The ideas beneath modern drug-discovery AI.

Van der Waals showed that molecules are not ideal points: finite size and weak attractions shape gases, liquids and molecular association.

Modern force fields, docking scores, molecular dynamics and ligand–protein packing depend on these non-covalent interactions.

Gibbs created the thermodynamic framework that links energy, entropy and chemical potential, making equilibrium and spontaneity quantitatively predictable.

Binding affinity, conformational stability, solvation, phase behavior and free-energy calculations all inherit this framework.

Boltzmann connected microscopic states to macroscopic thermodynamics, establishing how energy controls the probability of molecular configurations.

Conformational ensembles, molecular simulations, temperature scaling, sampling and energy-based generative models rely on this statistical view.

Arrhenius formalized how reaction rates increase with temperature and depend exponentially on an activation barrier.

Chemical stability, degradation, enzyme catalysis, metabolism and accelerated stability studies all use Arrhenius reasoning.

Fischer proposed that enzyme specificity arises from complementary shapes between a biological macromolecule and its substrate.

The metaphor seeded structure-based ligand design, pharmacophore thinking and the search for complementary binding pockets.

Ehrlich envisioned chemicals that selectively bind disease-causing cells or organisms while sparing the host and helped establish systematic chemotherapy.

Target selectivity, therapeutic index and mechanism-based screening remain central goals of drug discovery.

Hill introduced a compact mathematical description of cooperative oxygen binding by haemoglobin.

Dose–response curves, receptor occupancy, multisite binding and systems pharmacology still use Hill-type models.

Michaelis and Menten established a quantitative relationship among substrate concentration, catalytic rate and enzyme saturation.

Potency, enzyme inhibition, target engagement, metabolic clearance and mechanistic pharmacology routinely use this kinetic framework.

Their experiments showed that purified DNA could transfer heritable biological traits between bacteria.

Genomics, variant interpretation, gene therapy and sequence foundation models depend on DNA being the durable molecular carrier of biological information.

Hodgkin determined the structures of penicillin and vitamin B12 and later led the determination of insulin, proving that complex therapeutic molecules could be resolved atom by atom.

Structure-based drug design depends on the experimental structural tradition she helped establish.

Shannon defined information quantitatively and connected uncertainty to entropy, creating the mathematical language of coding and communication.

Sequence modelling, cross-entropy training, language models, mutual information and representation learning all use Shannon’s framework.

They exploited biochemical differences in nucleic-acid metabolism to design selective antimetabolites rather than relying only on unguided screening.

Mechanism-based design, pathway selectivity and iterative medicinal chemistry are direct descendants of this strategy.

They used chemical geometry and hydrogen bonding to predict recurring protein secondary structures before atomic protein structures were available.

Protein representation, fold recognition, structural priors and generative protein design all encode these recurring geometric motifs.

Their model explained how DNA structure permits faithful copying through complementary base pairing.

Sequence analysis, variant prediction, genome design and nucleic-acid therapeutics all rest on this structural logic.

They introduced an acceptance rule that samples complex probability distributions through a stochastic walk.

Molecular simulation, Bayesian inference, conformational sampling and probabilistic model calibration use descendants of this algorithm.

Their diffraction measurements and quantitative analysis supplied crucial structural constraints on DNA’s helical geometry and hydration.

Structural genomics and sequence-to-structure reasoning began with experimentally grounded molecular geometry.

Sanger determined the complete amino-acid sequence of insulin and later developed chain-termination DNA sequencing.

Biological foundation models exist because proteins and genomes became readable, comparable and computable at scale.

Waddington described development as movement through a constrained landscape of possible cell states.

Single-cell embeddings, trajectory inference, reprogramming and virtual-cell models often represent cell identity as movement through a learned state landscape.

Koshland argued that binding partners can reshape one another, replacing a rigid lock-and-key picture with dynamic molecular adaptation.

Flexible docking, conformational selection, protein motion and ligand-induced pocket changes are modern extensions of this idea.

Crick clarified the directional flow of sequence information among DNA, RNA and protein while distinguishing information transfer from metabolism.

Multi-omic models and sequence foundation models connect genotype, transcript and protein through this information-flow framework.

Kendrew resolved myoglobin and Perutz resolved haemoglobin, showing that proteins possess intricate, reproducible three-dimensional architectures.

Protein structure prediction and structure-based design became meaningful because experimental crystallography established the target reality to predict against.

Anfinsen’s ribonuclease experiments showed that, under suitable conditions, the amino-acid sequence contains the information needed to recover the native biologically active structure.

Protein structure prediction, inverse folding and generative protein design all assume that sequence strongly constrains structure and function.

They showed that gene expression is controlled by regulatory circuits rather than being constitutively active.

Target biology, perturbation models, transcriptomic response prediction and virtual cells all require an explicit model of regulated gene programs.

Hansch related biological activity to hydrophobic, electronic and steric molecular descriptors using statistical models.

Classical QSAR established the central premise that molecular features can predict potency and guide optimization—the conceptual ancestor of modern molecular machine learning.

Black designed beta-blockers and H2-receptor antagonists by starting from a defined physiological receptor mechanism.

Target-based pharmacology and mechanism-led lead optimization follow the same design logic.

No direct model connection is currently indexed.

The MWC model explained how proteins switch between collective conformational states and transmit regulation between distant sites.

Allosteric drug design exploits remote pockets to modulate function, selectivity and resistance without competing at the active site.

Dayhoff assembled the Atlas of Protein Sequence and Structure and developed PAM matrices to quantify evolutionary relationships among proteins.

Protein language models, homology inference, multiple-sequence alignments and evolutionary priors inherit her conversion of sequence biology into computable data.

Levinthal observed that a protein cannot find its native structure by exhaustively searching all possible conformations, implying structured pathways or energy landscapes.

Modern folding algorithms, energy landscapes, learned priors and diffusion models solve a constrained search problem rather than brute-force conformational enumeration.

They connected quantum mechanics and classical molecular mechanics so complex biomolecular reactions could be simulated at useful scales.

QM/MM, molecular dynamics, free-energy methods and physics–ML hybrid platforms descend directly from this multiscale strategy.

They developed a method for producing antibodies of a single specificity indefinitely by fusing antibody-producing cells with immortal myeloma cells.

Therapeutic antibodies, diagnostic antibodies and antibody discovery platforms became scalable and reproducible.

Tonegawa demonstrated that antibody genes are rearranged during immune-cell development, explaining how a finite genome generates vast recognition diversity.

Antibody language models and repertoire design operate on the sequence space created by V(D)J recombination and somatic diversification.

Hopfield showed that a recurrent neural network can store patterns as attractors in an energy landscape.

Energy-based learning, associative retrieval and modern attention mechanisms share conceptual roots with this statistical-physics view of computation.

Smith developed phage display and Winter applied it to evolve and humanize therapeutic antibodies.

Display-based selection created an experimental search engine for protein binders and remains a core validation partner for computational antibody design.

Arnold established iterative mutation and selection as a practical way to evolve proteins toward desired functions.

Generative protein design increasingly closes the loop with directed evolution and experimental selection to optimize function and manufacturability.

Lipinski and colleagues identified simple physicochemical patterns associated with poor absorption or permeation among orally active drugs.

Generative chemistry and lead optimization routinely use drug-likeness and developability constraints inspired by this work.

Baker and collaborators developed Rosetta into a system for predicting structures and designing proteins with new folds and functions, later extending the field into generative design.

Modern binder design, inverse folding and diffusion-based protein generation build on this computational-design lineage.

They reconstituted CRISPR–Cas9 as a programmable RNA-guided DNA cleavage system.

CRISPR enables target validation, disease models, perturbation atlases, functional genomics and gene-editing therapeutics.

Diffusion models learn to reverse a gradual noising process, converting random noise into structured samples.

Modern protein-backbone, molecular-pose and biomolecular-complex generators use diffusion to sample valid three-dimensional structures and designs.

The Transformer replaced recurrence with attention, allowing models to learn long-range relationships in parallel across sequences.

Protein, genome, molecule and single-cell foundation models use attention to learn dependencies across biological sequences and multimodal inputs.