Models trained on public repositories like ChEMBL achieve high accuracy merely by predicting the creator of a molecule from its structure rather than learning chemical physics. Replacing molecular formulas with a vector indicating the probability of the chemist author yields identical predictive performance on target activity.
Because DNA encoded libraries are built combinatorially from building blocks, millions of molecules differ by only one or two changes. This structured variation enables machine learning models to isolate exactly which chemical fragments drive target binding while keeping the rest of the molecule constant.
Many therapeutic target proteins naturally bind DNA, producing massive false positive readouts in DNA encoded library screens. Assigning dozens of distinct DNA barcodes to identical small molecules reveals whether target enrichment is driven by specific DNA nucleotide motifs rather than the attached chemical.
Traditional drug discovery groups run DNA encoded libraries against isolated, difficult targets in hope of a single hit, missing systematic library quirks. Screening every library across hundreds of targets surfaces batch-to-batch manufacturing variations, chemical motifs, and assay failure modes that single screens never detect.
In a 2,000-team Kaggle challenge using a 4-billion interaction dataset, models performed adequately when test molecules shared building blocks or core hubs with the training set. However, when evaluated against completely orthogonal chemical libraries with distinct chemistry, no model beat random chance.
Sorting public dataset molecules by molecular weight alone matches top machine learning models in binding prediction. This correlation arises purely from chemist behavior: as a starting molecule shows activity, chemists iteratively tack on functional groups to optimize affinity and solubility, inflating the size of active compounds.
Current co-folding and structural models memorize ligand-pocket pairings rather than understanding binding physics. When binding pockets are digitally mutated, filled, or deleted, structural models frequently predict that small molecules still bind in the non-existent pocket because of memorized positions.
Attention maps in binding models initially focus on non-functional, variable amino acid regions outside active domains, essentially creating protein nicknames to bypass learning physics. Screening systematically mutated targets with disrupted pockets forces cross-attention models to distribute focus evenly across actual binding residues.
Traditional lab-in-the-loop systems require slow robotic synthesis to test model hypotheses. By generating millions of physical binding interactions upfront, language models acting as medicinal chemists can hypothesize, select test compounds, and receive instant grounded feedback as if operating in a coding sandbox.
As neural network parameter counts grow, direct numerical models risk memorizing complex mutant variations and individual building blocks. Representing molecules, targets, and structural differentials in natural English provides an inductive bias and regularizer that helps agents learn causal structure-activity relationships instead.
Subtle protein misfolding often evades standard gel electrophoresis but ruins biological training data. By including chemical analogs of molecular chaperones like lumacaftor in the screening library, unexpected binding hits instantly flag improperly folded protein batches so they can be discarded before polluting ML datasets.
