Biosci 1540: Computational Biology Nathan Brouwer nlb24 at pitt edu

What is “Computational Biology”?

• Biology
  • Study of life
  • Often organized in a hierarchy
• Computation
  • Computer
  • computer use
  • computing
• Computational
  • computational science

3 Types of biologists

• Lab / Field / Clinical biologists
• Theoretical biologist
• Computational biologist

Lab / Field / Clinical biologists study how the world IS

• Classic type of biologists
• Develop questions / hypotheses…
• … collect data to try to answer them
• Works with hands
• Currently: good computer user (e.g. spreadsheet)
• recently: Some know how to code (Perl, Matlab, R, Python)
• uses normal statistics

Theoretical biologists study how the COULD BE

• Since early 1900s: Math/Theory
• Pattern in natures provide inspiration
• May not work with / analyze real data
• Uses calculus, works on whiteboard
• uses computers to solve hard math / make plots
• often doesn’t work with lab/field researchers
• often doesn’t use statistics

Computational biologists

• First appears in 1960s
• Increasingly common since 1990s
• Works directly with lab researchers on their data
• OR uses databases / creates databases
• works on problems for which a single human couldn’t…
  • solve at all (assemble a genome)
  • or at least solve efficiently (ID species in 1000s of hours of bird song recordings )
• works on data for which normal statistics fail

Computational biologists often work on…

• reconstructing the past
  • phylogenetics
• predicting the future
  • How bad will next pandemic be
  • How likely a species is to go extinct
• informing actionable decisions
  • Which chemicals should be tested as new drugs

Computational biology is an approach NOT a field of study

• microscopy is a way to study life
  • revolutionized biology!
• computers are a tool for studying life
• “computational thinking” is a way to solve problems

What are some of the hardest questions in biology?

The Thinker

One of the hardest questions:

How do proteins fold?

Two issues

• predicting 3D structure from primary sequence
• predicting steps of folding

Phospholipase A2

“How do proteins fold” has a unique place in biology

• Only biology question I know of to have its own annual competition
• CASP: Critical Assessment of protein Structure Prediction
• Most specific question in biology to have a NIH-funded research center (https://predictioncenter.org/)

REVIEW: Central Dogma

REVIEW: Expanded Central Dogma

REVIEW: Expanded Central Dogma

Levels of protein structure

• primary:
  • amino acid sequence
• secondary:
  • alpha helices
  • beta-sheets
• tertiary:
  • 3D structure
  • folded, functional proteins
• quaternary
  • multiple sub units
  • e.g. hemoglobin

Protein-x interactions also important

• Transcription factors (proteins) bound to mRNA
• Proteins bound to DNA (e.g. histones)
• “Molecular machines” …
  • Ribosomes: rRNA + protein
  • Ribosomes: how does mRNA pass through it

STRUCTURAL BIOLOGY is the empircal study of macromolecular structures

• Isolated lots of protein (or other molecular)
• Use high-tech gadgets
• Generate tons of data
• Use lots of math and theory
• … SOLVED STRUCTURE

STRUCTURAL BIOINFORMATICS is the computational study of macromolecular sequences

Goal: Can we …

• predict structure just from DNA sequence
• …or even learn the rules for the steps of folding?

Strucual bioinformatics uses several approaches

• ab initio
• Threading
• machine learning / deep learning / AI

Ab initio methods try to predict structure from just sequence

• Start with primary sequence
• Write a computer program that knows the rules of chemistry / physics
• Key: different conformations have different STABILITY
• Borrow a supercomputer 😀
• Have the program figure out the conformation with the highest stability

Ab initio methods are REALLY hard

• take lots of time
• only works on simple proteins
• pure ab initio methods ignore all the hard empirical work of structural biologists

NOTE: Computer simulations and working with powverful computers are a key feature of computational biology

• SIMULATION = having a computer work through a problem without a known solution
• Imagine a head-to-head computer game where BOTH sides are controlled by separate computers

Homology modeling uses evolutionary similar proteins to predict 3D structure

• Structural biologists have SOVLED structure for thousands of proteins
• Compare unknown protein primary sequence to primary sequence of of SOLVED proteins
• This involves the bioinformatics process of MULTIPLE SEQUENCE ALIGNMENT
• Protein sequences stored in databases such as Protein Data bank (PDB) and UniProt
• Figure out what structure makes the most sense for unknown

Number of known structure increaing exponentially

strucs <- c(194259,185472,172880,158882,147386,
            136217,125151,114337,105085,95523,
            86184,77429,69502,61757,54468,47565,
            40431,34024,28690,23541,19394,16400,
            13587,10960,8604,6548
,4984,3812,2871,1582,886,694,507
,365,291,238,213
,195,175,153,117,85
,69,53,42,36,13)

strucs <- rev(strucs)

yrs <- 1:length(strucs) + 1975
plot(strucs ~ yrs, type  = "b", pch = 16, xlab = "Year", 
     ylab = "Total Structures",
     main = "Growth of PDB over time")

Homology modeling only works for proteins w/ close evolutionary relatives (HOMOLOGS)

• Relies on the availability of solved structures
• …and solving structures is HARD!
• Proteins without a known relatives (HOMOLOG) can’t be predicted
• Many 3D structures in proteins can be produced by VERY different primary sequences

NOTE: Studying evolution / using evolutionary information is a common aspects of computational biology

• Computational methods are useful for studying things in the past
• Building phylogenetics trees is a classic challenge of comp bio
• Many bioinformatics approaches utilize evolutionary insights behind the scenes (e.g. BLAST)

NOTE: using lots of empirical data produced by others is a common feature of comp bio

• empirical:
  • related to the real / observable world
• empirical data:
  • data collected from the real
• contrast with …
  • theoretical predictions
• e.g., black holes
  • for a long time predicted but not no empirical data

NOTE: Using databases is a common feature of comp bio

• Sequence searching:
  • NCBI, Protein Data bank, UniProt
• Literature mining:
  • PubMed

THREADING pieces together solved subunits of many proteins to predict unknown structure

• …

Structural bioinformatics approaches can be combined

• methods can be combined
• NOTE: combining different a

CASP is an annual competition to predict the structure of proteins

“Critical Assessment of protein Structural prediction”

• Each year structural biologists solve new structures
  • (yay lab people! 👏)
• CASP participants given PRIMARY sequence of protein which will soon have their 3D solved structures published
• participants try to use existing data to predict what newly solved proteins will look like
• CASP judges compare predictions to solved structures

NOTE: Key theme of MACHINE LEARNING (ML) is “holding out” data from analysis so that a finalized model can be judged

• normal statistics:
  • use all your data
• machine learning: divide data
  • training set:
    • data you build your model with
  • validation set:
    • data you keep hidden until you think you have a good model

CASP uses the machine-learning (ML) strategy by holding out data from participants

• training set:
  • known solved 3D structures in PDB
  • all participants can use
• validation set:
  • new, unpublished 3D structures
  • no one can access yet (no yet published)

Incremental progress has been made on predicting 3D structures

• predicting secondary structures pretty easy
  • we’ll use ML technique in class to predict 2ndary structures
• predicting simple secondary structures for highly CONSERVED proteins is possible
• people generally frustrated by lack of major progress
  • until recently….

Hello AlphaFold!

Recently ML has been used to revolutionize the prediction of macromolecules

AlphaFold uses

• evolutionary information on homology
• known 3D structures
• “deep learning” algorithm

What is a “deep learning algorithm”

• …“Deep” = we’ll get back to this
• …“Learning” = “machine learning” (ML)
  • “machine” = computer (why can’t they just say “computer”?…)
  • “learning” = analyzes existing data to try predict new stuff
  • So, when you see “ML” you can insert “data analysis with a computer focused on prediction”
• …“Learning Algorithm” = trail-and-error-ish based computer program
  • NOT based primarily on math or theory

Aside: Math vs. algorithms

Math

• When you do a standard statistical procedure on a computer
  • calculus or other advanced math done behind the scenes
  • solution reached in 1 step
  • Lingo “closed form solution”
  • Every computer or program solving the problem will give you SAME answer

Aside: Math vs. algorithms

Algorithms

• The computer takes multiple attempts to solve problem
• Each attempt gives an approximate answer
• More attempts = probably closer to best possible answer
• Different programs / computers doing same algorithm
  • will take a different path to final answer
  • may even provide slightly different answers!

Back to “Deep learning algorithm”

What does the “deep” mean”

• “Deep learning” is a subset of ML
• Usually applied to a “NEURAL NETWORK”
• in particular a “DEEP neural network”
• Deep = many layers to the network

So, alpha fold is …

• A “deep learning algorithm”
• aka …
  • “A data analysis approach focused on a prediction” (learning)
  • “…done with a computer” (machine)
  • “…that takes multiple approximate steps to develop its best answer” (algorithm)
  • “and specifically uses a neural network with many layers”

ML / DL is gaining mainstream recognition, but

• computer scientists have been solving problems for biologists with ML for over 30 years
• Qian and Sejnowski 1988. Predicting the secondary structure of globular proteins using neural network models. Journal of molecular biology.
• Cited 1609 times (Google Scholar)

Application of ML/DL in biology

Genomics

• gene prediction in assembled genomes
• predicting features of genes
  • splice sites

Medicine

• medical image analysis
  • predict disease state from MRI

Ecology

• bird song analysis
  • determine which species occur in a live recording (Kitze lab)

(Current) limits of AlphaFold

These should all be tackled with time…

• Can only predict what has been seen -Doesn’t know the “rules” of protein folding
  • Perhaps there are protein features we’ve never seen…
  • Can’t predict a novel structure
  • … but the structural biologists are still working hard!
• Unstructured proteins
  • Many proteins have highly unstructured regions
• Interactions, complexes
  • Protein-protein interactions
  • Protein-DNA, protein-mRNA

How does Alaphfold (kinda) work

• PDB: 194,259 solved structures
• Structures split into 3 groups
  • Biggest: Training set
  • Validation set
  • Test set
• Training set
  • develop the algorithm with this
• Validation set (holdout 1)
  • Once you think its working, check it with this
  • Model given a score
  • get a sense for its failing
  • …go back to working on training set
• Test set (holdout 2)
  • Once you think you’re done
  • check one last time again test set
  • ideally - use test set once

Once AlphaFold was fully trained it was used for prediction

• Predicted structures for proteins for which we only have the sequence
• biologists can now use those predictions to inform their research
• … model will be revised, updated etc as more structures solved
• Code is online for everyone to use!

Analogies for the training / validation / test set

• Training set:
  • practice problems
• Validation set (hold out 1):
  • practice test
• Test set
  • The Midterm / final exam
• Novel data
  • MCAT, GRE etc

Key aspecs of compio apparent in AlphaFold

This is “data science” and the focus of this class:

• BIG DATA: 194,259 proteins PDB, millions of amino acids, thousands of sets of homologs
• DATABASE: all this data has to be stored, organized, searched, extracted
• UNCONVENTIONAL DATA (not mentioned yet): 3D structures
• EVOLUTION: homologous proteins compared using multiple sequence alignments (MSA) and phylogenetic trees
  • Building MSAs and phylogenetic trees are classic problems in comp bio
• MACHINE LEARNING
• MATH & STATS: How to judge success of model
• CODING & PROGRAMMING
  • working with LOTS of data
  • organizing workflow of steps
  • implementing algorithm (*some, but not all, comp biologists work at this level)