduces to a data swap, and we would report the single-cell conditioning as inessential and the selectivity objective as the sole novelty.

Proposed Experiments and Decisive Checks

The first experiment is retrospective and cheap: hold out entire perturbations from scPerturb (respecting the 300-cells-per-perturbation and 1,000-UMI-per-cell resolution floor), train the reward model on the rest, and test whether predicted per-cell-type E-distances on held-out perturbations beat two nulls — a structure-free predictor using only cell-type means, and a single-channel predictor blind to cell type. Success is calibrated prediction with per-cell-type resolution, measured by rank correlation between predicted and true E-distances on held-out compounds. The second experiment tests generation: condition the flow-matching generator on target signatures (M2-to-M1 macrophage shift markers Arg1 and Vegfa downward, CCL2 downward, B7-H3 downward, T-cell glycolysis and interferon-gamma markers unchanged) and evaluate the output on four axes — chemical quality on MOSES metrics (target: validity at 100%, FCD at or below FragFM's reported 0.58, uniqueness at or above GGIFragGPT's reported 0.860), selectivity S(m) against the 2-DG/oxamate reference points, recovery of known active chemotypes for the A2B receptor and CDK2 (validated against the Therapeutic Target Database as GGIFragGPT did, though no reported number for TTD-based recovery exists on Lacuna), and novelty outside the training set's actives. The third experiment is the low-budget optimization run: the CLaSMO-style latent-space Bayesian optimization variant under a strict 100-evaluation budget with the same dual objective, benchmarked against CLaSMO's protocol, which outperformed Graph-GA and MARS on 13 of 20 tasks at that budget (CLaSMO). The fourth and most decisive check is prospective and modest in scope: synthesize or source the top-ranked A2B-receptor antagonist candidates and test them in the macrophage polarization assay system established by Hansen et al., where MTA drives an Adora2b/STAT3-dependent M2d-like state reversible by A2B antagonism (MTAP loss correlates with an immunosuppressive profile in GBM) — but now scored in co-culture with CD8+ T cells, so that both channels of the design objective are measured in the same well. A candidate that depolarizes macrophages while leaving T-cell interferon-gamma output intact, where sodium oxamate does both damage, is the result that would establish design-time selectivity as a real capability rather than an in silico aspiration.