Lab Notes

The Nanolix Lab Notes

Technical writing on generative chemistry, ADMET prediction, scaffold hopping, and the practical side of computational hit identification — from the scientists who run these programs.

Abstract molecular landscape showing balance between exploration of novel regions and exploitation of known clusters
Methods

Exploration vs. Exploitation in Chemical Space: Getting the Balance Right

Over-exploiting a known scaffold cluster gives you analogs with marginal improvements. Over-exploring produces novel scaffolds with unpredictable properties. The right balance depends on your program goals.
Abstract graph neural network visualization processing molecular graph structure
Computational Chemistry

How Accurate Are GNN-Based Molecular Property Predictions in Practice?

Graph neural networks for ADMET prediction have improved dramatically on benchmark datasets — but benchmark accuracy rarely translates directly to prospective performance.

 Abstract Pareto frontier visualization showing multi-objective trade-off surface in molecular optimization
Methods

Multi-Objective Optimization and Pareto Fronts in Drug Discovery

You cannot simultaneously maximize binding affinity, minimize toxicity, and maximize synthesizability without trade-offs. Pareto optimization surfaces the real trade-off frontier.

 Abstract visualization of molecular data handoff between computational design and synthesis workflows
Practical Guide

Getting CROs to Synthesize Computationally Generated Candidates

Generative models sometimes propose synthetic routes that CRO chemists reject outright. Here is how to structure deliverables so synthesis is actually feasible.

 Abstract scientific visualization of prediction uncertainty and confidence interval distributions
Computational Chemistry

How to Interpret Confidence Intervals in Computational ADMET Predictions

A predicted log P of 2.1 ± 0.8 means something very different from 2.1 ± 0.2. Understanding model uncertainty changes how you prioritize candidates.

 Abstract molecular fragment growing into larger drug-like structure through guided expansion
Methods

Fragment Expansion with ML Guidance: From 200 Da to Drug-Like

Fragment hits are cheap to find but hard to grow while maintaining binding and physical properties. Here is the ML-guided expansion approach we use in practice.

 Abstract protein surface visualization with cryptic binding site highlighted showing previously unexplored binding pocket
Drug Discovery

Small-Molecule Design for Historically 'Undruggable' Targets

Targets once deemed undruggable are yielding to fragment-based methods and generative exploration of cryptic binding sites. What we have learned running these programs.

 Abstract synthesis route visualization showing stepwise molecular transformation pathways
Methods

Why Synthetic Accessibility Must Be a First-Class Model Objective

Beautiful binding affinity predictions are useless if the compound requires 12-step synthesis and non-commercial reagents. Here is how we bake accessibility in.

 Abstract molecular scaffold transformation showing structural change while maintaining binding pharmacophore
Computational Chemistry

Scaffold Hopping Under Property Constraints: A Practical Guide

When your lead compound has liabilities you can't design out, scaffold hopping is the answer — but only if you constrain property predictions from the start.

 Split visualization contrasting grid enumeration versus flowing generative exploration of chemical space
Drug Discovery

Generative Chemistry vs. Enumerative Screening: When Each Makes Sense

Enumerative virtual screening is fast and interpretable. Generative methods explore regions enumeration misses. The choice depends on your scaffold constraints.

 Molecular structure with property prediction overlays showing ADMET profile visualization
Computational Chemistry

Predicting ADMET Properties Before You Touch the Bench

What 11 ADMET dimensions matter most in early-stage hit identification and how ensemble prediction models handle each.

 Abstract visualization of sparse chemical space sampling versus vast unexplored molecular landscape
Drug Discovery

Why Brute-Force Hit Identification Fails Small Biotech Teams

HTS screens millions of compounds but covers a vanishingly small fraction of viable chemical space. Here is why that math doesn't favor small teams.