to phenotypes is highlighted by our observa-
tion of how the effect of a given single muta-
tion depends on genetic context. For example,
CBTPA and CBTP3 melanocytes in vitro were
more similar in expression to CBTP melano-
cytes than to CBTA or CBT3 melanocytes, sug-
gesting thatPTENloss modulates the effect
ofAPCandTP53loss on gene expression. In
another example, aTP53mutation produced
different effects depending on which other
mutations were present: CBT3 melanocytes in
vivo did not form sizable primary tumors, but
CBTP3 tumors grew faster than CBTP tumors,
reflecting an interaction betweenPTENand
TP53mutations ( 68 ) (although no significant
association between these two mutations has
been noted in patient melanomas) ( 16 , 24 , 69 ).
These nonadditive effects also extend to the
tumor microenvironment. For example, the
abundance of neutrophils in CBTA and CBTP3
tumors and their absence or near absence in
CBTP and CBTPA tumors was not attributable
to a difference in a single mutation, as the
two tumor groups shared many mutations.
Hodiset al.,Science 376 , eabi8175 (2022) 29 April 2022 11 of 14
A
B
E
CD
Predict on patient-derived melanoma pathology slides (TCGA)
Classification
Genotype:
Prediction:
Tile prediction:
CBTPA
CBTPA
CBT
Unassigned
CBTP
CBTP3
Correct
CBTP CBTP3 CBTPA Unassigned
Unassigned Incorrect
True positive rate
Genotype
False positive rate Predicted genotype
1.0
1.0 CBTPA
0.8
0.8 CBTP3
0.6
0.6 CBTP
0.4
0.4 CBTA
0.2
0.2 CBT3
0.0
0.0 CBT
True positive rate
False positive rate
1.0
1.0
0.8
0.8
0.6
0.6
0.4
0.4
0.2
0.2
0.0
0.0
False positive rate
0.0 0.2 0.4 0.6 0.8 1.0
False positive rate
0.0 0.2 0.4 0.6 0.8 1.0
Random (AUC 0.50)
CBT (AUC 0.99)
CBT3 (AUC 1.00)
CBTA (AUC 1.00)
CBTP (AUC 0.89)
CBTP3 (AUC 0.90)
CBTPA (AUC 1.00)
Random (AUC 0.50)
APC or CTNNB1 mutant
(AUC 0.58)
OxPhos (AUC 0.74)
Random (AUC 0.50)
TP53 mutant (AUC 0.52)
MITF-low or Common
(AUC 0.63)
Random (AUC 0.50)
PTEN mutant (AUC 0.53)
Predict on held-out melanoma models
Predict TP53 loss-of-function
p53-mutants or -associated expression programs
Predict PTEN loss-of-function
PTEN-mutants or -associated expression programs
Predict APC loss-of-function
Wnt-mutants or -associated expression programs
Confusion matrix
CBTPA CBTP3 CBTP CBTA CBT3 CBT 0
20
40
60
80
100
Percentage
of genotype’s
predictions (%)
(X) : # of
predictions
Mutant melanocyte
tumors
BBatch normalization RReLU SSoftmax
Tumor
section
Tile
2048 × 2048
Overlapping
patches
3 × 512 × 512
INPUT ENCODER
(pre-trained on ICIAR2018)
CNN ARCHITECTURE OUTPUT
(trained on xenograft tumor sections)
{CBT, CBT3, CBTA,
CBTP, CBTP3, CBTPA}
Predicted (tile)
Genotype
(probability vector)
Predicted (section)
Genotype
(probability vector)
Predicted (section)
Genotype label
S
Cross
entropy
Back
propag. TRAINING LABEL
Ground Truth (tile) Genotype Labels
3x 3x CBTA
(^64128256)
32
conv 3
×^3
2x
16
conv 2
×^2
conv 3
×^3
conv 2
×^2
conv^3
×^3
conv^2
×^2 conv 3 × 3 conv 3× 3
B
R
B
R
B
R
B
R
B
R
B
R
B
R
B
R conv 1
×^1
2x
2x
Feature
maps
64 × 64
2x
(^64128)
conv^3
×^3 conv 2 ×^2 conv 3 × 3
B
R
B
R
B
R
2x
conv 2 × 2 B
R conv 1
×^1
CBT • CBT3 • CBTA • CBTP • CBTP3 • CBTPA
Sum
vectors
Max
value
Fig. 6. Tumor genotype leads to distinct histological features that also
associate with genotype-linked expression states in patient melanomas.
(A) Computational approach to classify histological slides into engineered
genotypes. (B) Test set classification examples. Classification of individual tiles
(colored squares overlying tissue images), the aggregated classification for the entire
section (“prediction”), and the true genotype (“genotype”) for three examples.
(CandD) Successful prediction of genotype from histology in held-out mutant
melanocyte in vivo tumor section images. (C) Receiver operating characteristic
(ROC) curves of the prediction false positive rate (xaxis) and true positive rate
(yaxis) at each probability threshold for each genotype (color). The area under the
curve (AUC) is indicated for each genotype in the legend. (D) Percentage (color bar)
of samples from a given genotype (yaxis) that received each genotype classification
(xaxis). The percentage and number of such predictions are displayed within
each cell. (E) Inferring genotype and genotype-associated expression states in
patient melanomas (from TCGA) on the basis of images of H&E stained tumor
sections. ROC curves were obtained through: (left) predictingAPCloss-of-function
genotype (a Wnt pathway gene) and setting true positive labels to be either Wnt
pathway mutants or a Wnt-associated expression program; (middle) predicting
TP53loss-of-function genotype and setting true positive labels to beTP53mutants
orTP53-associated expression programs; or (right) predictingPTENloss-of-
function genotype and setting true positive labels to bePTENmutants (no
PTEN-associated expression program available).
RESEARCH | RESEARCH ARTICLE