HTS-Oracle X: AI-Guided Prospective Discovery of Small Molecule Immune Checkpoint Binders

Abstract

Targeting immune checkpoint protein–protein interactions (PPIs) using small molecules remains limited by the characteristically low hit rates of conventional high-throughput screening against these interfaces. Here we report HTS-Oracle X, a multimodal deep learning platform that integrates bidirectional cross-attention fusion of ChemBERTa SMILES embeddings with extended RDKit descriptors, trains on continuous biophysical binding signals rather than binary labels, and employs Monte Carlo Dropout uncertainty quantification for uncertainty-adjusted compound selection. Trained on 45,760 Dianthus TRIC-screened compounds per target under scaffold-aware cross-validation, HTS-Oracle X was applied prospectively to a 100,160-compound enamine library against CD28, TIM-3, and VISTA. From 150 model-selected compounds, 45 dose–response confirmed binders were identified, yielding enrichment factors of 234–408× over experimentally established random prospective baselines and 16 sub-micromolar hits. The top hits, HX-CD28-1 (K _D = 233 nM), HX-TIM3-1 (K _D = 249 nM), and HX-VISTA-1 (K _D = 345 nM), demonstrated on-target functional activity in immune cell and tumor coculture assays.

Immune checkpoint proteins have emerged as validated therapeutic targets across oncology, with the approval of relatlimab/nivolumab for LAG-3/PD-1 co-blockade and ongoing clinical evaluation of anti-TIGIT antibodies establishing the translational relevance of this target class beyond PD-1. − Beyond PD-1/PD-L1, a growing number of co-stimulatory and co-inhibitory receptors have been identified as clinically actionable targets whose combined blockade or activation can overcome therapeutic resistance and broaden the scope of cancer immunotherapy. Among the next generation of immune checkpoint targets, CD28, TIM-3, and VISTA have attracted particular interest owing to their nonredundant roles in T cell activation, exhaustion, and myeloid-mediated immune suppression, respectively, yet remain underexplored by small molecule approaches relative to their clinical potential. , Nevertheless, the broad, shallow protein–protein interaction (PPI) surfaces that mediate immune checkpoint signaling impose severe constraints on small molecule affinity and selectivity, and conventional high-throughput screening (HTS) against these interfaces routinely yields low hit rates, imposing substantial resource burdens with limited chemical output. −

Our group has developed an extensive immune checkpoint small molecule discovery program spanning LAG-3, − ICOS, , TIM-3, VISTA, and CD28, using pharmacophore-based virtual screening, affinity selection mass spectrometry, and antibody pharmacophore-derived hit identification workflows. These efforts provided the chemical and biological foundation for HTS-Oracle, a multimodal deep learning platform combining ChemBERTa SMILES embeddings with RDKit descriptors that achieved up to 176-fold enrichment for TREM2 and CHI3L1, and 8.4-fold enrichment for CD28. However, the original platform employed simple feature concatenation, binary hit classification, and random cross-validation splits, limiting both predictive accuracy and the interpretability of prospective performance estimates.

Machine learning frameworks combining molecular language models with cheminformatics features have demonstrated strong bioactivity prediction performance across diverse drug discovery applications. − A critical but frequently overlooked requirement for platforms claiming prospective utility is scaffold-aware cross-validation, which enforces structural separation between training and test sets and prevents the performance inflation that arises from random splits when structurally similar compounds appear in both partitions. Beyond classification, training on continuous biophysical binding signals rather than binary labels preserves quantitative information that can improve the rank-ordering of active compounds, particularly in challenging PPI contexts where hit rates are low and graded binding information is diagnostic.

Here we introduce HTS-Oracle X, which advances our HTS-Oracle platform through three architectural innovations: bidirectional cross-attention fusion between the ChemBERTa and RDKit branches, regression on continuous Dianthus normalized fluorescence signal (ΔF _norm) binding signals, and Monte Carlo Dropout uncertainty estimation enabling uncertainty-adjusted compound selection. Trained on 45,760 biophysically screened compounds per target under scaffold-aware cross-validation and prospectively applied to a 100,160-compound enamine library, HTS-Oracle X achieves enrichment factors of 234–408× over experimentally established random prospective baselines, with 26–34% validated hit rates and sub-micromolar binders for all three targets, including a 36-fold improvement in enrichment for CD28 over the original platform.

Training data sets were established specifically for this study by our Dianthus TRIC biophysical screening of 45,760 enamine compounds per target against recombinant CD28, TIM-3, and VISTA extracellular domains (Figures S1–S3), yielding 1,734 CD28 binders (3.79%), 2,147 TIM-3 binders (4.69%), and 1,382 VISTA binders (3.02%), with continuous ΔF _norm binding signals retained as regression targets rather than binary hit/nonhit labels (Figure A). HTS-Oracle X generates molecular representations through two parallel branches: a ChemBERTa branch producing 768-dimensional CLS token embeddings from SMILES, and an extended RDKit branch encoding Morgan fingerprints (radius = 2, 2,048-bit), MACCS keys (167-bit), topological torsion fingerprints (1,024-bit), and 25 physicochemical descriptors (3,264 features total; Figure C). A bidirectional cross-attention module enables each branch to contextually attend to the other before the regression head, which predicts continuous ΔF _norm values directly from molecular structure (Figure C). Monte Carlo Dropout (10 inference passes) provides per-compound uncertainty estimates, enabling selection by Score = Predicted ΔF _norm – 0.5 × Uncertainty (Figure C). Fifteen submodels (3 feature selection methods × 5 scaffold-aware folds) are ensemble-averaged for final predictions, with ChemBERTa embeddings precomputed once across all compounds for efficiency (Figure B). The trained ensemble was applied prospectively to a 100,160-compound enamine library, with the top 50 compounds per target selected by uncertainty-adjusted Selection Score for experimental validation (Figure D).

1.

Overview of the HTS-Oracle X platform for AI-guided biophysical high-throughput screening. (A) Training data set curation: 45,760 compounds per target screened by Dianthus NT.23 Pico TRIC against CD28, TIM-3, and VISTA, yielding 1,734 (3.79%), 2,147 (4.69%), and 1,382 (3.02%) confirmed binders, respectively; continuous ΔF _norm values are retained as regression targets rather than binary hit/nonhit labels. (B) Ensemble training framework: scaffold-aware 5-fold assignment distributes compounds across five folds; three feature selection methods (LASSO, PCA, Mutual Information) are applied independently across all five folds, generating 15 submodels averaged into a robust predictive ensemble. (C) HTS-Oracle X architecture: ChemBERTa transformer embeddings (768-d CLS token) and extended RDKit features (3,264-d) are independently projected to 256-d representations and fused via bidirectional cross-attention (4-head, LayerNorm, residual connections); a regression head predicts continuous ΔF _norm binding signals, and Monte Carlo Dropout (10 passes) generates per-compound uncertainty estimates (σ) for uncertainty-adjusted Selection Score ranking (ΔF _norm – 0.5 × σ). (D) Prospective screening workflow: the 100,160-compound enamine library is scored per target and filtered to the top 50 compounds by Selection Score (>99.95% screening burden reduction), followed by two-stage experimental validation (single-dose Dianthus TRIC then Monolith X dose–response), yielding 234–408× enrichment over random prospective baselines (0.083–0.111%), 30.0% overall hit rate (45/150 confirmed), and 16 sub-μM binders.

All performance reporting employed scaffold-aware 5-fold cross-validation. HTS-Oracle X achieved ROC-AUC values of 0.914 ± 0.038 (CD28), 0.931 ± 0.032 (TIM-3), and 0.886 ± 0.049 (VISTA), with Spearman R of 0.582 ± 0.071, 0.621 ± 0.064, and 0.512 ± 0.083, respectively (Table ; Figure ). All targets exceeded the ROC-AUC threshold of 0.85 considered indicative of useful virtual screening performance.

1. Computational Performance of HTS-Oracle X under Five-Fold Cross-Validation.

Target	Training Compounds	Binders (%)	ROC-AUC (mean ± SD)	Spearman R (mean ± SD)	Avg. Precision (mean ± SD)
CD28	45,760	1,734 (3.79%)	0.914 ± 0.038	0.582 ± 0.071	0.742 ± 0.052
TIM-3	45,760	2,147 (4.69%)	0.931 ± 0.032	0.621 ± 0.064	0.778 ± 0.047
VISTA	45,760	1,382 (3.02%)	0.886 ± 0.049	0.512 ± 0.083	0.681 ± 0.061

2.

Computational performance of HTS-Oracle X under 5-fold cross-validation. (A) ROC-AUC and (B) Spearman R across CD28, TIM-3, and VISTA (mean ± SD, 5 scaffold-aware folds × 3 feature methods = 15 submodels per target). Dashed line in (A) indicates the ROC-AUC = 0.85 virtual screening utility threshold.

HTS-Oracle X was applied prospectively to 100,160 compounds from the Enamine Hit Locator Library, generating uncertainty-adjusted ΔF _norm binding score predictions independently for each immune checkpoint target. The top 50 compounds per target were selected by Selection Score (Predicted ΔF _norm – 0.5 × σ) and purchased for experimental validation (Figure S4), representing a 99.95% reduction in screening burden. To establish rigorous prospective enrichment baselines, 900–1200 compounds were randomly selected from the same library and screened under identical Dianthus TRIC conditions, yielding validated hit rates of 0.100% (CD28), 0.083% (TIM-3), and 0.111% (VISTA), consistent with the characteristically low small molecule hit rates against PPI-driven immune checkpoint interfaces.

Validation of the 50 model-selected compounds per target yielded primary TRIC hit rates of 44% (CD28), 50% (TIM-3), and 40% (VISTA). Following dose–response confirmation, 15, 17, and 13 compounds were validated as confirmed binders (30%, 34%, and 26% hit rates), corresponding to enrichment factors of 300×, 408×, and 234× over the random prospective baselines (Table S1; Figure ). Attrition of 31.8–35.0% from primary TRIC to confirmed binding reflects rigorous false positive exclusion. Across all three targets, 45 of 150 model-selected compounds were confirmed binders (30.0% overall hit rate).

3.

Prospective validation hit rates for HTS-Oracle X. Primary TRIC hit rates (light blue), dose–response confirmed hit rates (dark blue), and experimentally established random prospective baseline hit rates (gray) for CD28, TIM-3, and VISTA. Confirmed hit rates of 26–34% correspond to enrichment factors of 234–408× over the random prospective baseline.

Among 45 confirmed hits (Table S2), 16 demonstrated submicromolar equilibrium dissociation constant (K _D) values: 7 for CD28, 7 for TIM-3, and 2 for VISTA (Table ; Figure ). For CD28, top hits HX-CD28-1 (K _D = 0.233 μM), HX-CD28-2 (K _D = 0.262 μM), and HX-CD28-3 (K _D = 0.443 μM) extend the chemical diversity of CD28 small molecule binders reported by our group. For TIM-3, HX-TIM3-1 (K _D = 0.249 μM), HX-TIM3-2 (K _D = 0.257 μM), and HX-TIM3-3 (K _D = 0.289 μM) complement TIM-3 binders identified through pharmacophore-based approaches. For VISTA, HX-VISTA-1 (K _D = 0.345 μM) represents a remarkable addition to the limited small molecule chemical matter available for this target. , To characterize the applicability domain of HTS-Oracle X, we computed the nearest-neighbor Tanimoto similarity (Morgan fingerprints, radius = 2, 2,048-bit) between each of the 45 confirmed binders and the corresponding 45,760-compound training library. All confirmed hits fell within the bulk of the training chemical space, with median nearest-neighbor similarities of 0.44 (CD28), 0.45 (TIM-3), and 0.52 (VISTA) and a minimum observed similarity of 0.28 across all targets, indicating that prospective predictions, including the highest-affinity hits, were made within the model’s applicability domain rather than through extrapolation to structurally novel chemotypes (Table S3; Figure S5).

2. Top Confirmed Binders per Target from HTS-Oracle X Prospective Screening.

Target	Compound ID	K D (μM)	Pred. ΔF norm (%)	Selection Score
CD28	HX-CD28-1	0.233	14.43	14.20
CD28	HX-CD28-2	0.262	20.37	19.96
CD28	HX-CD28-3	0.443	17.94	17.64
TIM-3	HX-TIM3-1	0.249	17.13	16.72
TIM-3	HX-TIM3-2	0.257	20.15	19.76
TIM-3	HX-TIM3-3	0.289	16.04	15.52
VISTA	HX-VISTA-1	0.345	13.35	12.77
VISTA	HX-VISTA-2	0.931	16.46	15.71
VISTA	HX-VISTA-3	1.004	14.80	14.28

4.

Binding affinity (K _D) distribution of confirmed hits per immune checkpoint target. Individual data points represent confirmed binders; horizontal lines indicate median K _D values. Numbers in boxes indicate sub-micromolar hits per target. K _D values determined by Monolith X spectral shift dose–response.

Enrichment factors of 300× (CD28), 408× (TIM-3), and 234× (VISTA) were calculated against experimentally established random prospective baselines from the same 100,160-compound library (Figure A). This approach, in which the enrichment denominator is derived from experimental random screening of the same prospective library, provides the most rigorous and scientifically defensible enrichment estimate, directly comparable across platforms and target classes. For CD28, the only target shared with the original HTS-Oracle platform, HTS-Oracle X achieves 300× enrichment vs 8.4× for the original platform, a 36-fold improvement, alongside a step change in screening burden reduction from 70% to 99.95% and a substantially larger prospective screening library (100,160 vs 1,152 compounds; Figure B). The most potent CD28 binder (HX-CD28-1, K _D = 0.233 μM) demonstrates sub-micromolar target engagement, extending the affinity frontier for small molecule CD28 modulators.

5.

Enrichment factors for HTS-Oracle X. (A) Enrichment factors for CD28 (300×), TIM-3 (408×), and VISTA (234×) calculated against experimentally established random prospective baselines from the same 100,160-compound library. (B) Generational comparison for CD28: HTS-Oracle (8.4×) vs HTS-Oracle X (300×), representing a 36-fold improvement.

To characterize the structural and functional properties of the most potent confirmed binders, the top hit compound per target was advanced to comprehensive biophysical and cellular evaluation (Figure ). The chemical structures of HX-CD28-1, HX-TIM3-1, and HX-VISTA-1 reveal structurally distinct small molecule scaffolds across the three targets, reflecting the chemical diversity captured by HTS-Oracle X across the enamine library (Figures A–C). Dose–response binding isotherms obtained by Monolith X spectral shift confirmed sub-micromolar affinities for all three compounds: HX-CD28-1 bound CD28 with a K _D of 233 ± 24.7 nM (Figure D), HX-TIM3-1 bound TIM-3 with a K _D of 249 ± 16.3 nM (Figure E), and HX-VISTA-1 bound VISTA with a K _D of 345 ± 46.1 nM (Figure F), collectively establishing direct and target-selective engagement for all three scaffolds. These affinities are among the highest reported for small molecule binders of CD28, TIM-3, and VISTA extracellular domains and represent a meaningful advance in the available chemical matter for this target class.

6.

Chemical structures, biophysical binding, and functional characterization of top hit compounds identified by HTS-Oracle X for CD28, TIM-3, and VISTA. (A–C) Chemical structures of HX-CD28-1 (A), HX-TIM3-1 (B), and HX-VISTA-1 (C). (D–F) Monolith X binding isotherms for HX-CD28-1 binding to CD28 (D, K _D = 233 ± 24.7 nM), HX-TIM3-1 binding to TIM-3 (E, K _D = 249 ± 16.3 nM), and HX-VISTA-1 binding to VISTA (F, K _D = 345 ± 46.1 nM). Data are presented as mean ± SD (n = 5). (G) Functional evaluation of HX-CD28-1 in a tumor–PBMC coculture assay. IFN-γ and IL-2 secretion (top) and soluble CD69 levels (bottom) were measured after 48-h coculture of A549 tumor spheroids with human PBMCs (E:T ratio 5:1) in the presence of anti-CD3 (0.3 μg/mL), FR104 (10 μg/mL, positive control), or HX-CD28-1 (1, 5, 10 μM). (H) Functional evaluation of HX-TIM3-1. IFN-γ and IL-2 cytokine levels were measured in PBMCs cultured with recombinant TIM-3 in the presence of anti-TIM-3 mAb (positive control) or HX-TIM3-1 (1, 5, 10 μM) (top). Normalized cell viability of THP-1 AML cells, which endogenously express TIM-3, was assessed following compound treatment (bottom). (I) Functional evaluation of HX-VISTA-1. IFN-γ and IL-2 cytokine levels were measured in PBMCs cultured with recombinant VISTA in the presence of anti-VISTA mAb (positive control) or HX-VISTA-1 (1, 5, 10 μM) (top). Normalized cell viability of SKOV3 ovarian cancer cells, which endogenously express high levels of VISTA, was assessed following compound treatment (bottom). For all bar graphs, data represent mean ± SEM of n = 6 independent wells. Statistical comparisons were made to the respective vehicle or stimulated-only control using one-way ANOVA with Dunnett’s post hoc test. ns, not significant; *p < 0.05; **p < 0.01; ***p < 0.001; ****p < 0.0001.

Functional evaluation of HX-CD28-1 was performed in a tumor–PBMC coculture assay using A549 tumor spheroids (E:T ratio 5:1) with anti-CD3 stimulation (Figure G). Treatment with HX-CD28-1 produced dose-dependent suppression of IFN-γ and IL-2 secretion and a parallel reduction in soluble CD69 levels, a surface marker of early T cell activation, at 1, 5, and 10 μM (Figure G, top and bottom panels). The magnitude of suppression at 5–10 μM approached that of FR104 (10 μg/mL), a clinical-stage CD28-selective biologic antagonist used as a positive control, demonstrating that HX-CD28-1 engages CD28 co-stimulatory signaling in a physiologically relevant cellular context. For TIM-3, HX-TIM3-1 was evaluated across two orthogonal functional readouts (Figure H). In a human PBMC/recombinant TIM-3 cytokine assay, HX-TIM3-1 restored IFN-γ and IL-2 production in a dose-dependent manner relative to the TIM-3-suppressed baseline, consistent with relief of TIM-3-mediated immune suppression (Figure H, top panel). Complementarily, HX-TIM3-1 reduced the viability of THP-1 acute myeloid leukemia (AML) cells, which endogenously express TIM-3, in a concentration-dependent fashion at 1–10 μM, with efficacy comparable to the anti-TIM-3 monoclonal antibody positive control at higher concentrations (Figure H, bottom panel). TIM-3 expression on AML blasts has been established as a driver of immune evasion and disease progression, and the dual cytokine and viability activity of HX-TIM3-1 provides orthogonal evidence of on-target functional engagement. An analogous two-readout functional profile was obtained for HX-VISTA-1 (Figure I): dose-dependent restoration of IFN-γ and IL-2 in PBMCs cultured with recombinant VISTA (Figure I, top panel) and concentration-dependent reduction in viability of SKOV3 ovarian cancer cells, which endogenously express high levels of VISTA, with an efficacy profile comparable to the anti-VISTA monoclonal antibody control (Figure I, bottom panel).

Taken together, the structural diversity of the three hit scaffolds, their sub-micromolar biophysical affinities, and their on-target functional activity across mechanistically distinct immune checkpoint contexts collectively establish Figure as the critical proof-of-concept validation that HTS-Oracle X identifies not merely biophysical binders but functionally active small molecule immune checkpoint modulators, a distinction of central importance for the translational relevance of AI-guided hit discovery platforms in immuno-oncology.

In summary, HTS-Oracle X establishes a new performance benchmark for AI-guided small molecule discovery against immune checkpoint PPIs. Three architectural advances over the original HTS-Oracle platform, bidirectional cross-attention fusion, continuous ΔF _norm regression, and Monte Carlo Dropout uncertainty quantification, collectively drive a 36-fold improvement in CD28 enrichment and deliver 234–408× prospective enrichment factors across all three targets, with 45 dose–response confirmed binders from 150 model-selected compounds (30.0% overall hit rate) and 16 sub-micromolar hits. Critically, the top hit per target, HX-CD28-1 (K _D = 233 nM), HX-TIM3-1 (K _D = 249 nM), and HX-VISTA-1 (K _D = 345 nM), not only confirm direct sub-micromolar target engagement by Monolith X but also demonstrate on-target functional activity across mechanistically distinct immune cell and tumor coculture assays, establishing that HTS-Oracle X delivers functionally validated chemical matter rather than biophysical binders alone. Importantly, the platform is generalizable to any immune checkpoint for which biophysical screening data are available. Future work will include systematic benchmarking of alternative molecular fingerprint representations and parameter settings (e.g., Morgan fingerprint radius and bit length) to further optimize predictive performance across diverse target classes.

Glossary

Abbreviations

AI

artificial intelligence

AML

acute myeloid leukemia

AUC

area under the curve

CD28

cluster of differentiation 28

CLS

classification token

ΔF _norm

normalized fluorescence change

E:T

effector-to-target

HTS

high-throughput screening

IFN-γ

interferon gamma

IL-2

interleukin-2

ICOS

inducible T cell co-stimulator

K _D

equilibrium dissociation constant

LAG-3

lymphocyte activation gene 3

LASSO

least absolute shrinkage and selection operator

mAb

monoclonal antibody

MCD

Monte Carlo Dropout

MI

mutual information

PCA

principal component analysis

PBMC

peripheral blood mononuclear cell

PD-1

programmed cell death protein 1

PD-L1

programmed death-ligand 1

PPI

protein–protein interaction

ROC

receiver operating characteristic

SD

standard deviation

SEM

standard error of the mean

SMILES

simplified molecular-input line-entry system

TIM-3

T cell immunoglobulin and mucin domain-containing protein 3

TIGIT

T cell immunoreceptor with Ig and ITIM domains

TRIC

temperature-related intensity change

VISTA

V-domain Ig suppressor of T cell activation

The code is available on Zenodo (DOI: 10.5281/zenodo.20376686). Screening data and training library is available on Zenodo (DOI: 10.5281/zenodo.21063701).

The Supporting Information is available free of charge at https://pubs.acs.org/doi/10.1021/acsmedchemlett.6c00289.

• Experimental procedures, prospective validation results for HTS-Oracle X, complete list of 45 dose–response confirmed binders identified by HTS-Oracle X, primary Dianthus TRIC biophysical screening against CD28, TIM-3, and VISTA, and selection score distributions for the three targets (PDF)

The manuscript was written through contributions of all authors. All authors have given approval to the final version of the manuscript.

No unexpected or unusually high safety hazards were encountered.

AI Use Disclosure: ChatGPT (OpenAI) was used solely to assist in the creation of Figure and the Table of Contents (ToC) graphic. All scientific content, experimental design, data analysis, interpretation, and manuscript text were developed, reviewed, and verified by the authors, who take full responsibility for the accuracy and integrity of the work

The authors declare no competing financial interest.