Week 13
Drug discovery: from computer-aided to AI-driven design
Design reverses prediction: instead of mapping a molecule to its structure or property, find a molecule that has a desired property. The week starts with the drug discovery and development process and the role of structural bioinformatics in it, then contrasts classical computer-aided design (combinatorial optimization of an energy function) with AI-driven inverse design using deep generative models. ProGen2 generates protein sequences with an autoregressive language model; Chroma generates structures and sequences with conditional diffusion.
Objectives
- Lay out the target-based drug discovery and development pipeline (target identification and validation, hit identification by HTS or virtual screening, hit to lead, lead optimization, preclinical studies, Phase I–III trials) and explain why attrition and cost are so high.
- Name where structural bioinformatics helps drug discovery: construct design for crystallography, target druggability, structure-based virtual screening, ligand efficiency during lead optimization, P450 and hERG modeling for ADMET, and SNP effects on drug response.
- Formulate computer-aided design as combinatorial optimization: choose residues, rotamers, or chemical fragments from a discrete set to minimize an energy or scoring function, and explain why the search space forces heuristic or pruned search.
- Contrast forward prediction (molecule to property) with inverse design (property to molecule), and pose inverse design as sampling from p(x | y) ∝ p(y | x) p(x) with a generative prior and a property model.
- Represent small molecules as SMILES strings or molecular graphs and proteins as sequences or 3D backbones, and match each representation to suitable generative models.
- Compare GANs, VAEs (encoder, decoder, ELBO), normalizing flows (exact likelihood by change of variables), diffusion models (learned denoising of a noising process), and autoregressive models on likelihood, sample quality, sampling speed, and training stability.
- Explain how ProGen2 trains causal protein language models (151M to 6.4B parameters) and samples with temperature and nucleus sampling, and why larger models did not always predict fitness better zero-shot.
- Explain Chroma's design choices: diffusion with correlated, polymer-like noise; a random-graph neural network that scales sub-quadratically; and conditioners (symmetry, substructure, shape, fold class, natural language) that make design Bayesian inference under constraints.
Reading
| Gu & Bourne 34 | Gu & Bourne's chapter on structural bioinformatics in drug discovery: history, the modern target-based process, generating protein structures, drug targets (324 for all approved drugs), hit identification, lead identification and optimization (ligand efficiency, ADMET, P450, hERG), databases, and personalized medicine. Focus on the process diagram and where structure enters each stage. |
| Goodfellow et al. 20 or Zhang et al. 16 | Goodfellow 20 surveys deep generative models: Boltzmann machines, directed generative nets, VAEs, GANs, and autoregressive networks. Zhang's generative chapter builds a GAN and a DCGAN in code (its number differs between editions). Focus on how each family defines, bounds, or avoids a likelihood. |
Papers
E Nijkamp et al. “ProGen2: Exploring the boundaries of protein language models”. Cell Systems 14(11), 2023. PDF
ProGen2 asks how model size and training data shape protein language models. The authors train causal transformers from 151M to 6.4B parameters on over a billion sequences from genomic, metagenomic, and antibody (OAS) databases, then test generation (AlphaFold 2 and IgFold check that samples fold) and zero-shot fitness prediction. Larger models fit held-out sequences better and generate diverse, plausible proteins, but are not always better at fitness prediction, so the data distribution matters as much as scale. It is the sequence-based, autoregressive example of generative design.
JB Ingraham et al. “Illuminating protein space with a programmable generative model”. Nature 623, 2023. PDF
Chroma asks whether one generative model can sample new protein structures and sequences, including complexes, and steer them toward desired properties. It uses diffusion with correlated noise that respects polymer statistics, a random-graph neural network scaling as O(N) or O(N log N), and conditioners for symmetry, substructure, shape, class, and natural-language prompts. Of 310 designed proteins tested, many expressed and folded, and two crystal structures matched the designs to about 1.0 Å backbone RMSD. It is the structure-based, diffusion example of programmable design.