Nurix Therapeutics, Inc. · NASDAQ:NRIX
Nurix Therapeutics — TPD Assay Summit 2022
30 slides from Nurix Therapeutics's conference presentation, Dec 8, 2022
Slide 1: Molecular Property Prediction Using Machine Learning
Molecular Property Prediction Using Machine Learning TPD Assay Summit San Diego, CA December 8, 2022
Slide 2: Important Notice and Disclaimers
Important Notice and Disclaimers This presentation contains statements that relate to future events and expectations and as such constitute forward-looking statements within the meaning of the Private Securities Litigation Reform Act of 1995. When or if used in this presentation, the words “anticipa
Slide 3: Presentation Outline
Presentation Outline • Degrader optimization is difficult due to their unique chemical space and the lack of predictive tools to guide development • Nurix has successfully developed machine learning models able to accurately predict key properties of degraders, such as solubility and permeability. •
Slide 4: Degrader Optimization Remains Largely Empirical
Degrader Optimization Remains Largely Empirical ...We're still in the "try 'em and see" stage of optimization. … It would be much, much nicer if we could stand at the whiteboard or look at a screen, pursing our lips thoughtfully and then pointing purposefully at The Compound To Make, but we ain't th
Slide 5: Degrader Molecules Are Outside the Domain of Applicability for Commonly Applied Intuition and Rules
Degrader Molecules Are Outside the Domain of Applicability for Commonly Applied Intuition and Rules Targeted Protein Degrader molecules occupy a property space well beyond the traditional Lipinski RO5 molecules. As a result, intuition and rules for predicting physicochemical properties developed bas
Slide 6: Degrader Molecules Are Outside the Domain of Applicability for Commonly Applied Intuition and Rules
Degrader Molecules Are Outside the Domain of Applicability for Commonly Applied Intuition and Rules Targeted Protein Degrader molecules occupy a property space well beyond the traditional Lipinski RO5 molecules. As a result, intuition and rules for predicting physicochemical properties developed bas
Slide 7: Poor Correlation Between Ksol and 1D Molecular Properties
Poor Correlation Between Ksol and 1D Molecular Properties Extended view of weak correlation between molecular descriptors and solubility. A more thorough evaluation of 1D properties commonly thought to affect or relate to the solubility of small molecules shows very low correlation to experimental d
Slide 8: More Sophisticated Cheminformatics Calculators Similarly Struggle to Generalize to Degraders
More Sophisticated Cheminformatics Calculators Similarly Struggle to Generalize to Degraders Well Used Comp Chem Software #1 Well Used Comp Chem Software #2 Log(Calculated Solubility) Log(Calculated Solubility) R2 = .37 R2 = .31 M = .45 M =.76 B = -4.7 B = -7.1 Log(Ksol) Log(Ksol) Popular CompChem a
Slide 9: Active Literature Addressing Property Prediction in Bro5 Space
Active Literature Addressing Property Prediction in Bro5 Space 9
Slide 10: Literature Models Fail to Translate to Degrader Experimental Data
Literature Models Fail to Translate to Degrader Experimental Data Log(Papp AB) Min SA 3D PSA Impact of Dynamically Exposed Polarity on Publication Data Nurix Data Permeability and Solubility of Chameleonic Drugs Beyond the Rule of 5 Sebastiano, et al., J Med Chem, 2018 Log(Ksol) Literature model for
Slide 11: Motivation for Using Machine Learning Models for Property Prediction in the Degrader Design Space
Motivation for Using Machine Learning Models for Property Prediction in the Degrader Design Space • The rules commonly used to guide optimization campaigns are not applicable to targeted degraders, which occupy a distinct chemical space • This frustrates rational design and optimization of degrader
Slide 12: Solubility Assays from a Drug Design Perspective
Solubility Assays from a Drug Design Perspective Kinetic solubility vs thermodynamic solubility Thermodynamic Kinetic Preferred in later stages of drug Pre-dissolved Solid compound + compound + Early phase of discovery to drug discovery aqueous solvent strategies In-vivo aqueous buffer studies Img s
Slide 13: Nurix Solubility Dataset and Ways of Framing the ML Problem
Nurix Solubility Dataset and Ways of Framing the ML Problem Binary Classification (Soluble/Insoluble) Multiclass Classification (High/Med/Low) Regression (Predict numeric Ksol value) Leveraging the Nurix solubility dataset, we evaluated both Regression and Classification models (and a host of implem
Slide 14: Dataset Partitioning for Multiclass ML Model
Dataset Partitioning for Multiclass ML Model Multi-class Ksol distribution Dataset Splitting for ML Dev 1592 Initial dataset 10 µM 100 µM 595 482 338 Validation Prospective Train molecules molecules molecules 145 89 Not a part of initial dataset Combined small molecule and degrader Proper dataset hy
Slide 15: Model Architecture of Best Performing Ksol ML Model
Model Architecture of Best Performing Ksol ML Model Input: SMILES Feed model low with C(#N)C1=Cc2n(c(- embeddings c3ncc(C(=O).... Predict class medium high • This is a variant of commonly used molecular property prediction ML model called Chemprop. • Represents a molecule as a graph made up of atoms
Slide 16: Understanding Model Performance: Confusion Matrix
Understanding Model Performance: Confusion Matrix Incorrect Predictions lie in off-diagonal boxes (we want small numbers here) Correct Predictions lie along the axis (we want big numbers here) Das, C., Sahoo, A.K. and Pradhan, C., 2022. Multicriteria recommender system using different approaches. In
Slide 17: Model Performance on Prospective Dataset
Model Performance on Prospective Dataset 17
Slide 18: Generalizability Across Diverse Chemical Space
Generalizability Across Diverse Chemical Space High accuracy in prospective prediction of the true label of TPD compounds Model shows a high level of generalizability across compound classes, and high accuracy for TPD 18
Slide 19: nanoBret Target Engagement Assay to Assess Permeability
nanoBret Target Engagement Assay to Assess Permeability Artificial membrane Cell monolayer assays nanoBret Target Engagement assay assays (PAMPA) (MDCK, Caco-2) Degrader 1 Degrader 2 Evaluate shift in binding to intracellular target in live vs permeabilized cells Endpoints: Relative Measures passive
Slide 20: Nurix Permeability Dataset and Framing the Problem
Nurix Permeability Dataset and Framing the Problem Binary Classification (Permeable/Impermeable) Multiclass Classification (High/Med/Low) Regression (Predict numeric TE AI value) Due to the small size of the Nurix permeability dataset, we evaluated a binary-classification model 20
Slide 21: Permeability Data as a Binary-Class ML Dataset
Permeability Data as a Binary-Class ML Dataset • Dataset distribution : 281 degraders representing multiple ligases • Data labeling: If Availability Index (AI) > 5, AI < 5 AI > 5 then non-permeable else permeable. 21
Slide 22: Model Architecture of Best Performing Permeability ML Model
Model Architecture of Best Performing Permeability ML Model Random Forest A simple decision tree that predicts permeability based on molecular properties 22
Slide 23: Performance of Permeability Model on Prospective Dataset
Performance of Permeability Model on Prospective Dataset Actual permeability class Predicted permeability class Model shows high levels of generalizability across different Ligases. 23
Slide 24: Investigating Model's Perspective for Important Features
Investigating Model’s Perspective for Important Features Positive correlation • Feature importance tells which (higher Perm) features contributed to model’s learning process. • Bars to the left are negatively correlated and to the right are positively correlated to permeability. • Feature importance
Slide 25: Permeability with Varying Molecular Properties
Permeability with Varying Molecular Properties TPSA CLogP Compound : dBET6 Compounds with higher Compound : dBET1 TPSA – 194.05 Predicted as – Poor permeability TPSA and greater CLogP ClogP: 3.68 are identified as poorly Predicted as – Poor permeability permeable by the model Compound : Iberdomide T
Slide 26: Accuracy in Predicting Across Overlapping Property Space
Accuracy in Predicting Across Overlapping Property Space Experimental Data Actual permeability class AI<5 (Permeable) AI>5 (Impermeable) 21 2 Low 69% accuracy in convoluted property space TPSA 18 23 High Low High CLogP Predicted permeability class Permeability ML model predicts correct class members
Slide 27: Permeability Prediction: Recent and Similar Approaches
Permeability Prediction: Recent and Similar Approaches • In the following section, we will see some recent efforts published after our method was initialized. • Similar to our method, most of these efforts make use of molecular properties as features to train the model. • Surprisingly, the top featu
Slide 28: Emerging Consensus Regarding Highest Importance Features
Emerging Consensus Regarding Highest Importance Features Reliable prediction of Caco‐2 permeability by supervised Using in vitro ADME data for lead compound selection: An recursive machine learning approaches : Falcón-Cano, G. emphasis on PAMPA pH 5 permeability and oral bioavailability et.al Jordan
Slide 29: Important Take Homes
Important Take Homes • ML methods can learn interesting patterns from the chemical data where generalized literature-based methods might lack accuracy. • Most of the published methods for property prediction are built in small molecule chemical space, and thus might not span well for degraders which
Slide 30: Thank you
n U hy Thank you