Should you take this bet?

• You get: 1 hour with my (French^*) dictionary
• Test time: I show you words, you tell me if they are French (Y/N)
• You win: €10 per correct guess (but: you pay me €10 for every wrong one)

T-cell diversity & tolerance: a problem of numbers

The immune system's T-cell repertoire must discriminate "self" vs "foreign":

• recognize >10¹⁵ foreign peptides ... ¹
• ... while tolerating the many "self peptides" of healthy human cells

¹ Sewell (2012). Nature Reviews Immunology.

Solution

Two-step process:

1. Make a repertoire with many cells
2. Filter out self-reactive cells: negative selection

Solution?

TwoThree-step process:

1. Make a repertoire with many cells
2. Filter out self-reactive cells: negative selection ("central" tolerance)
3. Plug the leaks (peripheral tolerance)

Solution?

Solution?

Negative selection is not just "leaky", but very incomplete

So if it can never accomplish tolerance, why do negative selection at all?

Yu et al (2015), Immunity

Tolerance ≠ discrimination.

Tolerance ≠ discrimination.

Which of these scenarios holds?

Mission impossible?

Goat cheese: learning by example

fièvre sounds/looks like chèvre

If we can learn which words look like French, can T cells in the thymus learn which peptides look like "self"?

Which of these scenarios holds?

Which of these scenarios holds?

Approach: Artificial Immune Systems (AISs)

AISs are machine learning algorithms inspired by the immune system (like NNs are inspired by the brain).

They can be used as general classification algorithms, but also as models of immunology.

• bio $\rightarrow$ ML
• ML $\rightarrow$ bio
• bio $\rightarrow$ ML $\rightarrow$ bio

A simple model of TCR-sequence recognition

Three ingredients needed:

A simple model of TCR-sequence recognition

Three ingredients needed:

A simple model of TCR-sequence recognition

Three ingredients needed:

A simple model of TCR-sequence recognition

Three ingredients needed:

Simulating negative selection in silico

Example: Xhosa recognition after negative selection on English

Negative selection allows discrimination

Negative selection on 500 English strings (t = 3),
Compare recognition of unseen English and Xhosa strings

Motifs per string, before vs after:	Most frequently recognized:

$\rightarrow$ Motifs distinguish strings they have not seen!

Why discrimination on unseen strings?

Motifs rarely react to both English and Xhosa:

Nodes:

• English
• Xhosa

Edge if >10,000 motifs in common

Concordance (same-language neighbors): 81%

Why discrimination on unseen strings?

This only works if strings are truly different:

Nodes:

• English
• More English

Edge if >10,000 motifs in common

Concordance (same-language neighbors): 50%

Two key requirements for discrimination

Two key requirements for discrimination

1. Appropriate specificity/cross-reactivity
2. Sufficient self-foreign dissimilarity

Complete specificity (t = 6)	Low specificity (t = 1)	Intermediate specificity (t = 3)

Two key requirements for discrimination

1. Appropriate specificity/cross-reactivity
2. Sufficient self-foreign dissimilarity

Summary: what have we learned from languages?

Proof of concept:
Yes, negative selection is a learning algorithm.

• Self-foreign discrimination even on "unseen" sequences
• Even when tolerance is incomplete

But this works only if:

• self and foreign are sufficiently dissimilar
• cross-reactivity is intermediate

Back to self versus foreign peptides

Back to self versus foreign peptides

Use the same model for peptides instead of strings:

TCRs are moderately cross-reactive

Model at t = 4 matches order of magnitude of experimental estimates

¹ Ishizuka et al (2009). J Immunol.
² Legoux et al (2010). J Immunol.   ³ Blattman et al (2002). J Exp Med.   ⁴ Alanio et al (2010). Blood.

"self" and "foreign" peptides are similar

"self" and "foreign" peptides are similar

 HIV peptides are embedded in
 clusters of self peptides.

  Xhosa/English: separate clusters.

$\rightarrow$ self-foreign discrimination will be difficult!

Self-foreign discrimination is difficult

... but remains possible:

What if thymic self peptides are non-random?

• random self peptides are often similar and will delete the same TCRs
  $\rightarrow$ this is inefficient
• some peptides are exchangeable, others are not
  $\rightarrow$ a training set with more non-exchangeable peptides might do better:

Removal of self-reactivity $\neq$ self-foreign discrimination! Does this work?

"Optimal" training peptides help T cells learn

How could the thymus be "optimal"?

Computed "optimal" set is enriched in rare AAs, depleted of common AAs.
Peptides with rare AAs tend to be less exchangeable:

$\rightarrow$ Could enrichment of rare AAs help self-foreign discrimination?

How could the thymus be "optimal"?

Simple bias already improves self-foreign discrimination.

The same is true for other pathogens

• Poor self-foreign discrimination after selection on random peptides
• Improved self-foreign discrimination after selection on biased peptides

Lessons learned:
self-foreign discrimination in the immune system

1. Tolerance ≠ self-foreign discrimination.
And may require different peripheral tolerance mechanisms.

Read more

Acknowledgements

computational-immunology.org

		Johannes Textor Radboud University, the Netherlands
		Can Keşmir Utrecht University, the Netherlands
		Rob de Boer Utrecht University, the Netherlands
		Judith Mandl McGill University, Montréal, Canada

Why better self-foreign discrimination?

Self peptides with low exchangeability less often resemble foreign peptides:

$\rightarrow$ non-exchangeable peptides efficiently remove self-reactive TCRs, but preserve foreign-reactive ones!