At Sravathi AI Technology, we believe that the next generation of AI-driven drug discovery will require more than data-driven predictions. It will require integrating the underlying principles of molecular physics with the learning capabilities of artificial intelligence.
We are pleased to share our latest research, PATHNet: A Thermodynamic Path-Integral-Inspired Neural Network for Protein–Ligand Interaction Prediction, now available as a preprint on bioRxiv.
Why PATHNet?
Protein–ligand interaction prediction remains a challenging problem in structure-based drug discovery. Conventional scoring functions often rely on predefined functional relationships, while purely data-driven approaches can face challenges in interpretability and generalization.
PATHNet explores a physics-informed approach that combines molecular mechanics-derived descriptors with a constrained neural-network architecture to learn complex, nonlinear protein–ligand interaction relationships.
Key technological innovations include:
🔹 Physics-informed molecular representation: Integration of electrostatic, van der Waals, solvation and ligand conformational-strain descriptors.
🔹 Non-Monotonic Piecewise Linear (NM-PWL) mapping: Learning complex nonlinear relationships between physical descriptors and binding-related outcomes.
🔹 Hybrid neural-network architecture: Combining an interpretable functional mapper with a bounded neural-network correction.
🔹 RG-inspired spectral regulation: Introducing a physics-inspired constraint to regulate neural-network parameter spectra during optimization.
🔹 Multi-start basin optimization: Exploring multiple optimization basins to identify better-performing model solutions.
Promising results across diverse protein targets
PATHNet was evaluated against three established computational approaches – GAABind, PIGNet and taqo – across seven DUD-E targets.
The model achieved a mean reported classification accuracy of 73.24%, compared with 53.89% for GAABind, 57.97% for PIGNet and 60.35% for taqo.
PATHNet achieved the highest reported accuracy on six of the seven evaluated targets.
These results provide encouraging evidence for the potential of integrating physics-grounded molecular representations with constrained neural-network learning.
Our vision
At Sravathi AI, we are committed to developing scientifically grounded AI technologies that address meaningful challenges in drug discovery.
PATHNet represents another step in our efforts to bring together computational chemistry, molecular physics and artificial intelligence to advance structure-based drug discovery.
We look forward to further developing and validating this framework across broader chemical spaces and protein target families.
📄 Read our preprint:
https://doi.org/10.64898/2026.10.11.758399
Authors: Ben Geoffrey A., Abhishek Singh, Sowmya Kanchan and Nagaraj M. Kulkarni.
This manuscript is available as a bioRxiv preprint and has not yet been certified by peer review.
#SravathiAI #DrugDiscovery #ArtificialIntelligence #PhysicsInformedAI #ComputationalChemistry #MolecularModelling #ProteinLigandInteractions #StructureBasedDrugDesign #AIinDrugDiscovery