AI summary by TwinMind
Lecture 5: Limits of Technologies
MIT OpenCourseWare
1h 0m 57s
Introduction to Technology Limitations
- Following Zach's enthusiastic microwave technology talk from 1-2 weeks ago
- New technologies initially generate optimism that all problems will be solved within years
- Reasons: need investors, public funding
- Eventually realistic expectations emerge requiring consideration of actual limitations
- Science aims to make predictions about systems
- Limitations affect predictive power in terms of accuracy, sensitivity, completeness
Noise Definition and Characteristics
- Noise defined as irrelevant/meaningless data occurring with desired information (Webster definition)
- Noise not always bad - can be important signal
- Cosmic background radiation discovery: two radio astronomers 50 years ago found "noise" from all directions, won Nobel Prize
- Cisplatin discovery: platinum electrodes created unexpected cell effects, led to chemotherapeutic agent
- Biological noise examples:
- Mutations crucial for evolution
- "Junk DNA" in genome sequencing - unknown function but may determine spatial gene distances
- Asymmetric cell division during differentiation - RNA/proteins distributed by chance between daughter cells
- Stochastic fluctuations important for complex system stability
Stochasticity in Biological Systems
- Genetic networks are stochastic systems
- Few hundred transcription factor copies per nucleus (sometimes only 50)
- Intracellular environment not free solution
- Slow reaction kinetics
- Experimental evidence from E. coli study (2 years ago, similar in yeast):
- Two proteins under same promoter with different fluorescent tags
- Expected same expression ratio if deterministic
- Found variable ratios (red, green, yellow cells) due to stochasticity
- Population average measurements add noise
- Grind millions/thousands of cells for genomics/proteomics
- Even single cell measurements interfere with/kill cells
- End up with population averages again
Measurement Accuracy and Statistics
- No measurement without noise - accuracy/sensitivity limitations expected
- Triplicates with identical numbers indicate systematic error (photometer miscalibration)
- Continuous variables expected to have spread around true value
- Statistics invented for biological measurements, addresses:
- True parameter values
- Bayesian analysis (most frequent in biology) - updating beliefs through repeated experiments
- Normalization to correct systematic errors
- Pattern recognition in large datasets to avoid chance correlations
Microarray Measurement Noise Sources
Affymetrix Chip Process
- Tissue → RNA extraction → RT (reverse transcriptase) → cDNA/cRNA production → hybridization
- RT step limitations:
- Stochastic initiation requiring primer
- Enzyme drops off during transcription
- Stronger signal from 3' end than 5' end
- Uses poly-A tail as initiator (random primers for bacteria)
- cRNA production issues:
- Dye incorporation not linear
- Different transcription efficiencies for different messages
- cRNA fragmentation not uniform
- Cross-hybridization problems
Two-Color Microarray Issues
- Equal amounts of labeled cDNA samples mixed
- Ratio measurements problematic:
- No truly blank spots - background noise creates false signals
- Complete down-regulation appears as ~100-fold change instead of infinite
- Signal compression from -infinity to +infinity range caught at ±100-fold
Affymetrix Probe Design Problems
- 25 base pair probes, 11-16 probes per gene
- Expected: all probes for same gene show equal intensity
- Reality: 4-5 orders of magnitude variation between probes for same gene
- Due to cRNA secondary structure, other factors
- Must estimate true expression from highly variable probe set
Normalization Approaches
Systematic Error Correction
- Intensity distribution analysis across samples
- Assumption: most measurements should have similar distributions
- One outlier sample suggests systematic error (dye incorporation, reader miscalibration)
- Solution: shift curves to same mean/median, renormalize all numbers
- Two methods:
- Assume most things don't change
- Use error model (additive at low concentrations, multiplicative at high)
Two-Color Array Normalization
- Red vs green ratios shouldn't show intensity dependence
- Reality: curved relationship due to different dye incorporation efficiencies
- Correction: straighten curve assuming equal red/green incorporation
- Can use housekeeping genes (shaky concept) or special control genes
- Normalization functions: global mean/median or more complex approaches
Noise Impact on Information Content
- Reliable detection limit: 2-fold differences (optimistic, non-cross-platform)
- Time series experiment design considerations:
- Yeast: no point measuring more often than 5-10 minutes
- Mammalian cells: 15-30 minute intervals
- More frequent measurements just capture noise
Sensitivity and Completeness Issues
Network Size Estimates
- ~5,000 proteins in yeast interaction network
- Conservative estimate: few hundred thousand bionodes per cell
- 10-20,000 genes × <10 post-translational modifications
- Current technology far from complete coverage
Dynamic Range Problems
- Holland's yeast study (few years ago):
- Gene expression over 6 orders of magnitude (0.01 to >10,000 copies/cell)
- Microarray correlation with RT-PCR only starts at 2 copies/cell
- Many important genes (transcription factors) below microarray sensitivity
- Entire low-expression range invisible to microarray
Inherent Predictive Limitations
Chaos Theory Example
- Edward Lorenz at MIT (1960s) - weather prediction with deterministic differential equations
- Small initial condition changes → completely different outcomes
- Even perfect measurements may not enable prediction due to sensitivity to initial conditions
- Biology robust (we're sitting here talking) - likely at edge between deterministic and chaotic
Statistical Analysis Challenges
- Cancer gene example: finding consistently misregulated genes across samples
- Combinatorics problem: 6 cell lines, 13 consistently misregulated genes
- Gene co-regulation destroys independence assumption
- Permutation analysis must preserve correlation structure
- Statistical confidence can differ by 5-6 orders of magnitude depending on method
Discrete Measurement Noise (DNA Sequencing)
Basic Sequencing Issues
- 0.1% error rate (previously higher)
- Single vs multiple individual sequencing for SNP vs error determination
- Solution: sequence extensively to determine true sequence
Gene Finding Challenges
- 3 billion base pairs in human genome
- Must find genes, introns, exons, transcription factor binding sites in "sea of letters"
- Problem: abundant "junk DNA" or intergenic regions
- False positives appear by chance in non-coding regions
- Gene prediction algorithms tend toward liberal side (overpredict)
- Require cDNA libraries for validation, but only expressed genes captured
Unexpected Biological Complexity
- 1,600 antisense transcription units in human genome (much higher than expected)
- Chromosome 21/22 transcription analysis: ~10× more transcription than expected based on exons
- Cannot predict from primary sequence alone
Transcription Factor Binding Site Discovery
- Looking for sequences like TGGACT (variable: TGCACT, etc.)
- Constraints help: within 500 bp of start codon, tend to cluster
- Still very weak signal in noisy background
- Many obvious binding sequences don't function (higher-order DNA organization)
Cross-Species Conservation Solution
- Combine information from multiple species (human, chimp, mouse, rat, yeast)
- Assumption: important sequences conserved across species
- Alexander's group yeast study: 4 closely related species
- High synteny (same gene, same location)
- 100% nucleotide conservation for important genes
- Variable conservation for others
- Method: generate all random sequences (XYZ with 0-21 random bases between)
- Statistical significance tests:
- Intergenic vs genic conservation
- Upstream vs downstream conservation
- Conserved sequences found in front of functionally related genes
- Many computationally predicted sites confirmed experimentally
Key Conclusions
- Important biological information hidden in sea of irrelevant information
- Computational approaches alone usually insufficient
- Cross-species conservation provides crucial biological context
- "Gold nuggets" emerge when combining computational and biological approaches
- Biological organisms built long ago - blueprints lost, everything experimental
