Eighteen months was the number we used to quote when partners asked how long lead identification takes. It wasn't a guess — it was built from hard experience across three programs. Screen a few million compounds, chase the actives, run SAR, watch half the series collapse when you add a methyl group in the wrong place, start over. Eighteen months, if you didn't hit anything catastrophically wrong.

The CAI-001 program changed that number to four. Not through some abstract efficiency gain, but through a specific set of decisions about where to use computation and where to stop trusting it.

The Problem With Traditional HTS

High-throughput screening is seductive because it's concrete. You have physical plates, real assay data, IC50 values you can actually touch. What it doesn't give you is context. You find a hit with a Ki of 340 nM against KRAS G12C and you have no idea whether it's going to aggregate, pan-assay interfere, or simply fall apart the moment you try to improve potency. The chemist's job historically was to figure that out by making things — a process that is slow and expensive by design.

The counterintuitive move with CAI-001 was running far fewer physical assays up front. We used our molecular modeling layer to pre-filter the virtual library before anything went to synthesis. Starting from a 14-million-compound make-on-demand catalog, we scored every entry for predicted binding to the switch II pocket of KRAS G12C, penalized anything with molecular weight over 450 Da, filtered on cLogP between 1.5 and 3.5, and removed any scaffold that our ADMET models flagged for CYP3A4 time-dependent inhibition. That took 14 million compounds to roughly 8,400.

Of those 8,400, we ran docking calculations using an ensemble of 12 KRAS G12C crystal structures from the PDB — not just the apo form, but multiple GDP-bound and covalent-inhibitor-bound conformations. Compounds that scored poorly in fewer than three conformations were dropped. We ended up with 312 compounds for physical synthesis and assay.

What the Numbers Actually Looked Like

Of the 312 synthesized compounds, 47 showed activity below 1 µM in the primary biochemical assay. That's a 15% hit rate — meaningfully higher than typical HTS campaigns, which run 0.01% to 0.5% depending on the target class. Hit confirmation gave us 41 reproducible actives. SAR analysis clustered them into five distinct chemical series.

Three of those series had fatal liabilities visible from the structure alone — one had a Michael acceptor that would react non-selectively with any proximal cysteine, one had a reactive aldehyde that we somehow missed in the pre-filter (a gap we've since fixed), and one had a quinoline scaffold with a known hERG liability that our models didn't adequately penalize. That left two viable series.

Analog synthesis on the two lead series ran in parallel rather than sequentially. Because we weren't chasing dead ends, the medicinal chemistry effort stayed focused. Lead optimization across both series generated 68 analogs over 11 weeks. The final lead compound — the one that became CAI-001 — had a Ki of 9 nM, a selectivity ratio of greater than 200-fold against wild-type KRAS, acceptable microsomal stability (human liver microsome t1/2 of 48 minutes), and no red flags in the seven-species AMES test.

Where Computation Helped and Where It Didn't

The honest answer is that computation helped most in compound prioritization and hurt us nowhere we couldn't recover. What it didn't do is replace any individual chemistry decision. The choice to chase the switch II pocket with a covalent warhead rather than an allosteric GTP-competitive approach — that was a human call informed by the clinical failures of earlier KRAS programs. The decision to tolerate slightly elevated lipophilicity in exchange for better cellular permeability was also human-driven, based on cell-based potency data that the models had no visibility into.

Where the platform genuinely compressed time was in eliminating the period between "we have a hit" and "we understand the SAR." Traditionally that's six to nine months of iterative synthesis with each round taking three to five weeks. With the platform running predictive SAR in real time alongside the experimental data, we could deprioritize whole regions of chemical space before spending synthesis resources on them. It doesn't feel dramatic. It just means you're never making the compound you already know won't work.

The Caveats Worth Stating

Four months is not a universal number. KRAS G12C has excellent structural data, an active covalent chemistry community, and a well-validated assay format. The same approach on a structurally novel target with limited crystal data would not compress as dramatically. Targets with poorly defined binding pockets, significant conformational flexibility, or assay technical challenges don't benefit as much from virtual screening.

We also had a team of medicinal chemists who understood when to trust the models and when not to. That judgment is not something you install with software.

Still, across four programs now, the pattern is consistent: pre-filtering with predictive ADMET and binding models before physical synthesis compresses lead identification. Not always to four months. But reliably by more than half.