Artificial Intelligence in Structural Biology
Sergei Grudinin
Nano-D – LJK, UMR 5224, Inria Grenoble - Rhône-Alpes - CNRS [email protected]
team.inria.fr/nano-d/
SD-IA et biologie, Dec 2, 2020 Sergei Grudinin
Binding Free Energy
!2
Outline
Machine learning Physics-based modeling
SD-IA et biologie, Dec 2, 2020
Sergei Grudinin !3
training, ~60,000 cases
test, ~ 10,000 cases
Classical machine-learning example
SD-IA et biologie, Dec 2, 2020
Sergei Grudinin !4
training, ~ 10,000 structures test
native non-native non-native
native non-native non-native
?
Can we transfer it to 3D protein structures?
SD-IA et biologie, Dec 2, 2020
Sergei Grudinin !5
structures features
Feature extraction
need to be orientation- and translation- equivariant!
Yes we can!
Classification (or regression)
SD-IA et biologie, Dec 2, 2020 Sergei Grudinin
min f .Regularization(f) + Misclassification(f(r), v c (r))
nonnative vectors
native vectors
A
native vectors
B
nonnative vectors nonnative
vectors
native vectors
A
native vectors
B
nonnative vectors nonnative
vectors
native vectors
A
native vectors
B
nonnative vectors
v 1
v 2
!6
SD-IA et biologie, Dec 2, 2020
Sergei Grudinin !7
Deep Learning
๏ Multiple layers that progressively extract features on different scales
๏ They can learn a hierarchy of representations that correspond to different levels of abstraction
๏ Deep learning is effective at problems with hierarchical and structured data. Deep learning is not particularly suited to problems with unstructured data
Amino acids
SD-IA et biologie, Dec 2, 2020
Sergei Grudinin !8
What is the right protein abstraction?
Sequence / MSA profile
many classical ML methods
Secondary structure elements Distance / HB / Contact matrix
state-of-the-art structure prediction methods
Molecular graph
Fout et al. NIPS 2017; Ingraham et al. NIPS 2019; Igashov et al 2020
Set of balls / Point cloud
classical statistical potentials
Gaussian mixture
Derevyanko et al. Bioinformatics 2018; Pages et al. Bioinformatics 2019;
Molecular surface
Correia, Bronstein et al. Nat Met 2020
3D tessellation
Igashov et al. Bioinformatics (submitted)
2020
DeepCOV: Analysing Residue Covariation using FCNs
8-10 Padded
Convolutional layers 5x5x64 Filters
ReLu Activation BatchNorm
Input: 21x21 covariance feature
channels
Convolutional Maxout Layer (dimensionality reduction to 64 channels)
Output:
Contact
Probability Map
from Prof. David Jones, UCL Dept. of Computer Science & The Francis Crick Institute, http://predictioncenter.org
covariance between MSA columns
CASP 13 :
SD-IA et biologie, Dec 2, 2020
Sergei Grudinin !10
Deeper Networks
(Deep Residual Networks. Kaiming He. 2016)
6 / 42
CASP 13 : Revolution of Depth
CASP 13 : Key Developments
SD-IA et biologie, Dec 2, 2020
Sergei Grudinin !11
๏ 1989, 1998, Yann LeCun's back-propagation and convolutional kernels
๏ 2006, Hinton’s Layer by Layer training of Deep Belief Nets
๏ 2010, Acceleration by GPUs (CUDA/theano)
๏ 2011, Rectified Linear Units
๏ 2015, Batch Normalization
๏ 2016, Residual nets
Key Developments
• 2006, Hinton’s Layer by Layer training of Deep Belief Nets
• 2011, Rectified Linear Units
• 2015, Batch Normalization
• 2016, Residual nets
• 2010, Acceleration by GPUs (CUDA/theano)
Turned out that this is all we needed to
change to train deep nets!
INPUT 32x32
Convolutions Subsampling Convolutions
C1: feature maps 6@28x28
Subsampling S2: f. maps 6@14x14
S4: f. maps 16@5x5
C5: layer 120
C3: f. maps 16@10x10
F6: layer 84
Full connection
Full connection
Gaussian connections OUTPUT
10
8 > <
> :
<latexit sha1_base64="DbItFP9JJG+nD7fI64MaaCH8dRg=">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</latexit><latexit sha1_base64="5Sz/GxuDddQlZcdUHRc5NskOPH4=">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</latexit><latexit sha1_base64="5Sz/GxuDddQlZcdUHRc5NskOPH4=">AAAEhXichVPdahNBFJ62UWv8S9U7bwZDoUgJ2bRovRALXqh3FYwtZGKZnT3ZDJmfZWY2bTqsL+C7CN7qk3jpmzi7CTWbCA7scuab853/E2eCW9ft/trY3GrcuHlr+3bzzt179x+0dh5+sjo3DPpMC23OYmpBcAV9x52As8wAlbGA03jypnw/nYKxXKuPbpbBUNJU8RFn1AXovLVLYki58izYsEXzS5OQ6x+oZIGft9rdTrc6eF2IFkL7uHH1++23x/7kfGfrPUk0yyUoxwS1dhD1Mjf01DjOBBRNklvIKJvQFAb0MKMZmKFPQUtwZlY0d1fYjk+u9mNtEjCvjqQceuuoSqjQCgpcMxZERSXYoa+KU+DdgCR4pE34lMMVuszwVFo7k3HQlNSN7epbCf7rbZC70dHQc5XlDhSbOxrlAjuNy0rjhBtgTsyCQJnhIXHMxtRQ5kI/mkTBBdNShjSIgWQQDb0nVXA+XIt2VBQ1nVjksKxU3hdaS/GGgqSOxisV9ka70G+V1lGZC8eNvqh3I5nyzC5KeDmvYWhHUCh7IHhsqJn5TFteDlBlcznOZ55Mp0UZKJkC80ROwKiuzFfzCc+JdsWgFxQFjBwRVKUCcDvax+0eJoanY0dMBdY9eJKFKSgwxhV5FAoaoHKwqAhe/sq9VY+ZBVZS13mffa+oc0tkjc+MtpWBg/8buJYO6i0aTMtVTcIwX5pcQBJ2QqQ6jMdY9la6VhY85B6WL1pdtXWh3+u87EQfovbxHpqfbfQEPUV7KEIv0DF6h05QHzH0FX1HP9DPxnaj0zhsPJ+rbm4sOI9Q7TRe/wFUH4+N</latexit><latexit sha1_base64="yHuXOmKcG72f0oVEy/n3F5GAMFg=">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</latexit>
this is all we needed to train deep
nets!
LeCun et al. Proc of the IEEE, 1998
SD-IA et biologie, Dec 2, 2020
Sergei Grudinin !12
CASP14
๏ Use of metagenomics and structural templates
๏ No DCA / PCA coevolution - raw MSA profiles
๏ Attention networks - learn dynamic ‘soft’ molecular graphs. Pass messages along edges that have matching ‘keys’
๏ Proper use of 3D geometry, stereochemistry, equivariance, and local 3D frames
๏ Uncertainty prediction. The same idea used by Rosetta / Baker team
๏ Use of ‘distograms’ instead of contacts and histograms
๏ Training of one model takes ~200 GPUs (128 TPUs and lots of memory!) for several
weeks. This must be multiplied by ~100. So, we get ~ 10 Jean Zay running for 1 month.
GenBank WGS
(whole genome sequencing)
Moore’s Law
SD-IA et biologie, Dec 2, 2020
Sergei Grudinin !13
The amount of data (genomic and structural) grows faster than our computational resources
ML in structural biology
SD-IA et biologie, Dec 2, 2020
Sergei Grudinin !14
We constantly need better and faster algorithms for integrating growing amount of data
The amount of data (genomic and structural) grows faster than our computational resources
5 Å
5 Å 0
NOLB NMA method AnAnaS - analytical
analyser of symmetries SAM - FFT-accelerated
symmetry assembler
C 2 axis C 2 axis
C 2 axis
y x
z
A B
C
A' C'
B'
R z (η)
T z (E)
Pepsi-SAXS / SANS methods
training
1. Geometric features extraction
non-native distributions
native distributions
2. Projection 3. Optimisation
scoring vector w
4. atom-atom distance-dependent 1D interaction potentials u
r
u
r
u
r
potentials
x kl q = r Z max
0 n kl (r) q (r)
scoring
5. Geometric features extraction
distributions
6. Projection 7. Ranking
x kl q = w
r Z max
0 n kl (r) q (r)
1.
...
N. 2
3
1 E 2 E 3
E 1
structure vectors
Pipeline for KB-potentials
Docking and fitting methods
ML in structural biology
SD-IA et biologie, Dec 2, 2020
Sergei Grudinin !15
We constantly need better and faster algorithms for integrating growing amount of data
The amount of data (genomic and structural) grows faster than our computational resources
We need to develop novel machine-learning
approaches specifically adapted to our data (rather than adapt the data to existing ML and DL approaches)
Do we understand physics?
yes no
G. Pages et al. Bioinformatics 2019, bty1037
M. Karasikov et al.
Bioinformatics 2019, btz122
ML in structural biology
SD-IA et biologie, Dec 2, 2020
Sergei Grudinin !16
We constantly need better and faster algorithms for integrating growing amount of data
The amount of data (genomic and structural) grows faster than our computational resources
We need to develop novel machine-learning
approaches specifically adapted to our data (rather than adapt the data to existing ML and DL approaches)
ML in structural biology
DL models are interpretable!
G. Pages & S. Grudinin, unpublished
SD-IA et biologie, Dec 2, 2020
Sergei Grudinin !17
We constantly need better and faster algorithms for integrating growing amount of data
The amount of data (genomic and structural) grows faster than our computational resources
We need to develop novel machine-learning
approaches specifically adapted to our data (rather than adapt the data to existing ML and DL approaches)
✓
<latexit sha1_base64="MqBtYtWURDS0UTR7zMyC7mDaPIs=">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</latexit><latexit sha1_base64="HSxwWxT6PykmSdcazj8XgvBcoVk=">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</latexit><latexit sha1_base64="HSxwWxT6PykmSdcazj8XgvBcoVk=">AAAEfHichVPNbtNAEN7SACX8tXBEQiuiSghKFbsIygFRiQvHViJtpdhU6/XEWbI/1u46bVj5yJ0rPElPvAfPwEPA2okgdpBYydL4m/nmf5KcM2P7/R9rV9Y7V69d37jRvXnr9p27m1v3jo0qNIUBVVzp04QY4EzCwDLL4TTXQETC4SSZvK30J1PQhin53s5yiAXJJBsxSqyHjiM7BkvONnv93X798KoQLITem++XRz8/P7w8PNtaJ1GqaCFAWsqJMcMgzG3siLaMcii7UWEgJ3RCMhiS5znJQccuAyXA6lnZ3W6xLZt82kmUTkG/3hcidsYSmRKuJJS44cyLkggwsatrL/G2R1I8Utp/0uIaXWY4IoyZicRbCmLHpq2rwH/phoUd7ceOybywIOk80Kjg2CpcNRKnTAO1fOYFQjXzhWM6JppQ69vdjSScUyWELyPSkA6D2LmoTs7537IXlGXDJuEFLBtV/wurpXx9QzJLklaHnVbWj1NmTVQU3DKtzpvTSKcsN4sWXsx76MfhDaoZcJZoomcuV4ZV+1H7XM7ziYum07JKNJoCdZGYgJZ9UbTr8epU2XIYekMOIxtxIjMOuBfs4F6II82ysY10DTYjuCj3W1BijGvyyDfUQ9ViEe6j/JXDdsTcAK2oq7wPLiyb3ApZ4VOtTO1g7/8O/kh7zRENp9Ulpn6ZL3TBIfU3wTPl12MswtbUqoZXtS9BCZ8QrYm/kG7XH2XQPsFVYRDuvtoNjvq9g300fxvoAXqEHqMAvUQH6B06RANE0Uf0BX1F39Z/dbY7TzvP5qZX1hac+6jxOi9+AyT+jxo=</latexit><latexit sha1_base64="MUE8LJPyt0ejA9dzEVYiuES72hk=">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</latexit>r
<latexit sha1_base64="yrJ2yfrW9xcVdvfVyRgDyc+KRx4=">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</latexit><latexit sha1_base64="23jKwH37r5FJFJq5f1RYrdcCMRg=">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</latexit><latexit sha1_base64="23jKwH37r5FJFJq5f1RYrdcCMRg=">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</latexit><latexit sha1_base64="38o/evtAuIDFPhXyXQC9ZCGpo60=">AAAEd3ichVNLb9NAEN62AUp4tIUjlxVRUYWqKnaRCAekSlw4thKhlWJTrdcTZ5V9WLvrtGHlX8AVfhx/hRO7bgSxg8RKlsbfzDfvyUrOjB0Of25t7/Tu3X+w+7D/6PGTp3v7B88+G1VpCmOquNJXGTHAmYSxZZbDVamBiIzDZTb/EPSXC9CGKfnJLktIBSkkmzJKrIcu9PX+YHgybB7eFKKVMECrd359sEOSXNFKgLSUE2MmUVza1BFtGeVQ95PKQEnonBQwIW9KUoJOXQFKgNXLun/YYVs2/3qcKZ2Dfj8SInXGEpkTriTUuOXMi5IIMKlryq7xoUdyPFXaf9LiBl1nOCKMWYrMWwpiZ6arC+C/dJPKTkepY7KsLEh6F2hacWwVDj3EOdNALV96gVDNfOGYzogm1PpO9xMJN1QJ4ctINOSTKHUuaZJz/rceRHXdssl4BetG4X9ltZavb0hhSdbpsNPK+knKoo2Kilum1U17GvmClWbVwtu7HvpxeIMwA84yTfTSlcqwsBqNz/U8X7tksahDoskCqEvEHLQciqpbj1fnytaT2BtymNqEE1lwwIPoGA9inGhWzGyiG7AdwSWl34IaY9yQp76hHgqLRbiP8leOuxFLAzRQN3lfXFy3uQHZ4FOtTOPg9P8O/kin7RFNFuEIc7/Mt7rikPub4IXy6zETcWdqoeGh9jUo43OiNfEX0u/7o4y6J7gpjOOTdyfRxXBwNlpd5y56gV6iIxSht+gMfUTnaIwoAvQNfUc/dn71cO9V7+jOdHtrxXmOWq8X/QZrcYic</latexit>y
<latexit sha1_base64="AEQNWXBbAD4xugaevfvCOzFwKzc=">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</latexit><latexit sha1_base64="yHo5MUNrp2rlqhkC1bJO5gezlW4=">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</latexit><latexit sha1_base64="yHo5MUNrp2rlqhkC1bJO5gezlW4=">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</latexit><latexit sha1_base64="Otxb5hS3wS61TlwqH4zRzHSYJQk=">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</latexit>
z
<latexit sha1_base64="dmNPyKvjqbKcwFWVUo+B2RO4bmI=">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</latexit><latexit sha1_base64="VNv5TvaRTo4bI7RCJqz/mjxuk+A=">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</latexit><latexit sha1_base64="VNv5TvaRTo4bI7RCJqz/mjxuk+A=">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</latexit><latexit sha1_base64="Gref2oJrWutTEZEUIILdsNdQEvQ=">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</latexit>
x
<latexit sha1_base64="39pI6bN6jmNimHchBDuwXUXmRq4=">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</latexit><latexit sha1_base64="6GjwOOD+CkuWh8KpTs/2LQfxK7c=">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</latexit><latexit sha1_base64="6GjwOOD+CkuWh8KpTs/2LQfxK7c=">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</latexit><latexit sha1_base64="UCrZrBRa6rIlpftTzgHj1C+lN/w=">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</latexit>
C
N
C N
O O
φ
C α
KORP-PL – the state-of-the-art virtual screening potential, Kadukova et al. Bioinformatics 2020
learning spherical kernels on 3D graphs, Igashov et al. arXiv 2020
N C α
R
N
C
O C α
R N
C O
C α
R
C N O
C α
R
N
C O
C α
R N
C O C α
R
C
O C α
R
✓
<latexit sha1_base64="3Uxwoi/CcjBq07QEVEZs45Auhwk=">AAAB7HicdVDLSgNBEOyNrxhfUY9eBoPgadkNidkcxIAXjxHMA5IQZieTZMzsg5leIYSAn+DFg4pXf8H/8ObfOEkUVLSgoajqpqvbj6XQ6DjvVmppeWV1Lb2e2djc2t7J7u7VdZQoxmsskpFq+lRzKUJeQ4GSN2PFaeBL3vBH5zO/ccOVFlF4heOYdwI6CEVfMIpGqrdxyJF2sznH9kpuoVgiju16J17Rm5NC2SsR13bmyJ29Zk5vAaDazb61exFLAh4ik1TrluvE2JlQhYJJPs20E81jykZ0wFuGhjTgujOZp52SI6P0SD9SpkIkc/X7xIQGWo8D33QGFIf6tzcT//JaCfa9zkSEcYI8ZItF/UQSjMjsdNITijOUY0MoU8JkJWxIFWVoHpQxT/i6lPxPanm7bLuXTq6ShwXScACHcAwulKACF1CFGjC4hjt4gEcrsu6tJ+t50ZqyPmf24Qeslw/5XpEO</latexit><latexit sha1_base64="/nv1i/0LeNCR+DxB5disc1cA8cQ=">AAAB7HicdVDLSgNBEJz1GddX1KOXwSB4WnZD4m4OYsCLxwjmAUkIs5NJMmb2wUyvEJb8gxcPKp4Ef8H/8CL+jZONgooWNBRV3XR1+7HgCmz73VhYXFpeWc2tmesbm1vb+Z3dhooSSVmdRiKSLZ8oJnjI6sBBsFYsGQl8wZr++GzmN6+ZVDwKL2ESs25AhiEfcEpAS40OjBiQXr5gW57rlMouti3HO/bKXkZKFc/FjmVnKJy+mCfx05tZ6+VfO/2IJgELgQqiVNuxY+imRAKngk3NTqJYTOiYDFlb05AETHXTLO0UH2qljweR1BUCztTvEykJlJoEvu4MCIzUb28m/uW1Exh43ZSHcQIspPNFg0RgiPDsdNznklEQE00IlVxnxXREJKGgH2TqJ3xdiv8n9aJVsZwLu1AtojlyaB8doCPkIBdV0TmqoTqi6ArdoDt0b0TGrfFgPM5bF4zPmT30A8bzB+l5koI=</latexit><latexit sha1_base64="/nv1i/0LeNCR+DxB5disc1cA8cQ=">AAAB7HicdVDLSgNBEJz1GddX1KOXwSB4WnZD4m4OYsCLxwjmAUkIs5NJMmb2wUyvEJb8gxcPKp4Ef8H/8CL+jZONgooWNBRV3XR1+7HgCmz73VhYXFpeWc2tmesbm1vb+Z3dhooSSVmdRiKSLZ8oJnjI6sBBsFYsGQl8wZr++GzmN6+ZVDwKL2ESs25AhiEfcEpAS40OjBiQXr5gW57rlMouti3HO/bKXkZKFc/FjmVnKJy+mCfx05tZ6+VfO/2IJgELgQqiVNuxY+imRAKngk3NTqJYTOiYDFlb05AETHXTLO0UH2qljweR1BUCztTvEykJlJoEvu4MCIzUb28m/uW1Exh43ZSHcQIspPNFg0RgiPDsdNznklEQE00IlVxnxXREJKGgH2TqJ3xdiv8n9aJVsZwLu1AtojlyaB8doCPkIBdV0TmqoTqi6ArdoDt0b0TGrfFgPM5bF4zPmT30A8bzB+l5koI=</latexit><latexit sha1_base64="fp18B+pRyUMe4szTWmAu/JGtAxQ=">AAAB7HicdVBNS8NAEN3Ur1q/qh69LBbBU0hKa9JbwYvHCqYW2lA22027dvPB7kQoof/BiwcVr/4gb/4bt2kFFX0w8Hhvhpl5QSq4Asv6MEpr6xubW+Xtys7u3v5B9fCoq5JMUubRRCSyFxDFBI+ZBxwE66WSkSgQ7DaYXi7823smFU/iG5ilzI/IOOYhpwS01B3AhAEZVmuW6Tp2o+lgy7TdC7fpFqTRch1sm1aBGlqhM6y+D0YJzSIWAxVEqb5tpeDnRAKngs0rg0yxlNApGbO+pjGJmPLz4to5PtPKCIeJ1BUDLtTvEzmJlJpFge6MCEzUb28h/uX1MwhdP+dxmgGL6XJRmAkMCV68jkdcMgpipgmhkutbMZ0QSSjogCo6hK9P8f/Eq5st0762au36Ko0yOkGn6BzZyEFtdIU6yEMU3aEH9ISejcR4NF6M12VryVjNHKMfMN4+AaGTj1E=</latexit>'
<latexit sha1_base64="io9MQQyiqiFibQVbTBcFlJClI28=">AAAB7XicdVBNS8NAEJ3Ur1qrVj16WSyCp5CktTW3ghePLRhbaEPZbLft0s0Hu5tCCf0RXjyoePUv+Du8efOnuG0VVPTBwOO9GWbmBQlnUlnWm5FbW9/Y3MpvF3aKu3v7pYPDGxmnglCPxDwWnQBLyllEPcUUp51EUBwGnLaDyeXCb0+pkCyOrtUsoX6IRxEbMoKVltq9KRbJmPVLZcu0nZrr1pBlOpWaXXc0cSvVc7uKbNNaotwovrTeAaDZL732BjFJQxopwrGUXdtKlJ9hoRjhdF7opZImmEzwiHY1jXBIpZ8tz52jU60M0DAWuiKFlur3iQyHUs7CQHeGWI3lb28h/uV1UzW88DMWJamiEVktGqYcqRgtfkcDJihRfKYJJoLpWxEZY4GJ0gkVdAhfn6L/ieeYrmm3dBgOrJCHYziBM7ChDg24giZ4QGACt3APD0Zi3BmPxtOqNWd8zhzBDxjPH0zrkes=</latexit><latexit sha1_base64="cMV1U9AcWEv+8JfQfRkiFB2Ekvg=">AAAB7XicdVDLSsNAFJ34rLVq1aUgg0VwFZK0tmZXcOOyBWMLbSiT6aQdOpmEmUmhhC79ADcuVNz6C/0Od36DP+G0VVDRAxcO59zLvfcECaNSWdabsbK6tr6xmdvKbxd2dveK+wc3Mk4FJh6OWSzaAZKEUU48RRUj7UQQFAWMtILR5dxvjYmQNObXapIQP0IDTkOKkdJSqztGIhnSXrFkmbZTdd0qtEynXLVrjiZuuXJuV6BtWguU6oVZ8/32eNboFV+7/RinEeEKMyRlx7YS5WdIKIoZmea7qSQJwiM0IB1NOYqI9LPFuVN4qpU+DGOhiyu4UL9PZCiSchIFujNCaih/e3PxL6+TqvDCzyhPUkU4Xi4KUwZVDOe/wz4VBCs20QRhQfWtEA+RQFjphPI6hK9P4f/Ec0zXtJs6DAcskQNH4AScARvUQB1cgQbwAAYjcAcewKORGPfGk/G8bF0xPmcOwQ8YLx8pypNR</latexit><latexit sha1_base64="cMV1U9AcWEv+8JfQfRkiFB2Ekvg=">AAAB7XicdVDLSsNAFJ34rLVq1aUgg0VwFZK0tmZXcOOyBWMLbSiT6aQdOpmEmUmhhC79ADcuVNz6C/0Od36DP+G0VVDRAxcO59zLvfcECaNSWdabsbK6tr6xmdvKbxd2dveK+wc3Mk4FJh6OWSzaAZKEUU48RRUj7UQQFAWMtILR5dxvjYmQNObXapIQP0IDTkOKkdJSqztGIhnSXrFkmbZTdd0qtEynXLVrjiZuuXJuV6BtWguU6oVZ8/32eNboFV+7/RinEeEKMyRlx7YS5WdIKIoZmea7qSQJwiM0IB1NOYqI9LPFuVN4qpU+DGOhiyu4UL9PZCiSchIFujNCaih/e3PxL6+TqvDCzyhPUkU4Xi4KUwZVDOe/wz4VBCs20QRhQfWtEA+RQFjphPI6hK9P4f/Ec0zXtJs6DAcskQNH4AScARvUQB1cgQbwAAYjcAcewKORGPfGk/G8bF0xPmcOwQ8YLx8pypNR</latexit><latexit sha1_base64="4ssD4Da9yVs2Qy2Wq8qTocxeopY=">AAAB7XicdVBNS8NAEJ34WetX1aOXxSJ4CklaW3srePFYwdhCG8pmu2mXbjZhd1MooT/CiwcVr/4fb/4btx+Cij4YeLw3w8y8MOVMacf5sNbWNza3tgs7xd29/YPD0tHxvUoySahPEp7ITogV5UxQXzPNaSeVFMchp+1wfD332xMqFUvEnZ6mNIjxULCIEayN1O5NsExHrF8qO7br1RqNGnJsr1Jz654hjUr10q0i13YWKMMKrX7pvTdISBZToQnHSnVdJ9VBjqVmhNNZsZcpmmIyxkPaNVTgmKogX5w7Q+dGGaAokaaERgv1+0SOY6WmcWg6Y6xH6rc3F//yupmOroKciTTTVJDloijjSCdo/jsaMEmJ5lNDMJHM3IrICEtMtEmoaEL4+hT9T3zPbtjurVNueqs0CnAKZ3ABLtShCTfQAh8IjOEBnuDZSq1H68V6XbauWauZE/gB6+0TVNaPtg==</latexit>convolutional neural networks on irregular 3D tessellations, Igashov et al., Bioinformatics (submitted) 2020
1 2 3
DL models are interpretable!
ML in structural biology
We can use better abstraction and better geometry!
SD-IA et biologie, Dec 2, 2020
Sergei Grudinin !18
We constantly need better and faster algorithms for integrating growing amount of data
The amount of data (genomic and structural) grows faster than our computational resources
We need to develop novel machine-learning
approaches specifically adapted to our data (rather than adapt the data to existing ML and DL approaches)
Current data allows to reconstruct and/or learn structural heterogeneity and motion manifolds
manifold learning
automatic selection of representation
4
7
6
5 3
F
ML in structural biology
DL models are interpretable!
We can use better abstraction and better geometry!
near future
SD-IA et biologie, Dec 2, 2020
Sergei Grudinin !19
We constantly need better and faster algorithms for integrating growing amount of data
The amount of data (genomic and structural) grows faster than our computational resources
We need to develop novel machine-learning
approaches specifically adapted to our data (rather than adapt the data to existing ML and DL approaches)
Which leads to predicting protein function!
because function is linked with the shape and the motion!
ML in structural biology
Current data allows to reconstruct and/or learn structural heterogeneity and motion manifolds DL models are interpretable!
We can use better abstraction and better geometry!
SD-IA et biologie, Dec 2, 2020
Sergei Grudinin !20
We constantly need better and faster algorithms for integrating growing amount of data
The amount of data (genomic and structural) grows faster than our computational resources
We need to develop novel machine-learning
approaches specifically adapted to our data (rather than adapt the data to existing ML and DL approaches)
30 40 50 60 70 | . . . | . . . | . . . | . . . | . A Y D V E A E AV L Q AV Y E T E S A F D L A M R I M W I Y V F C F N R P I P F P H A
A C D E F G
A C D E F
With the ultimate goal of routine computational protein design
The Future goals
Which leads to predicting protein function!
Current data allows to reconstruct and/or learn structural heterogeneity and motion manifolds DL models are interpretable!
We can use better abstraction and better geometry!
SD-IA et biologie, Dec 2, 2020
Sergei Grudinin !21
๏ Example 1 - long-range interactions
๏ Divergence theorem, we can learn on a manifold
๏ Example 2 - message-passing algorithms
๏ FMM was invented in 1987, and can be reused in learning on graphs and point clouds
๏ Example 3 - rotational invariance / equivariance
๏ Can be represented using Spherical Harmonics and Wigner rotation matrices
˚
V
( r · F) dV =
‹
S
(F · n) dS
<latexit sha1_base64="B1yKSxvMe2c5Ot9esh1C0tC/1o8=">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</latexit><latexit sha1_base64="uDgIF1jpuMOxpyED+cuF8b8Zixg=">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</latexit><latexit sha1_base64="uDgIF1jpuMOxpyED+cuF8b8Zixg=">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</latexit><latexit sha1_base64="bjnsR7Zy8xrfDM4F5QlHZxSVL0M=">AAAChnicbVHfi9NAEN7EX3fR056CPviyWIQKUhPhOH0QioL4eHK2d9AtZbKZtMttNmF3IoSQB/9M/wIf/RfctEXsnQOzfHwz38zwbVpp5SiOfwbhrdt37t47OIzuPzh6+Ghw/HjmytpKnMpSl/YyBYdaGZySIo2XlUUoUo0X6dWnvn7xHa1TpflGTYWLAlZG5UoCeWo5+CGUUob4sp11QmNOI1EArdOct8JAqoF3XMisJP6X/+wZq1ZreiVeZ7MPkSi3E86jVmTKVRoaR41GPtqT7A8xHe/l5120HAzjcbwJfhMkOzBkuzhbDn6LrJR1gYakBufmSVzRogVLSmrsIlE7rEBewQrnHhoo0C3ajVUdf+mZjOel9emP3rD/KloonGuK1Hf2p7rrtZ78X21eU/5u0SpT1YRGbhflteZU8t53nimLknTjAUir/K1crsGCJP87e1vQeTv9AP/iyoJ23qHkuh83wfTt+P04+RoPJx93Vh2w5+wFG7GEnbIJ+8LO2JRJ9is4Cp4Gz8LD8E14Ep5uW8Ngp3nC9iKc/AH8IcPi</latexit>
A B C D
R(↵, , )f (!) = ˆ X
l,m
Y l m (!) X
m 0
D mm (l) 0 (↵, , ) Z
Y l m 0 ⇤ (! 0 )f (! 0 )d! 0
<latexit sha1_base64="0iuqtefr7BU1YRIGwGWErEbBBbI=">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</latexit><latexit sha1_base64="0iuqtefr7BU1YRIGwGWErEbBBbI=">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</latexit><latexit sha1_base64="0iuqtefr7BU1YRIGwGWErEbBBbI=">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</latexit><latexit sha1_base64="0iuqtefr7BU1YRIGwGWErEbBBbI=">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</latexit>
