Week 6
Protein function prediction; machine learning
The course turns from structure to function. Function can be inferred from a structure by fold matching, binding-pocket analysis, 3D active-site templates, and conservation. Structural annotation scales this to whole genomes by assigning domains with HMMs. Since no single method works every time, function prediction becomes a learning problem. That leads to an introduction to machine learning (supervised learning, capacity, generalization, estimation) and to the first shallow model, the support vector machine.
Objectives
- Describe how protein function is defined and recorded: four-level EC numbers for enzymes, and the three Gene Ontology vocabularies (molecular function, biological process, cellular component).
- Infer candidate functions from a structure using fold matching (DALI, CATH, SCOP), surface clefts and binding pockets, 3D residue templates such as the Ser-His-Asp catalytic triad, and conservation mapped onto the surface (ConSurf). Explain where each method fails.
- Explain how Gene3D and SUPERFAMILY assign structural domains to genome sequences with HMMs, and what the results show about coverage, power-law family sizes, and domain combinations.
- Frame a function-prediction task as supervised learning by choosing features, labels, a hypothesis class, and a loss, and by keeping training, validation, and test data separate.
- Explain capacity, underfitting, overfitting, and generalization error, and how hyperparameters tuned on a validation set control them.
- Derive the maximum likelihood estimate for a simple model, describe the bias and variance of an estimator, and contrast point estimation with Bayesian inference using a prior.
- Derive the hard- and soft-margin SVM as a maximum-margin optimization problem, and identify the support vectors from the dual solution.
- Apply the kernel trick to make an SVM nonlinear, and choose among linear, polynomial, and RBF kernels for features derived from sequence or structure.
Reading
| Gu & Bourne 21, 22 | Ch. 21 surveys structure-based function prediction: function definitions (EC, GO), fold matching, pocket detection, residue templates, conservation, DNA-binding prediction, and combined servers such as ProFunc. Ch. 22 covers structural annotation of whole genomes: HMM-based domain assignment (Gene3D, SUPERFAMILY), coverage, and what domain distributions show about evolution. Focus on the methods sections of both. |
| Goodfellow et al. 5 | Goodfellow et al.'s machine learning basics: learning algorithms, capacity, overfitting and underfitting, hyperparameters and validation sets, estimator bias and variance, maximum likelihood and Bayesian estimation, supervised methods including SVMs and kernels, and stochastic gradient descent. Focus on 5.1–5.5 and 5.7. The chapter also underpins week 7. |
Lectures
| 11.1 | Protein function prediction transcript | 22:32 |
| 11.2 | Introduction to machine learning transcript | 32:41 |
| 12 | Shallow machine learning: support vector machines transcript | 29:43 |
| 1:24:56 |