Predicting Lipophilicity from SMILES Strings with PyTorch Geometric

Multitask learning enables simultaneous prediction of multiple ADME properties by leveraging correlations between related endpoints. The approach uses message-passing neural networks with SMILES inputs, stratified random splits, and ensemble methods. Critical to success is loss function engineering where weights are applied per-endpoint and per-compound to address data imbalances. When combining proprietary and public datasets, careful selection determines performance outcomes—proprietary data improves some endpoints (solubility, permeability) but degrades others (logD) due to assay differences. The key insight is that matching data characteristics to prediction tasks matters more than sheer data volume.

Inductive Bio trains models in groups of conceptually related properties. For example, the same model was trained for KO2 apparent permeability and efflux ratio models, trained on KO data but also on permeability assay data (MDCK) and non-permeability auxiliary tasks like logD. This grouped multitask approach improves performance by leveraging shared patterns across related properties.

This section covers the assessment of drug candidates through ADMET prediction. The Boiled Egg Model predicts intestinal absorption and brain permeation based on lipophilicity (logP) and topological polar surface area (TPSA). SwissADME predicts ADMET properties including absorption, distribution, metabolism, excretion, and toxicity. Lipinski's Rule of 5 evaluates oral bioavailability by checking molecular weight <500, logP <5, hydrogen bond donors <5, and acceptors <10. Osiris Property Explorer predicts toxicity risks including mutagenicity, carcinogenicity, and reproductive toxicity.

ADMET (Absorption, Distribution, Metabolism, Excretion, and Toxicity) is the foundational framework for evaluating drug candidates. This section demonstrates the complete computational workflow: obtaining molecules from PubChem, extracting SMILES strings, creating structural derivatives using Marvin Sketch, and predicting pharmacokinetic properties with SwissADME. SwissADME evaluates lipophilicity, water solubility (logS), GI absorption, and drug likeness following Lipinski's Rule of Five (molecular weight <500 Da, logP <5, ≤5 H-bond donors, ≤10 acceptors). Results use color-coded zones where light red indicates favorable properties. This systematic approach enables rapid exploration of chemical space before experimental testing, identifying whether structural modifications improve or degrade pharmacokinetic profiles.

ADMET prediction relies on key physicochemical parameters: (1) Aqueous solubility (LogS) - fundamental property affecting absorption and distribution, (2) pKa (partition coefficient) - indicates ionization state affecting membrane permeability, (3) LogP (octanol-water partition coefficient) - measures lipophilicity affecting distribution and membrane crossing. Permeability prediction is critical: (1) Blood-brain barrier permeability - assessed using in silico prediction, Caco-2 cell monolayers, and PAMPA assays for CNS-targeting drugs, (2) Intestinal permeability - major factor for oral absorption, predicted using Caco-2 assays and PAMPA. Drug transporter prediction involves P-glycoprotein (PGP), a major efflux transporter affecting drug absorption and distribution, predicted using computational tools to assess substrate/inhibitor status.

Molecular heat maps adapt the concept from image classification, where heat maps highlight pixels important for object recognition. In chemistry, they identify atomic contributions to predicted properties like binding affinity or logP values. If removing an atom increases predicted activity, that atom was decreasing activity; conversely, if removal decreases activity, the atom was beneficial. Explainable AI serves two purposes: building trust with synthetic chemists by providing understandable explanations, and uncovering chemical principles underlying structure-activity relationships (QSAR). Validating heat maps is challenging because we often lack ground truth for what makes molecules active. Two approaches exist: asking experts (imprecise and subjective) or testing against simple models with known behaviors. Crippen logP provides a quantitative benchmark—octanol-water partition coefficient calculated from fixed atomic contributions that sum to the total value. RDKit displays these as heat maps showing which atoms contribute positively or negatively to hydrophobicity. Since these are empirically derived values, they serve as ground truth for validating machine learning heat maps—any ML model trained to predict logP should produce heat maps matching these atomic contributions.

This section covers two complementary analysis techniques. First, the 'MLP' command calculates molecular lipophilicity potential maps where dark cyan areas represent more hydrophilic regions and gold/yellow areas represent most lipophilic regions. Second, B-factor analysis uses the 'color by attribute' command to visualize flexibility, showing lowest B-factors (most rigid) in blue and highest B-factors (most flexible) in red. The instructor demonstrates limiting surface display to atoms within 6.3 Å of the ligand and hiding ribbons/atoms to focus on the active site pocket.

Graph Neural Networks (GNNs) can predict lipophilicity (log D at pH 7.4) by representing molecules as graphs where atoms are nodes with features (atom type, formal charge, atomic weight) and bonds are edges with features (bond type), using adjacency matrices to capture connectivity, and processing this information through convolution and linear layers to learn molecular properties from the Campbell database of 4,200 compounds.

Black-box models face adoption barriers because scientists prefer interpretable models aligned with their cognitive understanding. This project prioritized explainability by reducing descriptors to seven most important ones while maintaining good accuracy. Key descriptors include: Filterate log S (computed log solubility, 69% contribution), Lipinsky (molecular mass, hydrogen bond acceptor abilities, octanol-water partition coefficient), SI (structure information content), Rncg (relative negative charges), and Rpcg (relative positive charges). These represent molecular properties and atomic arrangements influencing solubility.

GNNExplainer is an explainability method for graph neural networks that identifies relevant subgraphs and node features by maximizing mutual information between the original model prediction and a masked subgraph, using gradient descent to optimize a continuous mask applied to the computation graph of the target node; this approach provides counterfactual explanations by determining which nodes, edges, and features are most influential in the model's prediction while being agnostic to specific GNN architectures.

Emerging AI methods in cheminformatics focus on learning general rules about chemical structure-composition relationships from diverse databases of bioactive compounds. The REINFORCE framework combines generative models (learning to generate realistic chemical structures from SMILES notation) with predictive models (assessing whether generated compounds have desired properties). The generative model receives rewards when producing compounds predicted to have desired properties and punishments otherwise, progressively converging toward novel compounds with target properties. This represents a shift from virtual screening (selecting from existing compounds) to de novo design (generating entirely new compounds).

De novo design (generative modeling/inverse design) addresses the dual challenge of generating diverse molecular ideas and scoring them effectively. The approach guarantees projects-relevant synthesis by enumerating small synthetic derivatives, enables multi-parameter optimization biased toward desired properties, and incorporates extensive validation. Mature technologies now produce molecules comparable in medicinal chemistry quality to those generated by experienced chemist teams. Enterprise informatics platforms enable horizontal scaling by making models accessible to entire project teams, allowing medicinal chemists to directly run curated free energy calculations. Real-world implementation shows over 60% of calculations run by medicinal chemists, with projects achieving 1,000-fold potency improvements and 90% prediction accuracy. Analysis of collaborative programs shows 96% advancing past hit ID versus industry average of 80%, with over three-quarters advancing past lead optimization compared to 51% of traditionally operated projects. Three case studies demonstrate impact: an allosteric M1 inhibitor project progressed from hit identification to lead optimization entry in 10 months; a PLK1 oncology target project used protein residue mutation strategies to achieve gene family-wide selectivity; an L4β7 IBD target project achieved simultaneous optimization of potency, permeability, and PK properties.

Two primary paradigms exist for molecular design: virtual screening and generative methods. Virtual screening filters existing compound libraries (containing billions or trillions of molecules from chemical vendors) using property prediction models to identify promising candidates. Generative approaches aim to invent novel structures by inverting structure-property relationships to go directly from desired properties to candidate molecules. Both approaches face limitations: virtual screening is constrained by available compounds, while generative methods sometimes produce chemically unreasonable structures that cannot be synthesized. Bayesian optimization and uncertainty quantification help navigate these challenges.

Generative AI models enable virtual screening by exploring commercially available chemical spaces against target proteins to identify promising molecules. These models can predict which compounds from existing libraries might bind effectively to a protein pocket. The predicted binding interactions appear as dotted lines showing how the small molecule forms connections with the protein, allowing chemists to optimize the number and strength of these interactions to improve drug efficacy.

Hit identification employs two main computational strategies: virtual screening and de novo molecular design. Virtual screening involves computationally docking candidate molecules against target protein structures to predict binding affinity, using techniques like molecular docking (energy minimization), QSAR modeling (quantitative structure-activity relationship), and multimodal learning that combines molecular and protein features. De novo design generates entirely new molecular structures from scratch using generative models trained on known active compounds. Recent breakthroughs with AlphaFold and RoseTTAFold have revolutionized this field by enabling accurate prediction of protein 3D structures, allowing researchers to design molecules that fit precisely into protein binding pockets with unprecedented accuracy.