candle.md

author	title
Rick Stevens	CANDLE Project KPP-1 Verification

Science Challenge Problem Description

The CANDLE challenge problem is to solve large-scale machine learning (ML) problems for two cancer-related pilot applications: (1) predicting drug interactions and (2) predicting cancer phenotypes and treatment trajectories from patient documents. The CANDLE pilot application that involves predicting the state of molecular dynamics simulations is treated as a stretch goal. The CANDLE project has specific strategies to address these challenges. For the drug response problem, unsupervised ML methods are used to capture the complex, nonlinear relationships between the properties of drugs and the properties of the tumors to predict response to treatment and therefore develop a model that can provide treatment recommendations for a given tumor. For the treatment strategy problem, semisupervised ML is used to automatically read and encode millions of clinical reports into a form that can be computed upon. Each problem requires a different approach to the embedded learning problem, all of which are supported with the same scalable deep learning code in CANDLE.

The challenge for exascale manifests in the need to train large numbers of models. One need inherent to each of the pilot applications is producing high-resolution models that cover the space of specific predictions individualized in the precision medicine sense. For example, consider training a model that is specific to a certain drug and individual cancer. Starting with 1000 different cancer cell lines and 1000 different drugs, a leave-one-out strategy to create a high-resolution model for all drug-by-cancers requires approximately 1 million models. Yet, these models are similar enough that using a transfer learning strategy in which weights are shared during training in a way that avoids information leakage can significantly reduce the time needed to train a large set of models.

In practice, speedup related to weight sharing can be discussed in the context of the challenge problem in terms of work actually done and the naive work done. Consider work actually done $W_{D}$ as being the number of jobs $J$ multiplied by the actual number of epochs $E$ trained for all stages $s$: $$ W_{D} = \sum_{1}^{s}JE:. $$ A stage is a discrete and naive work done $W_{N}$ being the number of jobs in the last stage multiplied by the number of epochs needed for a model to converge $E_{c}$,: $$ W_{n} = J_{s}E_{c}:. $$ The team can then talk about speedup as being the ratio $W_{N}/W_{D}$.

Several parameters exist when considering accelerated model training via the transfer of weights. These include how many transfer events, how to partition the input data, and how many epochs before a transfer occurs. Additional considerations include what weights to transfer and whether to allow those weights to be updated in subsequent models.

Table: Challenge problem details

---

Functional requirement Minimum criteria

---

Physical phenomena and Deep learning neural networks for cancer: feed-forward, associated models auto-encoder, recurrent neural networks.

Numerical approach, Gradient descent of model parameters, optimization and associated models of loss function, network activation function; regularization, and learning rate scaling methods.

Simulation details Large-scale ML solutions will be computed for the three cancer pilots:

$\phantom{continue}$ Pilot1: Leave-one-out cross validation of roughly 1000 drugs by 1000 cell lines. This involves roughly 1 million models. Partition the drugs and cell lines into $n$ sets and train with those for $e$ epochs, then transfer the weights $w$ to the next set of models, expanding the number of models in each iteration. Each of the models at iteration $i$ can be safely (i.e., avoiding information leakage) used to seed models for iteration $i+1$, where the set of drugs and cell lines in the $i+1$ validation set was not in the training set of the model at iteration $i$.

$\phantom{continue}$ Pilot2: State the identification and classification of one or more RAS proteins binding to a lipid membrane; prediction over time of clustering behavior of key lipid populations that leads to RAS protein binding. RAS proteins are represented in sufficient resolution to model all pairwise interactions within and between proteins. Lipid membranes are represented as continuous density fields of tens of species of lipid concentration. Predictions are trained on the cross product of thousands of simulations, each of which is thousands of time steps, over multiple protein configurations, and performed for a large range of different concentrations.

$\phantom{continue}$ Pilot3: Predicting cancer phenotypes and patient treatment trajectories from millions of cancer registry documents. Thousands of multitask phenotype classification models will be built from defined combinations of descriptive terms extracted from 10,000 curated text training sets. To accelerate model training, the team will use a transfer learning scheme with weight sharing during training.

Demonstration calculation The computation performed at scale will be standard requirements neural network computations, matrix multiplies, 2D convolutions, pooling, and so on. These will be specifically defined by the models chosen to demonstrate transfer learning. The computations performed at scale will require weight sharing.

Demonstration Calculation

Facility Requirements

List significant I/O, workflow, and/or third party library requirements that need facility support (Note: this is outside of standard MPI and ECP-supported software technologies)

Input description and Execution

Describe your problem inputs, setup, estimate of resource allocation, and runtime settings

Problem Artifacts

Describe the artifacts produced during the simulation, e.g., output files, and the mechanisms that are used to process the output to verify that capabilities have been demonstrated.

Verification of KPP-1 Threshold

Give evidence that

1. The FOM measurement met threshold ($>50$)
2. The executed problem met challenge problem minimum criteria