Discovery of a Drug to Treat a Disease

Medical research on a disease may discover the disease's life cycle.

There may be a critical point in the life cycle that might be disrupted by a drug. E.g. a point where a chemical might bind and prevent the diseases action.

Search space of potential drugs is huge. Estimated $10^{40}$ drug like chemicals.

Using high throughput robot based screening in the region of $10^{6}$ can be measured.

Computer models are used at many stages of the discovery ``funnel''. They guide chemists as to where to look next.

Models predict as yet unmeasured properties.
They can screen ``virtual'' chemicals, i.e. chemicals that do not exist.
Virtual chemicals could be manufactured, if the computer suggested they might be interesting.

Discovery of a Drug to Treat a Disease

Rat P450 2B1,
Showing cell wall membrane attachment

See.

Why is P450 Important to Drug Discovery

• As drug proceeds down discovery funnel costs escalate
• Later a potential drug is abandoned the more expensive it is.
• A major cause of late failure is adverse interaction with other parts of the body (another is poor bioavailability).
• P450 is responsible for degrading most drugs on the market.
• Early prediction of P450 activity is sought by the pharmaceutical industry.
• Most companies are looking at computer based approaches.

GSK ``Blind Trial''

• GSK merger of Glaxo Wellcome + SmithKline Beecham 2000.
• Both GW and SKB had large libraries of chemicals which might be used as drugs. Little overlap between them.
• Large program to measure biochemical properties, esp. P450.
• Workshop to assess feasibility of using machine learning in drug discovery
• Blind trial. 12 Machine Learning techniques (mostly in GSK):

  Train on former SKB training data.

  Test on former SKB.

  Extrapolate to former GW.

GSK ``Blind Trial'' Three P450 IC50 data sets

• IC50 is concentration of chemical needed to inhibit P450.

  Continuous P450 IC50 measurement divided into three classes. Inhibitory, Substrate and Inactive.

• Three data sets:

  3000	(	1000,		1000,		1000)		training fSKB		(IC50 given)
  4570	(	91,		1563,		2916)		test fSKB		(no IC50)
  1932	(	114,		446,		1372)		extrapolation set fGW		(no IC50).

• For each chemical GSK calculated 121 ``features'' from the chemical's formula.

  Features are domain knowledge. They can be Boolean, categorical or continuous.

  They represent chemical properties. E.g. charge imbalance, acid or base, heavy metal.

Spread of IC50 in Training Data

IC50 relatively high quality but still noisy data.

Training data was deliberately set up to have equal numbers in each of the three classes.

The test (or interpolation) data set and the extrapolation set very unequal split.

GP Prediction of P450 Inhibition

• Evolve predictive model from high level features to P450 class (inactive, substrate, inhibitor).

• 121 chemical features used as supplied (albeit converted to floating point number) and IC50 value.

  (Might have been better to use $-\log_{10}$IC50 (pIC50) instead?)

  For GP, the training data was randomly split into train and verification (2:1).

• Two approaches:

  Classification via regression (least squared error).
  I.e. Evolve prediction of continuous IC50 value, then place in one of the three class bins.

  Direct classification (3 classes).

P450 IC50 Regression

• Individual consisted of one tree:

  Arithmetic operations, IFLTE, min, max, etc.

  121 features, 0,1,2,...,9 and 90 randomly chosen constants

  $\frac{2}{3}$ of 3000 labelled data used as training cases.

• Output of tree compared to supplied IC50 measurement

  Fitness$= 20000\times$   hits$- \sum_{i=1\ldots 2000} \vert f_i-$IC50$_i\vert^{2}$

  (a hit is a correctly classified example after thresholding)

  Size fair crossover, mutations. Population size 5000, 50 generations

  Best of five runs

• Best individual across the 3 classes, was simplified
  (size reduced from 65 to 45)

Classification

• Trained Individual consisted of 3 trees (1 per class)
• Final class prediction: from tree returning largest value.

  Arithmetic operations, IFLTE, min, max, etc.

  121 features, 0,1,2,...,9 and 90 randomly chosen constants

  $\frac{2}{3}$ of 3000 labelled data used as training cases

  fitness = hits

  Size fair crossover, mutations. Population size 5000, 50 generations

  Best of five runs

• Best individual across the 3 classes, was simplified
  (size reduced from 93 to 75)

P450 IC50 training v. test

Best of 5 regression runs and 5 classifications runs are plotted together with best simplified model of both GP approaches.

P450 IC50 (simplified) Regression Model

F1, f2, ... f8 are GSK domain specific features calculated for each chemical from its chemical formula. These 8 features were chosen by GP from the 121 available.

P450 IC50 (simplified) Classification Model

Active

Substrate

Inactive

Each tree yields a value. The tree whose value is largest of the three indicates the predicted class. 21 features were chosen by GP from the 121 available.

GSK Workshop Recommendations

• Genetic programming, artificial neural networks and logistic regression all showed reasonable predictivity, but far from ideal. Explore further.

• Follow up on GP methods. Make technology available (software and assistance)

• GP produced a more easily understood model,

  using small number of features,

  and those features made sense to p450 modelling experts.

Conclusions

• Chemists want to extrapolate. They want to predict properties of chemicals they do not know.

• Chemists like to see your models.

  If it fits on a slide put it on slide!

• In this trial, on real pharmaceutical data, genetic programming was the most able to extrapolate and yielded the only comprehensible models.

• Evolved simple models are plain text and can be readily ported into other tools (spreadsheets, C++, SQL) with negligible overhead.