What stinks about value investing is we never get to hitch our wagons to anything cool. As Jobs, Brin, and Musk are doing sci-fi stuff, we're stuck worrying about burglar alarm switches. When you've got something like this in your portfolio - because you read your Graham and Dodd and swore a blood oath to it - but the damn thing is just sitting there, always profitable and yet trading below cash, doing nothing month after month as the Nasdaq flies, and all of a sudden your iPhone is having conversations with you, and Elon is crashing space jets, Bezos is flying drones around like the world is some video game, and yet you're stoically hanging in there with this stupid, boring (but really cheap!) switch-maker... don't you idiots know about the value anomaly!
That's why I get obsessed when I find something fun that's got some value characteristics. But it's not just that I enjoy the research process more. It's that tantalizing vision Stephenson's talking about: that there exist companies that've hitched their wagon to potent, hard-to-copy young technologies, and that if you can identify them, you can hitch your wagon to theirs. And if you're right and you bet big (which the value characteristics let you safely do), you could - I'm not saying you'd want to - but you could sit back for the next decade or so, relax, and enjoy the scenery.
Using Computers To Find New Drugs
Simulations Plus (NASDAQ:SLP) makes pharmaceutical software. Its various tools predict how drugs will behave in the body. Certain ones are used early in the drug design process to filter through the gigantic haystack of never-before-synthesized carbon-based molecules for those most likely to be absorbable, non-toxic, excretable, etc. (the ADMET Predictor). Other tools are used later in the process - after the active ingredient has been selected, but before any physical testing has taken place - to guess how fast a drug will move through the body's organs, how fast concentrations will rise and dissipate, and what biological effect a given level of concentration will have (GastroPlus), which is used to pick what kind of filler to clump around a drug, how to administer it (throat, vein, eye), and to predict what dose will help fix the first literal or figurative guinea pigs without, you know, killing them.
SLP is a (or the - it's tough to say, because the few serious competitors are private or small bits of larger public companies) leader in predictive software (i.e., not databases) for drug design, which is nice, because winning early probably means winning permanently. Partly that's because once a scientist has learned how to use a complicated tool, he's not going to switch to a different one, but more than that, it's because the software improves with use. SLP employs just 60-70 people, but each and every customer is a potential source of new tricks, recommendations, and data. So, SLP's gonna be around awhile - if the industry grows 15%/year for a decade, then so should the company. But will it? I'll get to that later.
Value Characteristic = Downside Clarity = Sticky Licenses + Fast Trend Growth + Fat Margins
SLP's P/E is ~33 - but don't run just yet. Yes, there is downside risk - the valuation requires growth - but it's unusually easy to estimate what's the worst that can happen to SLP, and it's not all that bad.
Question: When an EPS report looms, which of your companies is scary? And why? Probably companies with reliance on a few large customers, lumpy orders, industry cyclicality, low and volatile margins, managerial turnover, complicated accounting, or just the plain fact that the company has a history of burning you. SLP's got none of that, as I'll explain.
It has ~300 customers (up from ~100 in 2007), including 19 of the top 20 drug companies and most of the world's largest public health agencies. None of its customers is financially important, and ~90-95% renew their licenses each year. And the source of SLP's ~15% trend growth is broad - it's the 15-25 new customers the company signs up each quarter.
Its after-tax net income margins are ~25%, and that's despite buying an outsourced simulation consulting firm last year - which, of course, diluted margins.
Walt Woltosz is the stud who founded this company in 1996. He owns 40% of it, is still its president and chairman, but nowadays, he spends his time trying to port SLP's software to brand new industries. And Woltosz's the kind of guy you want doing that - he was in the thick of it when simulation went ubiquitous in aerospace in the 1970s. He invented the first augmented speech tool used by Stephen Hawking, almost on a whim, after his mom-in-law came down with ALS; then he founded this company around those tools for the disabled, paired that business up with his real passion - simulation - and this time applied to the next frontier - medicine. When the legacy business stopped performing well, he sold it, made SLP the pure play it is today, and has been growing it profitably for 10+ years. In short: technical expertise, broad contacts, status (Hawking!), and proven business acumen (many smaller peers overestimated how fast this field would grow and failed). It's the little things too: he's always popping up on obscure drug design forums to answer questions not even pertaining to SLP directly - he's ~70, but I expect he'll be around awhile.
Oh, the accounting - the licenses are annual things, so the company recognizes revenue each year in a lump when they're renewed. Simple.
Anyhow, my point is that SLP is less risky than its P/E implies, because the "E" part is really sturdy. The ~$0.30 it'll earn in 2016 is a steady base from which to grow. Yes, if trend growth dropped by 10% immediately, EPS growth would stall (but not decline), and the multiple might drop to 20, implying 40% downside to $6.
But hey, what if your favorite company's sales growth stepped down 10% tomorrow? I bet it'd be close enough to 40% down, and that it's more likely to happen to them than to SLP.
When we start thinking of what's the worst that could happen over longer horizon - i.e. what's your 90% confidence interval on 2021 EPS for your favorite company? - the case for SLP being a relatively safe bet strengthens.
But yes, SLP is priced to grow. And it's got plenty of avenues to do that.
Background On Simulation
A big fraction of R&D budgets are spent on computer simulation in most any research-intensive field - I've heard 5-20% is normal in chips, cars, aerospace, etc. (Individual companies in chip CAD are bigger than this whole field.) But not in drugs. US drug companies spend $50 billion+ per year on research, and yet, as far as I can tell, less than 1% (probably far less) is spent on simulation. For context, SLP is the largest pure play public firm here, and its software revenue is barely $10 million, and GastroPlus, which dominates its niche, does ~$6-7 million and is used by fewer than 300 companies and probably 1,500 users (assuming 5 users per site license). Obviously, what I can't help imagining is what happens to SLP if this situation changes.
The problem is perceived value. Imagine, for the sake of argument, that there are 500 biologically relevant things we need to know about a candidate drug. One way to estimate those 500 numbers is to do physical experiments - have combinatorial chemists synthesize them and then helicopter-drop them into petri dishes containing the malaria parasite, or whatever. That works in principle, but there are a gazillion possible carbon-based molecules - thus, simulation. (SLP's ADMET Predictor can guess 150+ properties of 100,000+ compounds per hour.)
The number of those predictions a computer could make accurately in 1950 was 0; in 1996, it was probably still close to 0. Today, it's much greater than 0 and rising fast. It's now good enough that simulation is used ubiquitously for modeling how an already chosen drug will behave in the body - but for finding brand new drugs, it's more of a filter. It weeds out lots of hay, but leaves behind needles mixed in with lots of hay that resemble needles.
The fundamental question is: How much better can it get? To answer that, we need to know something about machine learning - using computers to make predictions - in general.
(Short version: It is the dearth of experimental data that's probably holding this field back - a problem that is recognized and being gradually surmounted.)
Machine learning depends on three things: computing power, the learning architecture (i.e., statistical regression), and data.
You can think of learning architecture as coming in two types. One kind emulates the conscious thought processes of human experts. You can code into a computer how a doctor diagnoses patients in the form of a decision tree or a bunch of if-then rules that ultimately conclude in a diagnosis, i.e., a probabilistic prediction of what's up with the patient.
The other type emulates not conscious thought, but rather, the basic neuronal structure of the brain. What's required to build these "neural nets" isn't experts telling you how they decide, but lots of raw computing power and lots of data. It is from the data, not experts, that neural nets learn.
Before 1990, neural nets were held back by their computational intensity outstripping available computing power. That's because their basic learning principle is dumb, but gets smart by repeating that dumb thing zillions of times. Kind of like how evolution works: start with something random, mutate, see if the mutation is beneficial, if so keep it, then mutate again.
"Seeing if the mutation is beneficial" means checking to see whether the net's prediction got closer to the known values of the "training" data you fed it. The way voice recognition nets are trained is by talking to them. Speech is converted into electrical pulses, the machine is told what words those pulse patterns represent, and then it iteratively builds a map from pulse pattern to phoneme/word.
Machine learning progressed slowly until 1990, but then computing power caught up, nets proliferated, and one long-standing challenge after another has been surmounted ever since, i.e., chess, checkers, voice and text recognition, even Jeopardy playing.
One key thing about voice recognition, which works good - compared to what SLP does, which works OK for some things, is data. Voice data is easy to get: people love to talk. But drug data is filed away at hundreds of organizations around the world in incompatible databases, and much of it is wrong, since doing what seems like the same experiment in two different petri dishes often produces different results.
Simulation skeptics think the problem goes deeper than data scarcity. They say that most useful drugs have historically been found in nature, and that the simulation-synthetic approach simply can't compete with hundreds of millions of years of evolution in the contest to produce complex biologically active drugs. Nature is where most drugs have been found, and drug companies should turn away from combinatorial chemistry and simulation and start searching rain forests and garbage cans again.
But the more optimistic alternative view - that simulation is held back not by insurmountable barriers, but by inadequate institutions for generating and sharing data to feed to computers - is plausible. Lots of 1990-era neural net pessimism has been proven wrong in other fields where data is easier to obtain, and progress is being made fast.
Dr. Robert Fraczkiewicz, team leader for ADMET Cheminformatics for Simulations Plus, said, "The prediction of ionization constants (called pKa's) for molecules is fundamental to the prediction of many other molecular properties. Our pKa prediction, which was already best-in-class, was based on experimental data for about 13,000 molecules. As part of the collaboration, Bayer HealthCare provided over 17,000 of their proprietary structures and data to Simulations Plus so we could retrain the models using the expanded chemical space represented by their chemistry over our original data. The result is that the new model provides the most accurate predictions available not only for our original data but also for new research compounds synthesized at Bayer HealthCare. It's important to note that the models were built only on a portion of the Bayer HealthCare data, and were validated by predicting the pKa values for a large, completely external set of Bayer HealthCare molecules."
It's important to remember that only SLP's early-stage drug finding tool uses neural nets. And the majority of its software revenue comes from Gastro-Plus, which works more traditionally, by modeling the body's organs as a series of compartments through which drug dissolved in blood flows. This tool's value is non-speculative; it has 2/3rd market share; the FDA is encouraging drug companies to use it; the FDA regularly collaborates with and pays SLP to beef it up; and it may be portable outside drugs, as Dow Chemical (DOW) and the EPA have purchased licenses to help predict the potential toxicities in plastics and the environment.
GastroPlus is an established tool. All that speculative talk about nets and the potential for weeding through haystacks more effectively does not apply to it.
But it's the neural net that's so exciting! As I mentioned, nets learn from data, and they don't much care what that data is. In drugs, it'll be numbers describing a chemical's weight, volume, polarity, the number of nitrogens it has, etc.
But as long as you can describe the object of interest with relevant numbers (inputs), and as long as you have outcome data to train it on (known outputs), the net might be able to build you a useful map.
Woltosz is spending his time these days trying to convince aerospace companies and medical researchers to fund ports of SLP's net to aerodynamic problems and correlating brain scans with mental diseases. This got a started a year or two ago, when his Auburn aerospace buddies showed him a problem that's was too computationally intensive to calculate using mechanistic physics equations. SLP took the data, trained its net, and got a good result fast. In both these fields, SLP is out speaking at conferences and trying to obtain funding.
Aerospace, medicine, chemicals manufacturing, and environmental analysis: none of these fields generate much - or any - revenue yet. None are required to justify SLP's valuation - they can, and are, winning in drugs alone. But they sure do have lots of potential growth avenues.
A more certain thing - the company's consulting business has a bright future. A year or two ago, it bought Cognigen, the first firm to offer simulation consulting services on an outsourced basis. It paid ~1x sales for a thing that's now growing even faster than its software business, whose margins are 10-20%, and that's out in the field showing customers how to get value from SLP's software. Basically, the company paid ~1x sales for an accretive independently profitable marketing arm.
SLP says it can take 1-2 years to fully train its new reps in the software, and that there's not enough people at drug companies capable of using it well. So I don't see any reason to doubt that the consulting business can grow fast and profitably indefinitely. (I also wonder whether consulting plus simulation means it's cheaper to get into the drug design business, so maybe lots of small startups - filled with tech savvy young people, not old chemists - will come along and broaden this market.)
Value investing isn't buying cheap stocks. It is doing whatever is effective and simple to apply. Tilting towards value happens to be one of those things. Preferring situations where the downside is quantifiable and tolerable is another. Betting on good people is another - recall that Buffett was a founding investor in Intel (NASDAQ:INTC), not because he or anyone else knew much about chips in 1967, but because he wanted to bet on Bob Noyce. Value investors aren't willing to pay up for growth unless it's accompanied by durable competitive advantage, and the moats in software that's complicated to use and improves in proportion to the user base are wide and deep. Finally, value investors recognize that at the end of one's career, it's often just a few names that delivered most of the returns. So, it pays to be on the lookout for situations where lollapalooza effects operate, i.e., many powerful things pointing in a single direction. And here we've got the right people running a leading positive-feedback loop business at the intersection of AI and biology - the two fields most likely to transform the world in our lifetime.
I'll pay 30x EPS for that, shut my eyes for a decade or so, and see where it goes.
Disclosure: I am/we are long SLP.
I wrote this article myself, and it expresses my own opinions. I am not receiving compensation for it (other than from Seeking Alpha). I have no business relationship with any company whose stock is mentioned in this article.