LiTENexus: An End-to-End Virtual Drug Discovery System Based on Quantum Chemical Grounding

Artificial intelligence has revolutionized drug discovery, yet current molecular representation models predominantly rely on data-driven pre-training capturing statistical correlations rather than underlying physicochemical laws. This severely hinders out-of-distribution (OOD) generalizability and mechanistic interpretability. To overcome this, we propose the Quantum Chemical Information Injection (QCII) mechanism, which dynamically encodes fundamental quantum laws into physics-informed neural operators, and its foundational implementation, the LiTEN-Base operator. Driven by QCII, LiTEN-Base utilizes a field reconstruction paradigm to capture deep spatial topological constraints and long-range polarization, bridging the microscopic quantum realm and macroscopic pharmacological metrics. Building upon LiTEN-Base's robust representations, specialized downstream capabilities achieve exceptional performance. For microscopic property derivation, the model achieves quantum-level accuracy comparable to the M06-2X method, accelerating computation by 2-3 orders of magnitude over standard Density Functional Theory (DFT). In macroscopic pharmacokinetic prediction, it establishes state-of-the-art performance, outperforming over 20 baselines on Biogen dataset, excelling in 25 of 33 PharmaBench metrics, and dominating 62 of 77 ADMETLAB 3.0 tasks. Crucially, it exhibits robust generalization in complex OOD spaces like natural products and cyclic peptides. Finally, in cross-modal virtual screening, extending the latent space via the Quantum Manifold Dense Retrieval (QMDR) paradigm achieves record-breaking hit rates on DUD-E (𝐸𝐹 1% of 44.06) and LIT-PCBA (𝐸𝐹 1% of 6.77). Collectively, these capabilities constitute LiTENexus, an end-to-end quantum-inspired AI-aided drug discovery (AIDD) system. Ultimately, LiTENexus provides a highly extensible, physics-grounded framework redefining the efficiency and accuracy of molecular design, screening, and property optimization.